April 25, 2025
How do humans and machines build models to enable problem-solving and innovation? This is the question that has shaped Kelsey Allen’s career. It’s guided her from high-energy physics to machine learning, and now to the Vector Institute, where, as the newest Vector Faculty Member, she will investigate the mechanisms fuelling adaptive learning, specifically with reasoning and problem-solving skills.
But for Allen, the journey into cognitive machine learning was anything but linear.
When she started her undergraduate studies in physics at UBC in 2008, she envisioned a future in high-energy particle physics. Yet, as she neared the end of her degree, she found herself increasingly drawn to the intersection of physics and machine learning.
“I wanted to apply machine learning techniques to actually try to discover new particle interactions,” Allen recalls. “And that was really theoretical at the time. Nobody thought that was an okay thing to do.”
In 2014, while working on an internship project at the University of California, Davis about modeling ant foraging behavior as neural circuits, she started to see how biological systems could be analyzed using machine learning. Around the same time, another undergraduate internship introduced her to using satellite imagery to detect ice and agricultural changes. But machine learning wasn’t yet the dominant force it is today, and Kelsey remembers being skeptical.
However, as she worked on her honors thesis in particle physics, she found herself frustrated by the constraints of the field. So much of the research at the time was focused on trying to find evidence for theories that researchers had already formulated and believed in. “I wanted to use machine learning to do the opposite—to look at the data and let it guide us to new discoveries, without a predefined theoretical framework.”.
That’s when she seriously began considering a PhD in machine learning —due to the limited knowledge of AI and its applications.
Kelsey’s decision to pursue a PhD at the Massachusetts Institute of Technology (MIT) was driven by the flexibility that the program offered; it allowed her to explore different areas before committing to a specific path. As a first-year PhD student, she took a class on computational cognitive science that was part of the Center for Brains, Minds, and Machines. “It was an incredible environment,” she says. “We had machine learning researchers working alongside neuroscientists studying monkey cognition and psychologists studying child development. It made me think about intelligence in a much broader way.”
During her PhD, Kelsey interned at Google DeepMind, which led to a full-time position after graduation in 2020. Working there for four years in San Francisco and London, UK, was eye-opening: “I got a lot more exposure to high-tech machine learning than I was able to do during my PhD. It was a different way of thinking about AI.”
Allen’s research career has taken her across three major AI hubs — Canada, the US, and the UK — giving her a unique perspective on how each region approaches AI.
“In the US, people were much more optimistic about technology being only a force for good,” she says. “It felt like people just wanted to push the capabilities as far as they could.”
By contrast, in Canada and the UK, researchers were more critical — considering not just AI’s potential benefits but also its risks. In both countries, Allen observed a stronger emphasis on socially responsible AI applications, such as using large models to facilitate democratic decision-making.
Much of Allen’s current research focuses on understanding intelligence by drawing parallels between human cognition and machine learning. One of her key interests is tool creation — a longstanding goal in cognitive science that provides insights into intelligence across species.
“One reason people care about language is that they think it’s unique to humans and very powerful. But language isn’t something you can test across species very well. There’s no continuous scale of ‘how much language does a dolphin have relative to a spider?’
Instead, Allen became fascinated by tool use, an ability once thought to be uniquely human but now recognized in other species too. “It was a very interesting area to work in that blended cognitive science and this fascination with these very complex behaviors that can tell us something about intelligence as a continuous spectrum instead of just like ‘do you have language or do you not have language?’
“It’s also very hard to get a robot to creatively use tools, which is why tool creation is such an interesting problem in AI,” she adds.
She has explored tool use further in her research, including in a publication on differentiable physics and tool-use planning that received the best-paper award at the “Robotics: Science and Systems” conference in 2018.
In fact, her interest in this area has even bled into her leisure time. A native British Columbian, she has a deep appreciation for the Museum of Anthropology at UBC, “an exceptionally cool museum that showcases the First Nations and Indigenous artifacts of the Vancouver area.” She is also fascinated by the intricate tools and artifacts on display, which offer a glimpse into the problem-solving ingenuity of Indigenous communities.
Another area Allen has explored is using games as research tools. While traditional psychology experiments often do not capture the complexity of real-world decision-making, games, she argues, bridge the gap between controlled research environments and complex human interactions, while also incentivizing participants to perform well.
“What I liked about games is that they allow you to go in between these extremely simple, controlled experimental settings, and the messiness of studying people’s behavior in the real world, while also engaging people to really care about doing the experiment from this intrinsic motivation perspective of just enjoying the experience,” she says.
Naturally, Allen also has strong opinions about some of her favourite board games, amusingly noting that she prefers ones “that are cooperative because they lead to less tense friendships long term.” Thinking of it from a research standpoint, however, she gravitates toward open-world games like Minecraft, which allow players to create new worlds and problem-solve freely.
Allen is excited to bring her unique blend of expertise in cognitive science, physics, and machine learning to Vector’s growing research community. One of her focuses will be developing more realistic “world models” — training AI on large-scale datasets to enable more robust decision-making and innovation. With support from Vector’s strong ecosystem, she hopes to refine techniques that enhance AI’s ability to make decisions in dynamic, real-world settings alongside humans.
As an Assistant Professor at UBC, Allen will also be teaching a graduate course that is cross-listed between psychology and computer science, a rare interdisciplinary offering. This course, she says, will allow students to see intelligence from multiple perspectives and explore how machine learning models build their own representations of the world.
Looking ahead, Allen hopes to teach courses on designing experiments to test different capabilities, like memory, decision-making, and perception, both in humans and foundation models.
“It’s really critical that we think about how to design experiments really carefully, to actually be able to pull out what [these systems] understand and what they don’t, and then have students do projects where they’re comparing [models] to human intelligence and getting students to understand experimental design that way.”
