Speaking of Psychology - Reading minds using brain scans, with Kenneth Norman, PhD

Episode Date: September 21, 2022

The idea of a machine that can read your thoughts sounds more like science fiction than actual science. But in recent years, it’s come closer to reality. Kenneth Norman, PhD, of Princeton University, talks about how scientists decode thoughts from patterns of brain activity, what we can learn about thinking, learning and memory from this research, how it could be useful in mental health treatment, and more. Links Kenneth Norman, PhD Speaking of Psychology Home Page Learn more about your ad choices. Visit megaphone.fm/adchoices

The idea of a machine that can read your mind sounds more like science fiction than actual science. After all, what could be more private and inaccessible to the outside world than what goes on in your own head? But in recent years, scientists have been coming closer to making this fantastical seeming idea into reality. They've developed brain scanning tools and methods that can interpret brain activity and from that activity. Decode many aspects of what people are thinking. in essence, reading their minds. So how does this technology work? How do scientists translate patterns of brain activity into thoughts?

What kinds of thoughts can they decode? How advanced are these methods? And what are the limitations? What research questions can they help scientists to answer? What practical and moral questions does this research raise? And where might it be going in the future? Welcome to Speaking of Psychology, the flagship podcast of the American Psychological Association that examines the links between

psychological science and everyday life. I'm Kim Mills. My guest today is Dr. Kenneth Norman, a professor of psychology and neuroscience and chair of the psychology department at Princeton University. In his lab, the Princeton Computational Memory Lab, he and his colleagues developed new methods to analyze brain scans. They use those methods to study learning and memory by decoding people. thoughts as they learn and remember. He has published more than 100 research papers and his work

has been funded by the National Institutes of Health and the National Science Foundation, among others. Dr. Norman also teaches an undergraduate class called fMRI decoding, reading minds using brain scans. And he has won several awards for his mentoring and teaching. Thank you for joining me today, Dr. Norman. I'm looking forward to talking about this fascinating research. I'm very happy to be here. Thanks. in the introduction that the whole topic can sound like science fiction or even a parlor game for that matter to people who aren't familiar with it. So let's start by getting some basic grounding in the science. Can you explain in a fairly non-technical way how this works?

When we say you're working on reading minds, what does that mean and what are the tools that you're using? The core idea is that different thoughts correspond to different patterns of neural firing in our brain. So if we want to be able to decode people's thoughts, we need to be able to pick up on, for example, the difference in the pattern of neural firing associated with you thinking about a strawberry or a frog or a tree or what have you. And so the way we do this is with MRI machines. And so when people talk about functional MRI, basically we're talking about tuning an MRI machine to detect brain activity. And the way we do that is by tuning it to detect levels of blood oxygen. And basically, the idea is that parts of the brain

that are more active use up more oxygen from the blood. So if we've tuned the MRI machine to detect that, we can get a sense of which parts your brain are more active than other parts of brain. Right. And so this is a very indirect measure of neural firing, right? We're not picking up on the electrical zapping of neurons, we're picking up on this kind of blood flow correlate of neural zapping. And so it's really an open question whether the signal we're getting out of this specially tuned MRI machine is going to be too blurry to detect these sort of nuanced differences in neural firing or whether it will be resolved enough to be able to do that. And so that's sort of the game that we've been playing has been trying to figure out whether this imperfect carolid neural firing can pick up on these relatively subtle neural patterns.

And to cut to the chase, the answer is it does it surprisingly well, right? And so the way that we do brain decoding is basically we ask people to think about one thing. So we can ask them to think about frogs, right? and we take a bunch of snapshots of their brain activity with the MRI machine while they're thinking about frogs, and then we could ask them to think about lizards. And we could take a bunch of snapshots of their brain activity when they think about lizards. And then we feed those snapshots, the frog snapshots and the lizard snapshots, into a computer program that has basically been optimized to try to find the differences in the patterns of brain activity,

sort of which brain areas are more or less active when you're thinking about frogs or lizards, right? That's what a pattern classifier is, right? It's a machine learning algorithm that is basically trying to find opportunistically any difference in the patterns that you feed it. And so if there's a reliable difference in the frog pattern and lizard pattern, it'll find it. So if you and I were both put into an MRI machine and we're, were told to think about frogs, would our patterns look essentially the same? They would look kind of similar, right? But not identical because you and I might have had different life experiences with frogs and lizards, right? And the way that our brain represents

thing is a function of our personal experience. But the ideas that our experiences will have been similar enough that there should be some transfer. between my brain patterns and your brain patterns. And concretely what that means is that if you train one of these pattern classifier algorithms on the frog versus lizard distinction in my brain, right, it'll do the best job at decoding my frog versus lizard thoughts, right? But it'll still possibly be above chance, right, at detecting your frog versus lizard thoughts.

The most striking example to me of similarity across people's brains was a former postdoc in my lab, who's now faculty of John Topkins. Janice Chen, when she was here at Princeton, ran a study where she scanned people while they watched the entire episode of the Benedict Cumberbatch Sherlock TV series. And then she had them just recall it in the scanner while their brains were being scanned. And she did this for a bunch of people. and what she showed, which was totally amazing to us at the time, is that you could train one of these pattern classifiers

to decode which scene of the TV episode a person was watching. So she trained it to decode based on my brain, which scene of the episode it was. And then she showed that it did an incredibly good job at transferring to other people's brains, right? So the code that's being used to represent a particular scene in this TV episode, which seems like a fairly fine-grained thing, appears to be common across people, which is not what we expected. FMRI has been around for several decades, at least since the early 90s. What's changed in recent years to make your work possible?

There are two things that could change, right? One thing is sort of the quality of the signal coming out of the machine, and the other thing is, how we analyze the data. And so both those things have changed, but the main innovations that have made this kind of thought decoding possible are on the analysis side, right? We're much more sophisticated in how we kind of chew on and analyze the data than we used to be.

Some of it is just like computers are faster and people have developed kind of better pattern classification algorithms, but part of it is we've just sort of conceptually we approach it in a different way. And so just to illustrate that, say that I want to figure out which out of all possible animals you're thinking of, right? There are a lot of different animals, right? There are hundreds or thousands, right, of different kinds of animals.

And so one way to approach that problem is I could try to train a separate thought decoder for every kind of animal, right? So I could have you think about bears and then think about not bears, right? And we try to find differences in those bear snapshots and not bear snapshots. And then we have a bear decoder. And then I could do the same thing with sharks and whales and beavers and, you know, marmots and what have you. And so you could get sort of one decoder for every kind of animal, but it's very laborious, right? We'd have to train, you know, thousands of decoders.

and no one wants to sit in the brain scanner, right, for as long as it would take to train thousands of decoders. But the alternative approach is instead of training one decoder per animal type, we can think that animals have different attributes, right? They can be big or small. They can be furry or not furry. They could live on land or water, right? They could be dangerous or not dangerous, right?

And the idea is that we could train a different decoder for each of those attributes, right? So we could sort of take brain snapshots while you're thinking of big animals and small animals, right? Or land animals or water animals, right? And the idea is that the number of dimensions along which animals vary, right, is, you know, it's more than three, right? But it's less than a thousand, right? So if you think about like the game 20 questions, right? The reason that it's 20 questions and not 100 questions, right? It's because you don't need 100 questions to figure out what someone's saying, right? And that's another way of saying. The number of dimensions along which animals vary, it's probably on the order of tens or a hundred or something. So instead of training a thousand decoders or 10,000 decoders, one for every animal, you can train 10 decoders or 100 decoders, like one for every dimension, right? And the idea is that if you have like 10 of these

dimension-specific animal decoders, right? You can ask someone to think of something, and then you feed that data into each of these dimension-specific animal coders, and I can think, oh, Kim is thinking about a big ferocious land animal that's furry and brown or something, right? And then that gives me a pretty good sense. Maybe it's a bear. Right. And so the core principle, again, the conceptual innovation is that if we come up with the right set of underlying dimensions and train a decoder for each those dimensions, then you can decode the whole space, right? I could, you know, the system that I just described to you is like a general purpose animal decoder, right? It'll tell you what the animal that Kim is thinking of, you know,

sort of where it sits along each of the relevant dimensions. And that's probably going to be enough for me to guess what animal is. Right. And people have applied that strategy to, for example, like, which of all possible nouns, right, someone is thinking of, right? And the idea there is, again, the number of dimensions you need or like, you know, questions you need to ask to pinpoint what noun. Someone's thinking of. People figured out it's maybe a couple hundred. And so you can think of any noun as sitting as like sort of a point in a couple hundred dimensional space. And so we can look at your brain activity and figure out where your thoughts are in that couple hundred dimensional space. And that gets us very close to figuring out what noun you're

thinking of. What are some of the practical uses that researchers are exploring for this? One that's gotten a lot of media attention is using it to communicate with locked in patients, people who are not able to communicate with the outside world, but may still be conscious and thinking. Is that research still ongoing? And are there other practical uses people are interested in? My colleagues, Martin Monti, Adrian Owen, several others have worked in that particular situation. The very clever strategy they came up with was if you have a locked-in patient, they tried to come up with mental activities that they thought would activate. very different parts of the brain, right? So it is that the part of the brain that's activated by

thinking about like playing tennis, right, and all the movements involved in playing tennis is very different from the part of the brain that's activated by thinking about the layout of your house, right? And so it is that they wanted to come up with instructions that would elicit thoughts that were really different from each other, right? And then they would ask one of these locked-in patients to, like, imagine playing tennis or imagine the layout of your house, and they would see these differences, right? Basically, the playing tennis pattern would come to life when this person who can't move or speak was asked to think about tennis and vice versa for thinking about the layout of the house. So that's the strategy they've used there. And they've used that to very good

effect and to show that people, you know, they didn't know whether these people had the ability to follow instructions or do things like that, and they showed they could, which is obviously incredibly important. That's a very far cry from having one of those patients, for example, be able to type with their brain. But the idea is this is sort of a general theme of how we've made progress in the field is that you don't need to do extremely fine-grained to code. to be able to get insight into what people are thinking. So you can sort of set up a scenario where these very sort of crude differences are actually informative about what's going on.

And so another example of something that we're trying to do in the practical domain that leverages these sorts of crude differences is neurofeedback. right. So the idea there is that we can take someone while they're in the scanner, right, and try to decode their thoughts and then use that to train them to do something better, right? And so an example of this that I really like is Megan DeBettencourt, who was a grad student here at Princeton in my lab, was very interested in how to train people who do a better job of paying attention, right? And so the idea is you're doing something. boring task and your thoughts drift, right? Or you're a truck driver, right? And you're spacing out, right? And that could be very dangerous, right? And you might space out and then find yourself

like in a ditch, right? Because you weren't paying attention to the road because you got tired. So it'd be really good to find a way to train people to do a better job of paying attention or at the very least sort of notice when they're starting to space out. Right. And so the task Megan design involved having to do this sort of very boring button pressing task where we showed them a display where there were faces on the screen and they were also sort of in a ghostly kind of way superimposed on those faces pictures of scenes. So people are looking at these composite displays of faces and scenes and we would tell them like just pay attention to the faces, right? And press a button whenever you see a female face and ignore the scenes. And we'd have people

do this boring button pressing task, right? And so Megan would be analyzing their brain activity and using one of these decoding algorithms, she could figure out the moment that their attention started to drift to the scenes that they were supposed to be ignoring. And the reason she used faces and scenes is because we knew, just like playing tennis and thinking about the layout of your house, faces and scenes elicit really distinct patterns of neural activity. And that gives us like a very very high degree of sensitivity to the exact moment where people started to process the scenes that they were starting to ignore. And then because we're doing this decoding in the moment, right, in real time while people are in the scanner, we could change the display the moment that

we detected this in attentiveness. And what Megan decided to do very cleverly is that the moment she detected with this real-time brain analysis that people were starting to, to attend to the scenes, she made the faces less visible. She made the task harder. And so the idea there is she's trying to sort of amplify the attentional lapse. So the moment that your brain activity starts to float toward what you're not supposed to be doing, she made it really salient to people that they were messing up by making the tasks they were supposed to be doing really hard.

So the idea is that if we do that, we're going to get people to notice, before they would have otherwise, right, that they're spacing out, right? And the hope there is that we would make them more sensitive, right, to these attentional lapses. And they'd sort of learn to detect that better and to be less likely to have their attention sort of float all the way off in the future. And she demonstrated in a nature neuroscience paper that was published in 2015 that training people in this kind of closed loop setup where you amplify their attentional lapses makes them better able. to sustain their attention over time, you know, which was a big advance. And I'll just mention as a sort of way of building on that, we started to run studies. This is my grad student, Ann Mennon, and we're collaborating with a vet Shalene, who's a depression researcher at UPenn,

to sort of apply this same technique to help people with depression learn to sort of unstick their thoughts from sad mental states, right? So one of the most salient symptoms of depression is once a depressed person starts thinking about something sad, they have a hard time unsticking their thoughts, right, from this sad mental state. They sort of ruminate and ruminate on these negative things. And so the modification we made to Megan's task is very simple, right? It's basically the same task I just described, except we made the faces sad, right? And then we told people to attend to these pictures of emotionally neutral scenes and make simple judgments about the scenes, like is it an indoor scene or an outdoor scene, right?

But then we just told them ignore the faces. But what we see when we put depressed people in the scanner and have them do this task is that their thoughts start to drift toward the sad faces. And once that happens, they get sort of stuck on the sad face. And so what Anne Mennon did in this study is basically the second that we detected with these brain decoding algorithms that they were attending to the sad faces, we made the sad faces really visible. And so again, the idea is to make it really salient that their thoughts had sort of rolled away from the scenes toward the sad faces with the goal of making them more sensitive to the sort of moment when this lapse is happening. with the idea that they could use that to get better at catching themselves before, you know, their mental state has gotten all the way into this sort of pit of sad thoughts that it's

hard for them to get out of. And so we're running experiments now to see if that training process basically helps them unstick. So it sounds like it's becoming both a diagnostic and a therapeutic tool in a sense. And I'm wondering, then are you moving toward, say, identifying people with maybe disordered thinking or violent thoughts and then being able to maybe replace those thoughts with healthier concepts? Yeah. I mean, two things there.

One of them is that diagnosis is tricky just because, you know, I mean, this is something, again, I mainly do basic research on learning and memory. so I'm getting a little bit out of my wheelhouse now, but I think that, you know, I can say with some degree of confidence that one of the changes in how people think about mental health and different conditions, right, is that they're very complicated, right? Like, depression is not just one thing and it overlaps a lot with anxiety and overlaps with a lot of other disorders, right? And so the idea is that the people who have depression, right, that's diagnosed in some way, it's a heterogeneous group. and it overlaps with a lot of other groups, and that makes it sort of hard. People are working very hard to do this diagnosis, right?

But it makes it, that's one of the most challenging things to do with brain data, right? Is diagnosis. And so it doesn't mean it's impossible. It just means that it's a hard problem, right, that people are working on. But I think that this space that neurofeedback belongs to of therapy, or just promoting learning more broadly, right, where you've got a group and they've got some particular way of thinking, right, like in depression, these sort of sticky negative thoughts and you want to help them learn to control these negative thoughts or to be able to pull themselves

away from that. Then I think that these sorts of brain decoding tools are going to be very, very useful. I think they give us a really powerful window into how a person is thinking in a particular moment. And in ways like this neurofeedback setup that I just described, we can try to choreograph experiences for them that will help them learn to do things differently. And, you know, this applies both to clinical populations, but also like education, right? The idea is that, you know, what it means to learn a subject in the course you're taking is you learn to organize your thoughts, right, in a way that adheres to the kind of ground truth of how things are, sort of getting scrambled up.

But to concretize this, like say you're taking a course in computer science, You don't know anything about computer science, right? And what it means to learn computer science is basically to learn sort of which concepts go together and which concepts don't. So you learn that if then statements, right, are a way of sort of controlling the flow of sort of what happens in a project. And four loops and while loops are also a way of controlling the flow. But, you know, variables are something different, right? And so there are all these new terms. You have no idea what you mean.

And you learn sort of like these terms, like if then and wall and four loops, go together. And these other terms don't. Right. You sort of learn what coheres and what's different, right? Or if you're learning about animals, right? You learn which animals, right, are dangerous and which animals aren't. The idea is that we can use these different brain decoding measures to sort of get a window. into what concepts people think are similar,

what things people think at a particular moment go together, and which concepts people think don't go together. Taking this computer science example, again, if we've got two concepts that go together and we look at some computer science students' brain and we see that the pattern of brain activity associated with those concepts are really different, then we know there's some learning to do, right? We want to, you know, devote extra effort to

helping them understand that these things go together. And so we can give them some, you know, lesson on how these things go together. And we could see whether that lesson is successful by scanning their brain after the lesson to see if those concepts that should go together elicit similar patterns of brain activity. So you don't have to take tests anymore, right? Your professor puts you in an fMRI and says, aha, you didn't really learn this. Right. So the reason that I was using this computer science example is that a postdoc in my lab, who now works at the computer company, Snap, Mayor Mechelam did exactly the study. So he took computer science students at Princeton who are taking introductory computer science,

and he scanned them multiple times over the course of the semester, and he looked at the brain patterns evoked by different computer science concepts and sort of how similar or different they were. And basically what he showed is that he could figure out how well a student knew a concept by basically comparing the similarity of patterns in the student to the similarity of patterns in the teaching assistant. And if this student had similar patterns when the teaching assistant had similar patterns and different patterns when the teaching assistant had different patterns of brain, that predicted that they would do well on tests of those concepts. Right, right.

So what you just said, you know, you could use a brain scanner instead of the test is true. Of course, it's, you know, cumbersome and much more expensive, right? And it, you know, is much more cost effective to just give the person a paper and pencil test. I think that the point I wanted to make is that there are a lot of circumstances with brain decoding where, you know, it would just be easier to ask the person to say what they're thinking, right? Or to give them a normal exam question, right? rather than trying to do this fancy brain decoding thing. And I think that the situation in which these brain decoding methods really shine is, you know, there are a lot of scenarios where they're interesting things going on,

and it's not feasible to just ask the person what they're thinking. Like, for example, we know from hundreds of studies that there are really important things having to do with learning that happen when you're asleep. So studies have shown that if people learn, you know, they study something for a test and then they sleep and then they wake up, right? They actually like forget less, right? If they had certain kinds of sleep during the interval between studying and testing. So something is happening, right, when you're asleep, right? Your memories are getting strengthened or maybe they're getting reshaped. And, you're, you know, It's this incredible, deep, sort of cool puzzle to try to understand what exactly is happening.

But you can't ask someone what they're thinking about when they're asleep. Right. Because they're asleep, right? You know, maybe you can wake them up, right? And ask them, like, what were you dreaming about? Right? And sometimes people can tell you, but you only get the sort of fading whips of people's thoughts. But, you know, one really cool application.

for brain scanning might be to decode what people are thinking when they're asleep, right? So if the idea is, you know, and people who built theories of what might be happening during sleep is basically that your brain is composing for you a kind of playlist, right, of things that it thinks it's important for you to learn about, right? So the idea is when you're awake, stuff happens and some experiences get marked as being important, right? And they get put in this playlist. And then your brain sort of loops through them when you're asleep. And as a result of this looping through, these things are marked as being important, you learn more about these things. Right.

And we want to know what's on the playlist. So that's a great example of something that we can do with brain decoding that you can't do, just with asking. And so Anna Shapiro, who's a former grad student in my lab, who's now a professor at UPenn, had a brain imaging study. It actually, I guess in this study, she wasn't looking at when people were asleep. She was looking at sort of what people think about when they're just kind of spacing out, right? The idea is this sort of looping through concepts that were marked as important happens when you're spacing out when you're awake in addition to when you're asleep. And she wanted to get some insight into like what's on the playlist, right, of things that you think about when you're

spacing out. And how does that relate to the learning experience you just had? And she used brain decoding and got this really cool result. It sort of makes sense that the concepts that people struggled with the most when they were trying to learn this new thing that she was teaching them about were the ones that appeared the most on their sort of mental playlist when they were spacing out, right? Which is very adaptive, right? You don't want to spend your time thinking about stuff you already know. You want to spend your time working through the stuff that you don't know very well, And she was able to get a very tangible clear evidence for this idea using brain decoding. Let me ask you a long-term question, which is whether the goal is some kind of a universal thought decoder so that there'll be an absolute lexicon of what we know thoughts consist of.

And is this science fiction or is this something that's in the realm of the possibility? It's sort of like decoding the human genome. I mean, right. It's a very exciting possibility. And I think we're very close to having some kind of a universal thought decoding. And, you know, it uses the principle I described earlier, which is that, you know, we think that thoughts lie in these, we call sort of low dimensional spaces, which is just like saying that you can ask, you know, a couple hundred questions to sort of zoom in. on what particular thing people are thinking about.

And you can make a decoder, you know, for each one of those questions. And then we can do that now. And that's a universal thought decoder. But it doesn't work anywhere near perfectly. It's like a very blurry thought decoder. We can tell where your thoughts fit in like this 300-dimensional space. But there's like a huge cloud of uncertainty around our estimate of what you're thinking about, right? And with functional MRI, that has to do with just intrinsic limits on the resolution of the technique.

I said earlier, right, we're not measuring the zapping of all the neurons in your brain. We're measuring this blood flow thing that's loosely coupled, right, to the neural zapping. It's both blurry in space, right? So we can't tell exactly which neurons are activating, right? Just sort of which millimeter by millimeter by millimeter cubes of your brain are most active. And it's also blurry in time, right? This blood flow thing that we're measuring unfolds slowly relative to the very precise zapping of neurons. Right.

And so that blur is not something we're going to be able to fix. you know, my favorite example of this when I teach is there was, I think, in like, 2008 or somewhere around then, this Newsweek headline that's like, mind reading is possible, right? Which sounds terrifying. And then the subheading is like, people can now tell with 70% accuracy, whether you're thinking about pliers or a wrench. Right. And like, that's cool. But that's like not the mental picture that comes to mind when you read mind reading is now possible. And, you know, I think that should be kind of reassuring to people who are worrying, for example, about abuses of this sort of technique. And if I had like one practical point to make, right, is that there's just

this enormous gap between perfect decoding and. above chance decoding, right? And so when we publish papers in scientific journals saying that we're doing mind reading, what I mean is above chance decoding? We have some insight, right? But for example, if you, everyone I think now is using these speech to text things on their phone. Yeah. You talk to your phone and it transcribes it. And the idea is that if the speech to text thing is wrong, even like one every 15 seconds, right, which would be like 99% accuracy in the words it's transcribing. It's super annoying. And that goes back to what I was saying.

If, you know, people are thinking, when can we have like people who are locked in just kind of type their thoughts with FMRI? And, you know, I think people could and are actually working on developing successors to FMRI, right? that might be more precise, right, less noisy. But with current techniques, we're not even close to, you know, 90% thought decoding accuracy or, you know, we're above chance. So I think, you know, in some, we already have a universal thought decoder. It just doesn't work super well. If people tried to use it as like a product, they'd be really annoyed. Well, plus, then you need to have an MRI machine, right, which is hundreds of thousands of dollars and it weighs a lot and you're not going to have one in your living room like a television set.

That's right. But there are other techniques for non-invasively measuring brain activity like there's a thing called EEG, right, which is electro-ensopulography, which is measuring electrical fluctuations on the scalp, which is even blurrier than MRI, but it's much cheaper. right? And so a lot of the commercially available brain computer interface headsets use EEG. Right. And again, the idea is that you're not going to be able to type with your brain with EEG, right? But it's very feasible with EEG to detect like how attentive you are. Like are you in a focused or unfocused attentional state or even sort of relatively crude. brain decoding things like, are you thinking about a face or a house? You know, and people in my lab have used EG. So Nicole Rafidi, who was a undergraduate in my lab way back when, used EG to sort of measure out how much competition there was between different memories, right? So sort of how hard people were working to try to remember foreign language vocabulary. And the idea is that that's very useful

because that tracks pretty well how well you've learned something. So the idea is that once you've learned something, your brain zips right to it, right? And wrong things don't come to mind. But when you're still kind of uncertain and you're just beginning to learn, for example, a foreign language, lots of wrong things do come to mind. Right. So if we have have this sort of cheap and easy eG correlate, right, of the extent to which wrong things are coming to mind, that tells you how well you've learned something. And the idea is that there could be maybe some closed loop flash card system that uses this EEG correlate of how quickly and easily the memory is coming to mind to sort of know which flashcards to show you. So if it detects

there's a lot of competition from this EEG signal, then it'll keep showing you that flashcard. But if it sees that your brain is zipping to the right thing, which we can do with the EEG, then it'll know that that memory is, you know, it's cooked, right? You don't need to show that flash card anymore. So that's another important lesson is that there's a lot you can do with coarse-grained brain decoding. And the depression neurofeedback thing I mentioned earlier, that's, you. using coarse-grained brain decoding, right? Are you attending to the phase or the scene? But we think

we can do really powerful things with that simple distinction. And so I think that we've been playing with brain decoding for a couple decades now. And I think at the beginning we had this idea of like anything's possible and we'll just see. We'll use all of our fancy machine learning decoding algorithms and we'll see what we can decode. And now I think with appropriate humility, We know the kind of limits on the technique, and we can try to tailor the applications of the technique to the limits that we understand. Well, Dr. Norman, I want to thank you for joining me today and telling me about your amazing research. It's really quite fascinating. Thank you. All right. You're very welcome. It was fun talking with you.

You can find previous episodes of Speaking of Psychology on our website at www. www. speakingof psychology.org or on Apple, Stitcher, or wherever you get your podcasts. If you have comments or ideas for future podcasts, you can email us at speaking of psychology at APA.org. Speaking of psychology is produced by Lee Weinemann. Our sound editor is Chris Kondyne. Thank you for listening. For the American Psychological Association, I'm Kim Mills.