This episode is part of “The Young American Scientists,” an editorially independent project that was produced with financial support from Regeneron.
Rachel Feltman: For Scientific American’s Science Quickly, I’m Rachel Feltman.
Today we’re back with another one of SciAm’s 2026 Young American Scientists. This group of groundbreaking researchers represent the future of science, technology and medicine. One of those honorees is Kauê Machado Costa, an assistant professor at the department of psychology at the University of Alabama at Birmingham. He studies how learning works on a neurological level.
On supporting science journalism
If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
Feltman: Thanks so much for coming on to chat with us today.
Kauê Machado Costa: It is my great pleasure.
Feltman: Before we get into the subject of your research specifically, I have to ask if you would tell me why you describe yourself as slightly cursed?
Costa: Over the course of my career, there have been only very few instances where I have made a prediction, right? I have a hypothesis, I made a prediction, and my prediction actually turned out to be correct. Almost every single time when I make a prediction and I do the experiments, I get the opposite results or something that was entirely unexpected, not even in the realm of my imagination at the time.
That’s why I facetiously called it a curse, but actually it’s also a blessing, because it means that my career has been very exciting, at least from my perspective. There’s always something new at the end of every experiment.
Feltman: Yeah. Well, and I think that’s so interesting because some laypeople who are, you know, maybe familiar with scientific findings but not so much the process of the scientific method might hear that and think, “Eh, well, he must not think he’s very good at doing science.” But it sounds like it actually facilitates you being really excellent at doing science. Could you tell us more about how it informs your approach to research?
Costa: Of course. So one of the ways that it strongly affects my approach to research, and that I tell everyone in my lab, is that you should always start a project, start an idea with a hypothesis, with a prediction. And that’s very important because if your prediction is wrong, you still have an informative result, right?
Many of the times you can start a project with purely exploratory aims, and there are some times where that is very much warranted and necessary. You will not know when you’ve actually supported or refuted a particular hypothesis. So having a very strong, very specific prediction at the onset of your experiment really, really helps you achieve not only a publishable but really an informative, a significant result.
I mean, for me, if I didn’t have that, most of my projects would have ended in dismal failure. I mean, most of them still do, in a way. There’s always, you know, experiments that fail for multiple reasons. That’s part of the scientific process. But it would’ve been even worse, let’s put it that way.
Feltman: Yeah. Well, in one of the instances where things didn’t go quite right—I would love to hear more about this—you actually uncovered some issues with a very common model species that perhaps other scientists should be aware of.
Costa: Yes. This was a very interesting case toward, like, the last quarter of my Ph.D. So I was using this very common, probably one of the most used transgenic lines in neuroscience research, DAT-Cre mice. These mice, they allow us to do genetic manipulation specifically in dopamine neurons, so in neurons that express this protein called DAT, the dopamine transporter.
So they’re standardly used all across my field. Now, I was testing the effect of trying to knock out this particular gene. And when I ran all of my mice—and I was blind to it, right? So this is part of the rigor. You’re blind to the genotype when you’re running the experiments. When I was uncovered, and I could actually analyze the data, I noticed that my control mice, they were acting kind of funny. They were acting weird, and that was severely affecting the conclusions that I was trying to make, and that was very puzzling.
So at this point, you know, many people could say, “Well, this is very strange. I am going to switch. I’m going to ignore this. I’m going to do something else.” But I thought that that could actually mean something important. My Ph.D. adviser at the time was happy to indulge in my strange obsession, so I went, and I dug deeper, and I really tried to investigate what were the mechanisms of what I was seeing. And what I ended up discovering was that this widely used mouse line, at least in the kind of substrains and versions that we were using, they had a very particular, very strong sex-dependent phenotype, because the native expression of dopamine transporter was reduced, was impaired.
So essentially, these mice, and especially the females of these mice, they were a model of ADHD. So they were hyperactive. They had lower DAT function. And what we wrote up in the paper—which actually was a one of the top 100 neuroscience papers downloaded in the journal that year—what we say is that “well, you really need to look into that if you’re trying to see behavioral effects or any kind of effect of a particular genetic manipulation, because just that strain already has this background phenotype.”
Feltman: Very cool. So let’s talk about your work on learning. What was kind of the prevailing school of thought on how we learn, and what have you uncovered in your research?
Costa: So at the risk of using a very gross oversimplification, you can basically divide the ideas about learning or the frameworks of learning into two big groups. One big group, the idea is that when you learn about something, when you learn about an action or a cue, some event that you observe in the world, what you are learning is fundamentally how good or bad that event is, right? So the value of a particular cue of a particular action. This means “How motivationally relevant is the outcome that is associated with that particular cue or particular action?” And in this big view, which in the current computational lingo is often referred [to] as model-free learning, you don’t have to have a detailed representation of the world, all you need is to learn how good or bad something is. To update what is called a value function. So that’s one view of it.
Another view is that when you explore the world, what you do is you create a representation, a simulation of the external world in your mind, right? And you learn how individual events are associated with each other. You can estimate the probability that something will happen based on what you just observed.
And in this type of learning, which in the modern lingo is often called model-based learning, you create a rich representation of the world. And this has multiple advantages in the sense that you can use these representations to make inferences about things that actually have not happened based on what you have experienced before, while in the model-free type of learning, you can only really learn, you can only make decisions based on the actual experiences that you have had. On the other hand, model-free learning is much more simpler to implement, while model-based learning requires a lot more energy, a lot more computational resources.
Feltman: Got it. So model-based learning is essentially building an internal simulation of how reality works and model-free learning is pretty much just trial and error, where you’re keeping track of what works and what doesn’t.
So where does your work come in? You know, do your findings support that two-system framework, or are things more complicated than that? Costa: Now, in my work related to that field, there’s two findings that I like to highlight where I recently showed that dopamine prediction errors or dopamine signals, which have long been associated with model-free learning—so it’s thought to represent the difference between predicted and expected reward value, so something that is very much related to reward. I had experiments where I actually showed that dopamine signals, they actually represent prediction errors about things that don’t have reward value, and in that sense, these signals, they approximate much more a model-based prediction error than a model-free prediction error.
In another study, I investigated the role of an area called the orbitofrontal cortex, a very interesting part of [the] frontal cortex. Now in humans, it’s right here above the eyes, and this area had long been associated with model-based learning. All right? So there’s multiple studies showing that neuronal activity in this area represents kind of like these associations between elements of the world, especially those that are related to the execution of tasks. There has been a rather influential hypothesis that this area kind of stores this type of cognitive map presentations for the execution of tasks.
So my initial hypothesis was that if you inactivated the orbitofrontal cortex, this OFC area, that all model-based learning would be disrupted, and there would be kind of a default to a model-free learning system, right? Assuming that you have the two parallel systems in the brain, and when one can’t work, you basically default to this other strategy. What I found was that not only did model-based learning get kind of disrupted when you inactivated the OFC, but that this effect was actually very specific, meaning that the rats, they could still create a model of the world, but this model was confused.
So essentially they built a confused, imprecise model. And so from there came our proposal that what the orbitofrontal cortex is doing is not necessarily mediating all of model-based learning, but that it’s particularly important for linking specific events to each other, that when this area is not working right, you end up having confused, degraded or imprecise models.
And I like to imagine that a lot of behavior can be explained or some things that we consider maladaptive behavior, including disease states, they are better explained not necessarily from, like, this opposition between model-free and model-based learning, but rather they could be the use of models of the world that are differentially structured, right? That you can imagine that we’re all working based on our own interpretations or representations of the world but that someone may have a representation that is very detailed, very accurate, very adapted to the task they’re performing, while someone else may be having a representation that is not as precise, that is actually confusing different associations, forming a rather distorted view of the world.
Feltman: And what would the potential implications be, both, you know, in terms of just sort of understanding human behavior but, you know, also in potential applications?
Costa: So there’s a lot of possibilities there. If I start, maybe, with the most fundamental implication, is that maybe this dichotomy between model-based and model-free is a bit overblown. So maybe instead of thinking of model-based versus model-free, we’re thinking of models of different complexity that can be deployed or differentially recruit different brain areas. I think this could have a lot of implications for understanding mental illness, right?
So, for example, in another project, another study that I published, I also showed that the orbitofrontal cortex, this area, was very important for—essential, actually—for a process called latent inhibition. And latent inhibition is a measure of attentional filtering, basically, “How do you ignore information that is irrelevant?” So we show the orbitofrontal cortex is important for that, and people in—with schizophrenia, they have notorious deficits in latent inhibition, right? And the idea there is that they don’t filter out information efficiently, so they basically attribute relevance to almost everything that they see, so they form spurious associations that eventually lead to hallucinations and cognitive disorder.
So what if we had computational tools or at least, like, a general framework in which we could interpret the hallucinations and kind of the cognitive deficits that we see in schizophrenia based not necessarily on a general dysfunction of model-based learning but actually as the creation of a disordered model that has its own particular structure? We might be better able to pinpoint what are the cognitive processes that go awry in mental illness. So this I think is important for computational psychiatry.
Another potential example is in the neuroscience of addiction, substance use disorder. So if you think that—there’s a general theory that, uh, thankfully I think has fallen out of fashion, that as a consequence of substance use disorder, you have also this transition from a model-based strategy to a model-free strategy, right? From someone that forms and uses a detailed representation of the world versus a over prioritization of rewards and values.
But you can think that actually instead of having this transition from model-based to model-free, what you have after substances or drug abuse is the creation of a disordered model. And you can think about, uh, the importance of that because it explains, for example, the success of behavioral strategies like contingency management in the treatment of drug abuse, right? And which is much harder to explain if you think that you have this, let’s say, an overdominance of model-free strategies. These are controversial topics that I’m sure a lot of my colleagues would disagree [on], but it’s where my thoughts are leading based on my previous work.
Feltman: Very cool. And my last question is just, you know, what are some questions that you’re really excited about answering in your field?
Costa: So one of the things that we are doing in my lab quite a lot is trying to understand what is the informational content of those dopamine teaching signals that I talked about before.
So if it is indeed the case that dopamine signals are carrying information that is beyond, you know, reward prediction errors or beyond rewards, and it’s doing something that is really more akin to a model-based prediction error, then what are the dimensions of information that contribute to these signals?
Another question that I’m very interested in—and this relates to a recent work that I published with Zhewei Zhang, a fellow postdoc at the NIH [National Institutes of Health], where I was working with Geoffrey Schoenbaum, my postdoc advisor—so we found out that if you record another neuromodulator, acetylcholine, together with dopamine, that their interactions, they vary a lot depending on whether dopamine seems to be responding to something that is about reward versus processes related to motivation. So maybe part of the secret to this difference in information encoding comes from the interaction of different neuromodulators.
So I’m also very interested in how different neuromodulators interact in learning to try to see if that increases information capacity. And then, as I mentioned before, thinking about the orbitofrontal cortex, I am interested in trying to tease out “What are the environmental properties or the conditions in which you create more detailed or simpler or more precise versus more generalized models of the world?”
And I think, finally, I am very interested in how all of these, you know, rather abstract computations are actually enacted by individual neurons.
So while it may seem like a lot of aims, I hope it’s also clear that they all center around a general question, which is trying to understand: What are the mechanisms, both informational and cellular and molecular, that determine what do we learn? Like, what exactly do we incorporate into the brain, into our mind, when we are exposed to different events in the world?
Feltman: That’s all for today’s episode. We’ll be back on Friday with one more Young American Scientist special—this one all about surprising new questions in cancer research.
For more on this year’s Young American Scientists, don’t forget to check out the latest issue of Scientific American. You can also head over to our YouTube channel to see video profiles of some of our winners.
Science Quickly is produced by me, Rachel Feltman, along with Fonda Mwangi, Sushmita Pathak and Jeff DelViscio. This episode was edited by Alex Sugiura. Marielle Issa and Aaron Shattuck fact-check our show. Our theme music was composed by Dominic Smith. Subscribe to Scientific American for more up-to-date and in-depth science news.
For Scientific American, this is Rachel Feltman. See you next time!
