Q&A with Anne Collins
If there is one thing the past year of the COVID-19 pandemic has taught us, it’s that people learn and adapt to new situations all the time. We are never really starting from scratch.
“We have a whole set of strategies that we rely on to adapt quickly compared to all other non-human animals,” says Anne Collins of the University of California, Berkeley. “And the area of reinforcement learning to model and understand some of these strategies and behaviors has been very helpful, but it’s also very limited.”
Collins wants to better understand how it is that people can so quickly and flexibly transfer knowledge to adapt to new situations. “We know it’s really hard because people in AI still struggle with getting artificial agents to do this human-like behavior,” she says. Her work couples controlled lab experiments with new computational models, to better capture and understand this unique human learning.
A co-recipient of the Young Investigator Award, Collins will discuss this innovative work at the CNS 2021 virtual meeting this March. I spoke with Collins about her research and its potential applications, including for AI, her thoughts on quantitative skills, and how her work has been affected by the pandemic.
CNS: How did you first become interested in cognitive neuroscience?
Collins: I absolutely have not had a classic trajectory. I did my undergrad degree in France, and I started in math and engineering there. I actually didn’t know about the fields of cognitive sciences and cognitive neuroscience until the end of my undergraduate degree. I discovered them at my school when a class was offered as a broadening topic. I had always been fascinated by the mind but that’s when I discovered it was possible to study the mind in a scientific and quantitative manner. I went on to do a master’s degree in cognitive sciences and saw that my quantitative skills were very helpful.
CNS: Many people don’t associate quantitative skills with cognitive neuroscience. How are quantitative skills important to you?
Collins: Yes, I know people might not see those skills as a core; they might think of psychology more as a core. I think there are 2 ways in which quantitative skills are essential. The first, which many people are aware of, is statistics. In cognitive neuroscience, we handle big or complex data and having the quantitative skills to analyze them with more precisions and depth is essential. The probably more important second point is related to computational modeling — models that allow us to put theories into quantitative terms and that allow us to provide a more mechanistic understanding for phenomena, as well as to make precise predictions. That’s the part of cognitive science that relates to AI. There is plenty of good cognitive neuroscience that is done without those quantitative skills but I think we can go much further when we apply those skills.
CNS: What will you be presenting in your Young Investigator Award talk?
Collins: I’ll be talking about the role of executive function in learning. What I have been uncovering more and more is that there are many different ways in which executive function plays a role, and they tend to be hidden when we do traditional modeling.
Learning is often modeled with very simple reinforcement learning algorithms; they are flexible and do a good job of capturing behavior but they also hide the contributions of all the systems. For example, they can hide what executive function tells us about what matters in our environment that we should learn about — what is relevant in the environment. Usually when we use reinforcement learning models, we take a lot for granted about how we define the state of the environment and what constitutes a choice, and that’s something I have been trying to study. And what I am most excited about recently is how executive function contributes to what we think is rewarding, which is another hidden part of the traditional models.
Usually when we use reinforcement learning models, we take a lot for granted about how we define the state of the environment and what constitutes a choice, and that’s something I have been trying to study.
CNS: Can you talk a little bit about how your work relates to real-world learning and rewards?
Collins: In reinforcement learning, we often say that something that is rewarding will lead you to repeat the previous choice, like training a little kid to do something by rewarding them with chocolate. Those food-type rewards are primary reinforcers. And there are also secondary reinforcers we work with in the lab, like points, money, a smiley face, etc., things that we have learned to associate with something positive with the same effect as primary reinforcers in the brain. But in your everyday life, how often do you get chocolates or points or even money for doing something?
What’s much more frequent is that when you are learning to do something new, you set yourself a goal and if you reach that goal, you internally reward yourself. For example if you have never played a game of sudoku, you have a goal to arrange the numbers in a very specific way and then you finally manage to do it and you have a feeling of internal reward. And that’s from a very random, abstract goal that has no reward in and of itself. That’s an example of things that happen all the time in everyday life. [See our Q&A with Amitai Shenhav about his related work.]
CNS: What are the potential applications of your work?
Collins: I consider myself a fundamental researcher first and foremost, and my focus is on cognition to understand the mental processes. I think it’s important to take what information we can from neural processes to inform mental processes. That said, there are tons of downstream applications. For AI, for example, we can learn from how humans learn fast; the models we come up with to explain that can help develop better artificial agents, and that’s something that is happening now. I think there are also applications in terms of better understanding individual differences, be that in a typical or developing or neuroatypical population. If we better understand how people learn, we can then try to optimize it.
CNS: How has your research been affected by the pandemic?
Collins: The bread and butter for my lab is behavioral experiments and computational modeling, and imaging data is secondary to that. We happened to have just collected a new set of EEG and fMRI data before the initial shutdown. Since then, we’ve had to shift to online data collection, through online learning games, but it’s been mostly an issue of scaling it up. So in principle, the impact of the pandemic has been fairly minimal on the research itself. That is not to say it hasn’t had a huge impact on our work, however. With two small kids at home, it’s had a huge impact on me and then downstream on the lab, and I am just one example. And those impacts are important to acknowledge.
CNS: What have you learned in your career so far that could be helpful to early-career cognitive neuroscientists?
Collins: Two things: One is not to underestimate the importance of quantitative skills; they really get us much further.
The other one is not science related: A decent work-life balance is essential. I’m French, and I believe that you feel better if you have a life outside of work and take breaks. We all do better work if we’re balanced and rested.
-Lisa M.P. Munoz