In the latest installment of our Innovator Spotlight series, we interview Chris Fisher, a Cognitive Modeler at Parallax Advanced Research. With a passion for unraveling the complexities of human cognition, Chris works on developing cutting-edge models that enhance decision-making and problem-solving in various fields. Chris shares insights into his approach to research and development, the research problems he’s working on, and the exciting projects on the horizon.
Background and Career Path
Q. Could you share a bit about your background and what led you to pursue a career in cognitive modeling and computational science?
My career path has been very unusual and circuitous. During high school, I was the quintessential under-achiever: I spent much of my time hanging out with friends and skateboarding late into the night rather than studying. I did the bare minimum to graduate. One aspect of school I did like was what was termed at the time Industrial Arts, which included classes like carpentry and machining. My knack for building things with my hands lead to an internship my senior year of high school as a surface grinder at a local tool and die company where I ground electric motor components to their final dimensions, with tolerances of ± .0002 inches in some cases (much smaller than the width of a human hair). After graduating early, the internship turned into a full-time job. Six months later I was laid off during the dot com recession. I enjoyed the challenge of learning to surface grind, but there was not much room for continual growth in this line of work. I know the monotony of continuing this job for decades would be unfulfilling. For this reason, I enrolled in community college for computer programming, where I also took a wide range of elective classes, including psychology. At the time, programming did not click, but I did find interest in psychology, particularly topics such as reasoning and decision making. Against the advice of numerous people, I changed my major to psychology and transferred to a local university where I completed a BA. During this time, I developed an interest in statistics, which set the foundation for my future interest cognitive modeling in graduate school at Miami University. Although statistics was interesting to me, I found it limited in many ways because it merely quantified effects rather than providing insights about the cognitive processes underlying the effects. One thing I found useful about cognitive models is their ability to sharpen research questions. Thinking formally, and programming a simulation requires a person to make assumptions explicit. Otherwise, the code would not run, or your model would do something unusual, or erratic. I decided to teach myself the necessary programming and mathematical skills to use cognitive models in my research. Interestingly, I came back to programming full circle.
Q. What drew you to Parallax Advanced Research, and how has your role evolved over time?
My wife Mary Frame PhD is a fellow Parallaxian (who I believe coined the term) and is my connection to Parallax. Based on her experience, I know Parallax had a work culture and work-life balance that is compatible with my goals. Once I had an opportunity to work at Parallax, I applied and was fortunate enough to be hired. It is a great place to work, with talented employees, and compassionate managers who act with integrity. I do not regret my decision at all. I hope to work here for many years in my same role as a cognitive modeler. I enjoy doing science, writing code, and solving problems. I don’t see my role at Parallax changing anytime soon.
Research and Achievements
Q. Your recent research on likelihood functions within the ACT-R framework has garnered significant attention, even earning you a prestigious award. What sparked your interest in this area of cognitive modeling, and how does it enhance our understanding of human cognition?
As I mentioned above, I have interests in statistics and cognitive models. You can think of a likelihood function as a statistical bridge that connects a cognitive model to data. A cognitive model makes predictions about some aspect of human behavior or cognition, and those predictions need to be compared to data which have structure but also some degree of randomness. The work on developing likelihood functions for ACT-R was at the intersection of my interests in cognitive modeling and statistics, making it a rewarding project to work on. Using likelihood functions has two primary benefits: (1) it enables the use of Bayesian statistical techniques, which allows the modeler to quantify uncertainty in parameters and predictions made with the model, and (2) it increases the speed and efficiency with which a cognitive model can be applied to data (at least in cases where the math is tractable).
Q. For those unfamiliar, could you briefly explain ACT-R and why it’s so important for modeling human cognition?
ACT-R is a type of cognitive model called a cognitive architecture. You can think of a cognitive architecture as a general theory of cognition, spanning areas as diverse as memory, perception reasoning, decision making, and problem solving just to name a few. The basic idea behind cognitive architectures is to synthesize findings across these areas into a single theory which can be tested by developing a model of any task (e.g., remembering information, driving a car, managing limited resources etc.) with a common set of principles. ACT-R has modular architecture, consisting of several memory systems, perceptual systems, goal tracking, and other systems. ACT-R describes information processing bottlenecks and how each of these systems interact to produce human cognition and behavior on a wide variety of tasks. One reason ACT-R is important is because, as a general theory of cognition, it can scale up to more complex tasks, such as those found in many DoD applications.
Figure attribution: Fisher, C. R., Morris, M. B., Stevens, C. A., & Swan, G. (2024). The role of individual differences in human-automated vehicle interaction. International Journal of Human-Computer Studies, 185, 103225.
Q. Earlier this month, you submitted a paper introducing a new method for testing core assumptions of cognitive architectures. Could you tell us more about this research and why you believe it will have a significant impact?
Earlier this month, my colleagues Joe Houpt, Othalia Larue, and Kevin Schmidt submitted a paper introducing a new method that addresses a fundamental challenge: the proliferation of many cognitive architectures (approximately 49 in total), and a lack of formal methods for testing their core assumptions. If we look at the prediction landscape, there are many regions where different cognitive architectures make similar predictions using different explanations. This leaves us with the challenging task of figuring out which cognitive architecture to use.
To address this problem, we developed a framework called Systems Factorial Technology Global Model Analysis (SFT-GMA) to test core assumptions of cognitive architectures. Using this framework, we derive invariant mathematical relationships from core architectural assumptions and develop an experiment to test the mathematical relationship. The basic logic is that if people violate a predicted relationship or do something impossible under the assumptions of the cognitive architecture, we can rule out that architecture and make progress towards narrowing down the number of cognitive architectures researchers are using in applied and basic research. When applying our method to ACT-R, we identified several important constraints on the way the architecture can process information and complete a task. One constraint is that it cannot process visual stimuli in parallel regardless of values assigned to parameters and how information and rules are structured in its memory systems. Looking back into the literature, we found a study showing that people often process visual stimuli in parallel, especially after becoming familiar with the task. This suggests that there are fundamental limitations in ACT-R’s visual processing, which could lead to incorrect inferences in some cases.
Caption: An illustration of the SFT-GMA framework. The basic steps for applying SFT-GMA are listed across the top. The grey rectangle represents the dimensions of the SFT model space projected onto two dimensions. This represents all possible model classes. A hypothetical cognitive architecture occupies a subspace represented as a blue rectangle. Observing an outcome outside the blue rectangle (including the x) would falsify the cognitive architecture. The bottom right shows analyses of hypothetical data consistent with different SFT models.
Why is this important? Any time a person uses a cognitive architecture, they bring to the table a set of assumptions which needs to be tested thoroughly. By testing core assumptions, we can rule out some cognitive architectures and have more confidence that remaining cognitive architectures provide reasonable insights. Although our new method is used primarily in the context of basic science, it is easy to see how placing cognitive architectures on sound theoretical footing has downstream effects for applied research, whether that is improving operator workflows, assessing workload implications of user interface designs, or intelligent teaming.
Q. How does this new method build upon your previous research, and what challenges did you face while developing it?
This SFT-GMA builds on my previous research by incorporating formal mathematical approaches to theory testing—something that has not been predominant in the cognitive architecture field. One challenge was identifying the core assumptions underlying an architecture. Since cognitive architectures like ACT-R do not typically formalize their core assumptions and mathematically prove properties of the architecture, it took considerable effort to bridge that gap. This challenge pushed me to find ways to apply mathematical rigor to a field that historically has not embraced it, ultimately making the testing of assumptions more systematic and precise.
Innovation and Impact
Q. Cognitive models like ACT-R are used in a variety of fields. How do you think your work could influence industries outside of traditional cognitive research, such as AI, healthcare, or defense?
The likelihood methods that my colleagues and I helped develop could have a significant influence on industries outside of traditional cognitive research, such as AI, healthcare, and defense. One of the main advantages of these methods is that they make applications easier and require far less computational power to test models and generate predictions. This opens up the potential for real-time applications, which is crucial in many of these fields.
For instance, in defense, there are numerous applications. The DoD, for example, is interested in individualized prediction models that can quickly compute future performance, workload, or how a person might react if given an additional task. These methods are designed to handle such computations efficiently, which makes them ideal for high-stakes environments where rapid, accurate predictions are critical.
Q. Your research focuses on enhancing the accuracy and precision of cognitive simulations. How do you think this will change the way we design intelligent systems or everyday technologies?
Improving the accuracy and precision of cognitive simulations will have a profound impact on how we design intelligent systems and everyday technologies like personal assistants or smart systems. Having a more accurate model of how people think and make decisions would allow for more natural, intuitive interactions between humans and these systems.
In defense, these improvements could lead to smarter, more reliable decision-support systems that help operators manage complex scenarios with greater efficiency. But the real game-changer is the likelihood methods. These are especially good at characterizing data for individuals, or even in a hierarchical model that leverages similarities between individuals to inform potential outcomes. This means systems could be personalized to predict how specific users might react in certain situations, or how groups of people with similar traits might behave, making these technologies far more adaptive and responsive to individual needs.
Looking Forward
Q. What’s next for you in terms of research? Are there any new projects or areas you’re particularly excited to explore?
The immediate direction for my research is to continue applying this method for testing core architectural assumptions. The goal is to use it to whittle down the list of cognitive architectures by systematically testing their assumptions. I've already applied it to one model—ACT-R—and was able to identify some of its limitations. Now, the next step is to compare the remaining cognitive architectures, starting with the most popular cognitive architectures, as that might have the largest immediate impact. It's an exciting process, as this could help narrow down the field and highlight which models are truly viable for future applications.
Q. How do you stay inspired and innovative in a field that is constantly evolving? Are there any trends or developments in cognitive science and modeling that you’re keeping an eye on?
To stay inspired and innovative in a constantly evolving field, I make it a point not to get stuck in a research silo. In graduate school, and academia more generally, there’s a tendency to focus on solving a single problem, but I believe it’s important to maintain a broader perspective. It’s easy to overlook potential connections when looking at a research problem too narrowly and failing to see how different fields might be interconnected.
In terms of trends, I’m always looking out for developments that encourage a more interdisciplinary approach to cognitive science and modeling, where multiple perspectives come together to tackle complex problems. Staying open to new ideas and avoiding tunnel vision is key to staying ahead in this field.
Advice and Reflection
Q. What advice would you give to young researchers or students interested in pursuing a career in cognitive modeling?
One piece of advice I’d give to young researchers or students interested in cognitive modeling is to learn good software design principles. There is an allure to hack things together well enough to get code working because it seems like you are saving time. However, in the long run, these short cuts are often shortsighted and makes it more difficult to reuse code and understand what it is doing. Taking the time to learn software design principles well will make your work more efficient in the long run.
My other piece of advice is one I have largely not followed: invest time social networking and learning how institutions work. When people think of science, they tend to imagine a person with disheveled hair sequestered in a small office feverishly scribbling equations on a whiteboard and typing manuscripts on a computer surrounded by piles of books empty coffee cups, but this is only one aspect of science (the only one I like). However, in practice, science operates more like a business (applying for grants, managing personnel, writing reports etc.) and requires lots of marketing and social networking for most people to be successful.
Q. Looking back at your career so far, what are some key lessons you’ve learned about innovation and research?
In my experience, developing novel research ideas is often an iterative and circuitous process, with failures and dead ends often encountered along the way. The end product may look very different from the initial idea you sketched out, and that is OK. One approach I sometimes use to start the process is to search for research questions where several of my interests intersect, even if they seem unrelated. I think this provides an opportunity to innovate and push boundaries.
###
About Parallax Advanced Research & The Ohio Aerospace Institute
Parallax Advanced Research is a 501(c)(3) private nonprofit research institute that tackles global challenges through strategic partnerships with government, industry, and academia. It accelerates innovation, addresses critical global issues, and develops groundbreaking ideas with its partners. With offices in Ohio and Virginia, Parallax aims to deliver new solutions and speed them to market. In 2023, Parallax and the Ohio Aerospace Institute formed a collaborative affiliation to drive innovation and technological advancements in Ohio and for the nation. The Ohio Aerospace Institute plays a pivotal role in advancing the aerospace industry in Ohio and the nation by fostering collaborations between universities, aerospace industries, and government organizations, and managing aerospace research, education, and workforce development projects. More information on both organizations can be found at Parallax and OAI websites.