Professor Timothy T. Rogers is the principal investigator of the Knowledge and Concepts Laboratory currently in the Department of Psychology at the University of Wisconsin-Madison and moving to the Institute for Cognitive Neuroscience at University College London in summer 2026. His research areas include Biology of Brain and Behavior, Developmental, Perception, Cognition, and Cognitive Neuroscience. Tim was in conversation with Shrankhla Pandey, a PhD student at the Department of Computer Science and Technology, University of Cambridge.
Could you share the key challenges your work addresses and why are these important to address for the future of Language Sciences?
We use a lot of jargon in our science. My work focuses on what we call semantic memory or semantic cognition, that just means the ability to understand language and interpret events in the world. This includes perception, recognizing objects, and making inferences about their properties, abilities we often take for granted as part of everyday mental life. I am particularly interested in how these abilities arise from neural systems, how biological computations carried out by networks of neurons give rise to these complex mental phenomena. For example, I can perceive where I am, recognize the objects around me, and engage in this conversation with you, all while maintaining a coherent understanding of what is happening in the world. That capacity, which we believe is uniquely human, has been studied for last 20-25 years using artificial neural network models, the same technology that underpins contemporary AI. These models serve as tools for developing hypotheses about how the brain might support such cognitive functions.
This research matters for several reasons. First, it's fuelling these new technologies, for good or for ill. But on a more human level, it addresses critical challenges in mental and neuropsychological health. For instance, individuals suffering from progressive dementias gradually lose their ability to understand the world due to neuropathological processes that damage the neural systems responsible for these functions. We'd like to understand better how the healthy system works so that we can understand how to remediate effects of disease. Moreover, there are a variety of sort of social goods that can come out of that for people, even beyond the innovation of artificial technologies that seem to be able to understand the world.
What shaped your journey from undergrad in psychology into the interdisciplinary field of language sciences?
As an undergraduate psychology major, I remember finding the questions posed in my courses extremely interesting, but many of the hypothesized solutions felt deeply unsatisfying, partly because they didn’t connect to what we knew about the brain then. For example, we discussed concepts like capacity limits in working memory, which is a fascinating phenomenon: you can’t hold more than about seven items in mind over a short span of time. But simply calling that a “bottleneck” or a “capacity” didn’t explain anything; it was just describing the phenomenon. Why should a brain with billions of neurons only be able to remember seven things for a period of time? That seemed like a question that had to relate back to the biology of the system.
At the time, cognitive neuroscience was just emerging, though in my department at the University of Waterloo it was called biological psychology, a term still used in some places. I took that course and learned the basics of how neurons work. I had also dabbled in computer programming, and I remember thinking: you could write a simple program to mimic what a neuron does, and then allow those neurons to interact. When I spoke to my undergraduate advisor, he pointed me to Parallel Distributed Processing, a seminal collection of works that introduced neural network models as tools for bridging biology and cognition. I realized my little idea had already been deeply explored by scientists, and I was excited to learn more.
That connection, a set of computational tools that were brain-like in some ways but allowed us to simulate simplified versions of the processes studied in psychology, provided just the right blend of biology, computation, and cognition to make progress. When I began looking for graduate schools, I was fortunate: Jay McClelland, one of the pioneers of these ideas, happened to be recruiting students that year. He accepted me into a cross-disciplinary training program where I took the same neurobiology courses as neuroscience majors, while they took psychology courses alongside me. Computer scientists were doing both. It was an early effort to integrate these three disciplines, and my career has benefited enormously from that approach.
I started graduate training in 1995 at Carnegie Mellon University in the Center for the Neural Basis of Cognition program, one of the first cross-disciplinary initiatives connecting cognition, neuroscience, and computation, similar to the Institute for Cognitive Neuroscience at UCL. I got my PhD in 2000 and then joined the MRC Cognition and Brain Sciences Unit (CBU) for a postdoc. The CBU was another early adopter of cognitive neuroscience, deeply connected to neuropsychology and clinical applications, not just helping patients and caregivers understand neuropathology and disease, but also innovating ways to study these populations to inform our understanding of both healthy and damaged brains.
My time at the CBU was critical for linking cognition and computation to clinical and pathological studies. At the time, my mentor Karalyn Patterson and neurologist John Hodges, working with my long-time collaborator Matt Lambon Ralph, had pioneered a tight integration between clinical and caregiving mission on the one hand and the basic research mission on the other. I really can't speak highly enough of their efforts. They developed a system to visit patients with unusual forms of frontotemporal dementia, progressive conditions affecting the anterior temporal and inferior frontal lobes, every six months in their homes. These patients had cognitive impairments very different from Alzheimer’s or other common dementias. By building relationships with patients and caregivers and evaluating them on a comprehensive suite of behavioural tests, we could track changes longitudinally, to see how their behaviour was changing over time and identify sites of selective impairments and islands of preserved functions, and study syndromes across populations rather than treating each case in isolation.
This approach enabled a case-series methodology and allowed us to provide patients and caregivers with tailored, up-to-date information about their cognitive state and what to expect, specific to these rare disorders rather than generic Alzheimer’s advice. It was a major innovation that couldn’t have been done before. There were challenges: traveling to hospitals was a huge barrier, and these patients were often socially isolated as friends and family withdrew when language and comprehension declined. I remember that John and Karalyn arranged regular meetings for the caregivers to come in and meet one another and talk about their own experiences. Of course, this work was costly, requiring funding for travel, staff, and infrastructure. It was only possible under the MRC unit model, which provided long-term funding stability. Sadly, that model is now being dismantled in the UK, which I think is a real loss (read more here https://www.bmj.com/content/387/bmj.q2878.short ).
You’ve been successfully working across disciplines for more than two decades. What advice would you give to researchers who want to collaborate beyond their own discipline?
One general piece of advice is that people often feel intimidated about attending talks outside their department because they worry they won’t know enough, “I’m not a computer scientist; I don’t really understand that field.” But the truth is, nobody knows everything. Effective collaboration doesn’t require being an expert in every domain; it requires having just enough knowledge about another area to hold a meaningful conversation on a shared topic of interest.
When I was at Wisconsin, we created a cross-disciplinary PhD training program based on this philosophy, which we called “just enough cross-disciplinarity.” We organized events where people from different departments could talk in plain, non-jargon language about their work. The goal was to help participants understand the kinds of problems others work on and the skills they bring to the table. That way, someone might think, “If I could apply that technique I learned about in a CS (computer science) talk to my MRI dataset, we could do something really exciting.”
One of the biggest barriers is jargon. People studying similar phenomena from different disciplines often use specialized language that isn’t accessible to outsiders. Having one or two people in the room who can pause and clarify, “When you say this, do you mean that?”, helps everyone get on the same page and makes collaboration possible.
So my advice is: attend seminars in other departments. Don’t worry if you don’t understand everything, you’ll absorb more over time. And from the perspective of the Language Sciences group at Cambridge, you have outstanding scientists, students, and postdocs within silos, and there’s clear interest in building cross-disciplinary ties. To the extent possible, create events, like the CLS symposium, that allow for “fortunate collisions,” where people interact informally over coffee, beer, or shared seminars.
When I arrived in Madison, I met computer scientists and engineers with similar interests—for instance, Rob Nowak (an expert in optimization theory) and Jerry Zhu (an expert in machine learning). Instead of separate meetings, we decided to bring everyone together in one room, and make it fun. We started a seminar called HAMLET (Human, Animal and Machine Learning: Experiments and Theory). It was held on Friday afternoons, and the speaker had to bring beer for everyone. The atmosphere was informal: instead of polished talks, speakers shared ongoing projects and challenges. The mix of people allowed everyone to learn about other disciplines in a relaxed setting.
If you’re in academia, collaboration should feel enjoyable. Create opportunities people look forward to, where they can chat outside their discipline, and the science will grow from that. Especially in England, where going to the pub after work is a natural social activity, these informal settings help break down barriers. They make it easier to say for example, “I don’t know what you mean by L1 regularization,” and have someone explain it to you.
Is there an interdisciplinary project you’re especially proud of?
It’s hard to choose just one, so I’ll share two projects I’m particularly excited about.
The first is in my main area of research. Working with computer scientists and engineers, we have developed methods for decoding information across the entire brain at once, rather than analysing individual regions separately. Functional brain imaging has undergone a major revolution with the application of machine learning to understand how information is encoded in different regions. However, brain imaging studies face a significant statistical challenge: we often have thousands of neural activity estimates but far fewer observations, creating a severe overfitting problem.
When I was at Wisconsin–Madison, I began collaborating with the optimization team in the engineering department, experts in finding sparse signals in vast amounts of noisy data. Through discussions about how information might be distributed in the brain and why standard approaches fail for decoding, we developed a suite of tools that fit decoding models under assumptions about distributed information within and across individuals. These tools now allow us to decode information that we know is present in neural activity but could not detect with previous methods. I’m excited because, for the first time, we can identify information networks rather than isolated areas, which could fundamentally change how we think about brain organization.
The second project is more recent and addresses a very different question: why do people hold strongly conflicting beliefs about phenomena, even when exposed to similar information? For example, some believe human activity causes global warming, while others think it’s a hoax. Social cognitive psychology has long used computational models of belief propagation, but traditional agent-based models assume overly simplistic rules, such as averaging opinions, which we know from research are inaccurate. People discount some sources and overweight others.
What’s new now is the arrival of large language models (LLMs). In our project, we simulate individuals as LLM-based personas. We can instruct a model to role-play as, say, a 45-year-old Midwestern farmer sceptical of vaccines, or a Black woman veterinarian in Chicago, and then generate tweets reflecting those beliefs. We create populations of such personas and let them interact, measuring how attitudes shift over time. The goal is to tune these simulations to better capture real human responses to challenging information, creating what we call an “AI terrarium”, a miniature society for testing strategies to promote accurate information and counter disinformation.
This work involves collaboration not only with computer scientists but also with colleagues in journalism. It raises important questions about implicit bias in LLMs and how underrepresented populations in training data affect outcomes. Interestingly, our initial findings show that LLM personas converge very quickly on scientific consensus, unlike humans who often resist change. This likely reflects reinforcement learning processes that push models toward ground truth, making them less useful for simulating real-world dialectics. Here, cognitive psychology plays a critical role: we know how to measure human behaviour and design experiments that mirror these simulations. Last year, we published a paper showing that across topics, from vaccination to other contentious issues, LLM-based groups rapidly align with consensus, unlike human groups.
I’m eager to see how this line of work evolves. Many similar efforts are happening behind closed doors in political campaigns and industry. By pursuing this research openly, we hope to clarify how these technologies can be used responsibly and for public good.
Do you use large language models (LLMs) in your day-to-day?
Yes, I do. In fact, during my sabbatical in Cambridge last year, I worked on a project where, for the first time in a long while, I handled every aspect myself, designing the experiment, programming, and collecting the imaging data. This meant learning a lot about the imaging processing environment at the MRC in Cambridge, which differs from our setup in Madison. Normally, my students handle much of this work, so I had to teach myself various Python packages and software tools for tasks like combining multiple echoes into a single image.
Initially, I tried asking colleagues for help, but they were often super-busy and I felt bad about disturbing them all the time. So I thought, “Why not ask GPT?” It turned out to be like having a teaching assistant (TA) at my shoulder, available 24/7. It wasn’t always perfect, but it was close enough that I could use my own understanding to adapt its suggestions. Since then, I’ve found it invaluable, especially for quickly getting up to speed with new computational techniques.
I don’t use it for writing or for generating entire workflows. I use it the way I would use an on-call TA: when I don’t understand something, I explain the problem, and it provides explanations and code snippets I can adapt. For me, coding requires a clear mental picture of the entire workflow. If you let an LLM generate everything, it’s like trying to understand someone else’s code. But as a collaborative tool, it’s fantastic.
One of the major challenges in cross-disciplinary education and science is scalability. Everyone is busy, especially senior PIs whose time is consumed by multiple responsibilities. In our training program, we experimented with team-based meetings instead of one-on-one sessions. For example, we would have two PIs, one from computer science and one from psychology, along with a group of postdocs, graduate students, and undergraduates collaborating on the same project. The conversations still happened, Rob explaining something to me, me explaining something to Rob, but in a group setting where others could ask questions and clarify. This worked well as long as we had a handful of projects that interested both faculty and students. However, when the program grew and we needed 25 such projects, it became unwieldy.
I’m optimistic that large language models (LLMs) can help here. Instead of needing to be in the room with advisors, an LLM can act as a 24-hour TA and even as a communication facilitator. For example, last year I faced a problem that a colleague described as a “bipartite graph.” I had never heard of that term. During my sabbatical, I asked GPT to explain bipartite graph matching, and it did so in a way that made me realize how my problem mapped onto that concept. I learned that this is a well-known engineering problem with established algorithms and toolboxes for optimization. From there, I was able to code the solution myself. What I really needed from my colleague was the keyword, the conceptual bridge to the right set of techniques. GPT provided the rest, saving me from multiple meetings. My problem involved comparing two brains: each had a set of voxels carrying information, and I wanted a metric to measure similarity between selected voxels across individuals. My approach was to imagine pairing each voxel in brain A with one in brain B and minimizing the sum of distances across all pairs, a classic bipartite graph matching problem. Years ago, combining such an abstract concept with a new domain like functional brain imaging would have seemed impossible without deep conceptual understanding. This ability to recombine concepts in novel ways is what we often mean by “conceptual innovation.”
It’s fascinating, and challenging, to articulate how human intuition differs from what models do. LLMs are now performing tasks we once thought required profound understanding, and pinpointing that difference will be very difficult.
Which aspect of your work do you see impacting Language Sciences?
Right now, there is considerable interest in the research community in understanding how large language models, neural network models trained on massive language corpora and forming the backbone of current AI technologies, are similar to, and different from, the ways humans understand and produce language. Much of the work in computer science is driven by making these systems function effectively, given the technologies we have. My work in cognitive science and cognitive neuroscience brings a different perspective: identifying where these enormous models diverge from human cognition and exploring whether innovations in cognitive science can inform machine learning.
For example, we have conducted recent studies examining large language models in tasks designed to estimate the mental structure underlying human knowledge in a given domain. One major concern with these models is their ability to provide accurate information, they aim for ground truth answers, drawing on vast internet data. Humans, by contrast, are not perfect memory systems; we exhibit systematic memory distortions and phenomena that reveal how human memory, language, and concepts work. Interestingly, when we run the same tasks on large language models, their accuracy often produces representational structures that look very different from those observed in humans. In other words, the kinds of decisions people make, sometimes deviating from ground truth, allow us to uncover meaningful structure in experience and language, something current models do not replicate.
We see clear differences between human cognitive models and transformer-based language models. This suggests potential for bidirectional impact: not only machine learning influencing cognitive science, but also cognitive science informing AI. There are aspects of human cognition, what we consider real understanding, that are not captured by contemporary models. Incorporating these insights could reshape both language science and emerging technologies that are transforming the world.
We already have research studies underway testing these hypotheses, and interest in this area is growing rapidly.
What broader changes, societal or technological, do you think your research could help shape?
A blue-sky goal would be to understand the connection between neural and cognitive mechanisms well enough to address a wide range of neurological and neuropsychiatric problems that currently have no solutions. People often talk about neuroprosthetics, systems involving sophisticated electrode arrays implanted in the brain to read neural activity and use neural networks to interpret it, translating signals into planned motions. This is already happening in animal models and some human trials for individuals who have lost motor function. For example, if the motor cortex is intact, we can interpret its signals and convert them into commands for actuators in a robotic arm, a scenario that feels like science fiction but is becoming reality.
Motor control is one area where we have some understanding of how primary motor cortex activation relates to intended movement. But consider the challenge of restoring language function in someone who has had a stroke, or memory in someone who has lost their hippocampus due to viral encephalitis. These are grand challenges that require a fundamentally different way of thinking about brain function than the traditional view of isolated regions performing discrete functions. The field is moving toward understanding whole networks, how distributed and coordinated patterns of activity across brain regions give rise to thoughts, ideas, memories, and intentions.
Solving these basic science questions could pave the way for extending neuroprosthetics technologies to a much broader range of cognitive functions. It might sound like science fiction, chips in the brain to restore episodic memory, but serious efforts are underway. I was once part of a team invited to submit a proposal to DARPA, the U.S. agency that funds cutting-edge research, aiming to build an artificial hippocampus. The team believed they understood enough about hippocampal function to attempt restoring episodic memory in individuals who had lost it.
These ideas have the potential to transform healthcare and human cognitive functioning in profound ways. But achieving them will depend on developing a much richer and deeper understanding of the connections between brain systems and cognition than we currently possess.
As we are looking ahead to 2050, what do you hope we will have achieved in managing and understanding these technologies, and what kind of impact do you expect cognitive science to have?
Like many others, I am deeply concerned about the rapid development of technologies whose capabilities we do not fully understand. These systems are extraordinarily powerful, and my hope is that cognitive science and language sciences will play an important role in establishing constraints and guidelines for how such technologies are used and deployed.
To give some context, Geoffrey Hinton, a computer scientist who collaborated with cognitive scientists in the 1980s and 1990s and pioneered many of the technologies we now see worldwide, has essentially dedicated the latter part of his career to warning about poorly understood systems. Similarly, at the recent Cognitive Science Society meeting, the winner of the Rumelhart Prize (the most prestigious award in cognitive science), Douglas Medin, devoted his acceptance speech to emphasizing that many societal problems can be understood as cognitive problems: how beliefs are formed, how persuasion works, why people respond differently to the same evidence, and why we divide ourselves into competing social groups.
If I am hopeful about the future, because our discipline can contribute to making these systems more human-like and to using cognitive neuroscience and psychology to better understand what these technologies are doing and how they work. This understanding could help guide us in preventing the misuse of AI for influencing elections, spreading propaganda, or making rapid yet ethically fraught decisions about national defence. My aspiration is that science becomes part of the effort to ensure that the enormous benefits these technologies can offer are realized responsibly.
What is your ambition for the field of language sciences as we look toward 2050?
One of the most urgent ambitions for the field is to understand how to build and enforce guardrails for technologies that are rapidly becoming embedded in everyday life, prioritizing broader societal concerns rather than just engagement metrics. From my experience, I have many colleagues in industry, and I don’t think leaders at companies like Google or Meta are solely driven by profit. There is genuine interest in understanding the ethical, legal, and societal implications of these technologies.
However, one persistent challenge I’ve observed over the years is that people in computer science often feel they have sufficient expertise to address these questions. I don’t think that’s true. This is where cross-disciplinary collaboration really matters. We need experts from the humanities and social sciences, people who think deeply about what makes a good society, what can go wrong, and how human nature’s darker aspects manifest. These scholars have relevant expertise, but often lack the technical language to communicate effectively with computer scientists.
I have a close colleague in computer science who once said, “I don’t know what fairness is, just give me a number I can optimize.” Concepts like fairness, bias, and equity are complex and deeply studied in the social sciences and humanities. Reducing them to a single metric for optimization is alien to those fields. Bridging this gap is critical. We need hands-on collaboration that captures the attention of multibillion-dollar industries, where every model costs millions to train and the stakes are enormous. These companies won’t change based on a single professor’s opinion; what’s needed is a coordinated, cross-community effort to communicate major concerns in ways that earn trust and influence engineers and computer scientists working on these systems.