Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
The Philosophy Behind AI and the Mind | Columbia Professor John Morrison
11 Plays1 month ago
In this episode of The Healthcare Theory, we sit down with Dr. John Morrison, a professor at Columbia University and the architect behind its Cognitive Science program.
He shares how his background in philosophy led him to explore the mind through neuroscience, artificial intelligence, and computational modeling. We dive into his research on perceptual confidence and how both brains and machines may operate probabilistically, from perception to decision-making. He walks us through his work using transfer learning to uncover which algorithms neural networks use to process information. We also discuss the unique role philosophers can play in shaping emerging sciences like neuroscience. Whether you're curious about cognition, AI, or the philosophy of mind, this episode unpacks the deep questions driving this interdisciplinary field.
Recommended
Transcript
Introduction to the Podcast
00:00:00
Speaker
Welcome to the Healthcare Theory Podcast. I'm your host, Nikhil Reddy, and every week we interview the entrepreneurs and thought leaders behind the future of healthcare care to see what's gone wrong with our system and how we can fix it.
Meet Dr. John Morrison
00:00:14
Speaker
Today we're speaking with Dr. John Morrison, who is a professor of philosophy at Columbia and at the Mind-Brain Behavior Institute. He led the creation of Columbia's Cognitive Science Program, and we're super excited to have you on, Dr. Morrison. Thanks for coming on.
00:00:27
Speaker
Thanks for having me.
Philosophy to Cognitive Science Transition
00:00:29
Speaker
Of course. And before we get into your work and what you're doing today, you have a pretty unique background blending philosophy and neuroscience and cognitive science all into basically one career. um What kind of drew you into those fields in the first place?
00:00:41
Speaker
Well, I started off as a philosopher. That's what PhD is. And when I was doing philosophical research, it became more apparent that results from neuroscience and cognitive science were relevant.
00:00:53
Speaker
And so I spent some time learning them. And when I got tenure, I pivoted more um, uh, significantly in the direction of ah cognitive science and neuroscience. So I started spending my time at the Zuckerman Institute and developing collaborations with people in the science.
Research Focus: Cognition and Computation
00:01:11
Speaker
I mean, the general big question that interests me is, is how the mind in particular, how cognition and thought emerge from brain. Um, and also how thought can be modeled.
00:01:23
Speaker
So in cognitive science, we model thought in a very computational way. We think about the mind as a kind of computer. And then in cognitive neuroscience, we try to think about how those computations are realized or implemented um in the brain itself.
00:01:36
Speaker
So those are the sets of issues that I'm interested in. And ah that's what made my research more interdisciplinary. Yeah. And like tying back to your story, I mean, there's probably... Was there like a certain point in time or a certain event that happened that really made you realize that you want to kind of gain knowledge from these different fields, like a computational approach, a more neuroscience based approach? Like, is that something that happened over time or was there a specific moment that really sparked that um interest when you're on tenure?
00:02:03
Speaker
It was gradual. I started going to more and more talks and getting more and more interested in what cognitive neuroscience could show. i mean, in the last decade or so, neuroscience has really matured a lot and we have a ah many more examples of computations being realized in the brain.
Building Columbia's Cognitive Science Program
00:02:20
Speaker
And I guess in my interest in computational modeling of the mind itself was a longstanding interest. um And so those those two things combined.
00:02:32
Speaker
Of course. Yeah, that makes sense. um And of course, like this is kind of a new field. Like I think only really recently are people starting to realize that like understanding the mind, the brain requires the intersection of different fields from like anthropology to philosophy and neuroscience, computation. There's so much that goes into that.
00:02:49
Speaker
um For you, like, I mean, this is maybe a bit of an, and was this like an untraditional path for you? Was this like a natural thing that kind of you went down? Like what kind of, cause you of course built Columbia's cognitive cognitive science program from the ground up.
00:03:01
Speaker
Um, Like what was that journey like and how much of resources and knowledge was there at the intersection of this field at the time? Well, I should first mention, so cognitive science is itself a very interdisciplinary field involving AI and psychology and um philosophy and linguistics. And that goes back to the 1950s.
00:03:19
Speaker
So it's always been a pretty interdisciplinary thing. And, ah you know, in 1950s neuroscience was at a very early stage. There was just... say, Hubel and Weasel doing experiments, for example, showing that there are some neurons that are selected to inputs, but certainly nothing as sophisticated that would have much to contribute to cognitive science, and that's something that's changed.
00:03:39
Speaker
So the interdisciplinary nature of it has has really been there for a very long time. um And um the question about how the program started, well, of of course, there are many people who were contributing to the start of the program. I just took the lead.
00:03:55
Speaker
And um so cognitive science programs have been widespread in lots of universities. So ah since I think, say the 80s and 90s and Columbia and Barnard were just very slow and just had never gotten it started. So I wasn't starting something that didn't exist at the universities. I was just starting something ah you know, that was common other places.
00:04:19
Speaker
And it's a really great program because what it does is is it requires a kind of core competence across the field. So students have to take an intro to CogSci course, then they have to take a philosophy course, linguistics, psychology, neuroscience, and then two computational methods, usually a computer science courses, and then they choose their own specialization and they can cobble together courses from lots of different departments.
00:04:40
Speaker
So um that's kind of what sets cognitive science apart from other programs is that it's highly interdisciplinary.
Complexities of Brain and AI
00:04:48
Speaker
yeah Yeah. It definitely builds on like Columbia's core, of course, and people can understand like based off their core courses, what track they want to go down.
00:04:54
Speaker
And of course, like we can definitely see that you've been always the most interested in philosophy. And one of the things that you've been researching is like getting an understanding of their brand from like a more abstract perspective.
00:05:05
Speaker
And like to you, like, why is this important? And like, why should people kind of be caring about this, like what makes us so important in the way that we understand the mind and cognition?
00:05:16
Speaker
Well, both the brain and artificial neural networks are and the one hand, very important, but also really, it's unclear what they're doing. So we can even one understand what they're doing at a so neuron by neuron or node by node level, it's very hard to understand what they're doing at a more abstract human level.
00:05:36
Speaker
So um The big challenge where the big challenges is to try to come up with that kind of understanding. So it's just like if you had all the credit card transactions in the country right now, um that really wouldn't tell you very much about what's happening. You'd want to have some more higher level ah description about like what is the growth in the ah you know, what is the. um ah What is the deficit? What is inflation? All these sort of more macroscopic ah measures about what's going on. And those are the ones that scientists and policymakers, economists and policymakers use when and setting up laws.
00:06:09
Speaker
So likewise, we want something like that for the brain. We want to not just look at node by node, neuron by neuron. We want to understand what it's doing at a more macroscopic level. And that's a very hard conceptual question. So you know like in economics, it wasn't initially clear what those macroscopic measures are.
00:06:24
Speaker
ah right microscopic measures should be. And maybe we don't even have the right ones now. But in neuroscience, it's been even less clear. So the two the two basic ones from cognitive science are we should look at what is being represented and we should look at what more abstract high level computations exist.
00:06:44
Speaker
So to try to understand the brain is somewhat analogous to a computer and that there's the hardware, but there's also this level of software that we want to understand. So what is the software that the brain is running?
00:06:55
Speaker
And then for neural network artificial neural networks, the same thing, even though we train them and we have perfect access to them, ah We don't really understand what they're doing. um And so we want to have some more high level, traditional human accessible understanding.
00:07:11
Speaker
and And that's the kind so when I got tenure, I decided i wanted to work on what seemed to me one of the most fundamental projects were, and this seemed to me ah fundamental project that I could possibly have something to contribute to because and the reason the reason i might have something to contribute to it is that it's largely a conceptual question about what the right concepts are and philosophers are very good at or at least trained to be good at um sort of conceptual work and figuring out what the right definitions are and how to properly define them
00:07:43
Speaker
And that's why you see philosophers really get involved in sciences at the very beginning. So, you know, with the rise of mechanistic science in 17th century or quantum mechanics or ah beginning of computer science, um or even the starts of linguistics, there was a large dialogue between the scientists and philosophers. And the philosophers were helping them answer foundational questions, particularly conceptual questions.
00:08:07
Speaker
So anyway, I think neuroscience is a relatively new
Neural Networks and Algorithms
00:08:10
Speaker
science. So we didn't even know that neurons were responsible for for thought and for 100 years ago so it's and then the recording techniques took a long time to catch up so it's a very very very new science it's just at the beginning and so that's exactly the time when a philosopher like myself might have something to contribute and it's not a coincidence that right now what you're seeing is that ah the rise of philosophers interested in these questions and then the receptivity to them on the side of neuroscience um so there wasn't really much of a field in the philosophy of neuroscience say 20 years ago but there's um
00:08:41
Speaker
more so in the last 10 years and then really starting to pick up right now. Of course. Yeah. And it's something interesting where neuroscience, like just 50 years ago, we didn't really have a great understanding of how like there's networks in our brain, just like how there is in a computer.
00:08:56
Speaker
And like, it's not just so you don't have just individual brain regions doing something. It's all integrated networks working together. it's very interesting. Also hard questions. And of course, like on that note, like you're trying to define concepts like um like representation algorithm in the brain and you're trying to define those concepts and that's not easy to do right it's a very again foundational thing and it requires a certain skill set so how do you go about talking something so foundational just like defining those concepts and building out those questions when you're working with these researchers
00:09:28
Speaker
Well, a number of things are helpful. One is you want to go to as many talks as you can, and you want to speak with as many people as you can just learn. And the other thing that helps is teaching because teaching forces you to organize your ideas.
00:09:41
Speaker
And so I've started teaching courses on these topics, mostly for my benefit, but maybe they benefit student benefit too, just because it is a good way of forcing myself to think about how to organize the ideas and how to communicate them properly.
00:09:58
Speaker
um But like for now, like I had some ideas about how to define algorithm implementation. And so now I've been doing a lot of neural network modeling. um And so that's another way in which you can try to um clarify your idea. So it's if it's a big job, there's just, you do lots of things simultaneously and you hope that thinking you come up with something interesting.
Perception and Probability in Cognition
00:10:20
Speaker
Okay, and the the research you just kind of talked about and neural network modeling and kind of that area, can you kind of kind of talk us through like, what does that actually look like day to day? Like, what are you actually doing? And then what are the outcomes you're looking for, hoping for, and the questions you're trying to answer?
00:10:36
Speaker
Well, the big picture is that I'm looking for experiments that tasks basically where you can train an artificial neural network. And then there are two different algorithms that it could have learned. And then the question is, how do you figure out which algorithm it's using to complete the task?
00:10:50
Speaker
And then my general strategy is to look at other tasks and to see how quickly they can learn them. And that'll be some clue about what algorithm had been using all along. So basically, if it only has to make minor modifications to an algorithm, it will be really fast at learning some new tasks and slower it up.
00:11:05
Speaker
And then other tasks, we have to make big changes, it'll be slower at using learning those. That's called transfer learning. And so the idea is to use transfer learning to recover what the originally learned algorithm was.
00:11:18
Speaker
And so the day-to-day is to come up with experiments that have that form, which is that there's some input-output mapping it has to learn, and then there are multiple algorithms it could have learned to use, and then using transfer learning to help decide which is the one it's actually using.
00:11:32
Speaker
So my day-to-day right now is mostly like struggling with GPU clusters and and the and checking out lots of running as many experiments as I can and to see what are the interesting results.
00:11:44
Speaker
And i I have some interesting results so far. So I just got back from Cosign, which is a computational neuroscience conference where I presented some of them. And yes, that's my day-to-day. Yeah. And can you talk to us about what were the initial results? Like a simple, like simply like kind of what were results so far? And like, how do you think that'll impact like your work on the mind and understanding of that too?
00:12:05
Speaker
Well, the general, so transfer learning is a behavioral approach to figuring out what the algorithms are. You're looking at how quickly you can do something, which is nice because it's, uh, um It's something it's it's's something easier to export to the brain. And also, um anyway, um so what are the results? Well, the results were that the approach works, at least in the experiments that I came up with, involved very simple models and very simple tasks.
00:12:34
Speaker
And... um And then I did that for some al tasks where the had to and learn algebraic functions, and I did some tasks where had to ah ah classify images. And then sort of next stage is to scale up to language models, was starting with very small model models like BERT, and to see if it can work there. And in collaborations with some ah computer scientists and neuroscientists, I have some very preliminary results that are at least encouraging that for some tasks involving learning grammatical structure that it might work.
00:13:05
Speaker
so Yeah, it's it's sort of hard to go through the results specifically without showing you
The Value of Academic Research
00:13:10
Speaker
figures and all that. kind Yeah, that's just the gist of it. Yeah, unfortunately about podcasts is that we don't get to show pictures like YouTube and stuff like that.
00:13:17
Speaker
Yeah. And there are some other things we've been doing that are pretty interesting. We'd love to kind of dive into those too, like perceptual confidence. um It's a pretty unique term. I think most of us have like a know what those terms mean and and and a vacuum, but when you combine them together in the work that you're doing, can you kind of talk us through like, what is so important about perceptual confidence and what does that really tell us about how we perceive the world and how we understand this?
00:13:43
Speaker
The basic idea is that perception doesn't just present things as being one way. Sometimes it presents several possibilities and weights them with something like a probability. And so one example would be, you know you see a friend coming from the distance and you might say at some point,
00:13:59
Speaker
ah oh, that looks like it might be my friend, Isaac. And then they get closer, like, oh, that's probably him. and then really close, you're like oh, hi, Isaac. And so the thought is that what's happening is that visually even is that there's some possibility that's being introduced visually that it could be somebody. And then that possibility is given more and more higher, higher probability um as they get closer and closer.
00:14:22
Speaker
So um and the thought like lots of examples, like things happen very quickly or in the periphery or in the dark where it's like our perceptual system isn't just able to figure it out, but it doesn't leave it open either. It just says, well, here are some possibilities and some are more are more likely than others.
00:14:38
Speaker
So this is unlike the traditional view of perception, where you think about it like a painting, where it's just giving you one definite description of the way things are. and And this view, it's something much more complicated.
00:14:50
Speaker
so And this does dovetail with some work in neuroscience because there's some reason to think that early perceptual system is making use of probabilities, including like Bayesian kind of framework where you have some prior probabilities about what to expect. And then you have some estimate about how good your evidence is, some probabilistic estimate, and then that's combined to give you a result And then ah there's lots of lots of evidence that at the level of cognition or thought that we are using probabilities to make decisions, like when doing betting behavior, for example.
00:15:23
Speaker
But the traditional picture is this kind of um hourglass form. You start out with probabilities, but they're collapsed into just one possibility, and that's what you perceive. And then they're expanded out again at a level of cognition where um probabilities has come back into play. And so part of the big picture I have is that actually, no, the mental perception and cognition are probabilistic all the way through from the very earliest stages up until um action.
00:15:51
Speaker
And how do you kind of research and define that then? Because I think that is, it's very interesting where we kind of have, essentially like we're kind of reiterating probability and over over and again. And i assume like the brain comes up to like, some abstract concept of the expected value. like Is this is this this is my friend or not? Should I say hi to them? And then from that level, like how do you kind of understand and research that to kind of like publish a study and then hopefully like get this to the real world?
00:16:16
Speaker
um How about the real world? but the initial the initial yeah so The initial idea was just introspectioning on my own experience. And then I i gave some arguments that um we're supposed to show how this would give us a really elegant and simple explanation about why we're sometimes uncertain at the level of cognition or belief.
00:16:38
Speaker
and um But then I wrote another paper where I looked at some neuroscience evidence, basically that the same uncertainties or probabilities that are in the very early visual system are also present and available at the level of cognition. And then they are the argument was that it's just much easier to see why the same probabilities would still be there if they had been preserved through conscious perception.
00:17:00
Speaker
And that kind of argument is very tentative, in part because it involves consciousness and who knows where exactly consciousness arises in the brain. And we, you know, it's we ah also depends on speculation about the role of consciousness.
00:17:13
Speaker
But anyway, that's a more neuroscience kind of um like account. Yeah, and I think it's it's super interesting how in research, of course, it's not necessarily about like trying to accomplish something and impact the world. world It's about trying to answer big questions, answer questions that you're passionate about.
00:17:31
Speaker
And you basically spent your entire career in academia. i mean, did you ever consider going into the industry and what kind of has research and teaching in academia brought you in your time? I mean, i want to push back little bit on the real world look ah claim because i think these are about the real world.
00:17:47
Speaker
ah In fact, they're really fundamental claims about the real world. Just because they don't have commercial applications doesn't mean they're not about the real world. So like ethics is another field of philosophy that's very important and very much about the real world.
00:17:59
Speaker
But I wouldn't, doesn't have commercial applications. That doesn't mean it's not about the real world. um So that's really what universities are for is that we're, they're, there needs to be some people who are asking the basic questions and um and then preserving our knowledge about what other previous people have said about the basic questions, ah because it's just really important as a culture and a society for us to have that information and to maintain it. And it does, as a side benefit, have lots of, um you know, can also sometimes impact people that are trying to run businesses and things like that. But that's not its source of value.
00:18:35
Speaker
um Yeah. So I think that's, i mean, just just in the same way arts are valuable, even if they're not like selling lots of records or something like that. So it's just, there are things that are important that transcend um commercial applications. So um there was a second part to your question, but I didn't want to leave that that point untouched.
Understanding the Brain: Passion and Insights
00:18:56
Speaker
Yeah, of course. It definitely makes sense. I guess I kind phrased that question wrong. um But yeah, of course, and'm like i'm I was curious of a for you personally, what kind of has answering these questions brought you and why are you so kind of passionate about them and what's made your passion? Because people are different. Of course, people always love like answering different questions. been speaking people who really want to answer, like, how can we use policy as laborer to improve health care?
00:19:17
Speaker
Some are using like perception and ways to understand like AI and the way understand the mind. But for you specifically, like what kind of have these questions brought you and has teaching and academia brought you?
00:19:31
Speaker
Well, two of the the greatest mysteries for us right now is how does the brain give rise to thought and what exactly are these artificial neural networks doing?
00:19:41
Speaker
And really can we use these artificial neural networks as a guide to what the brain is doing? For me, these are all just ah very fundamental questions sort of the about biological and artificial systems that are just absolutely essential to everything that we do.
00:19:56
Speaker
And so, yeah that I'm, So my drive, i don't know if what I'll come up with will be of any use to anybody, but my drive is to try to understand them in a way that we using concepts that opens them up, opens up the black box to us about what they're doing. um And sort of a secondary goal is that you know neuroscience, I said, is a very new science. It's kind of unclear what the basic concepts should be.
00:20:23
Speaker
And I would like to contribute to neuroscience by giving it concepts that would be productive and useful to it. ah and And secondarily, even if the my exact proposals turn out not to work out, to encourage people in this field philosophy of philosophy and neuroscience, these are important questions that need to be addressed.
00:20:42
Speaker
So, um and then I just find that the kind of puzzles and questions that the research gives rise to intrinsically interesting and and um and you know, it leads me like in the last, has led me to have to learn how GPU clusters work, which is not something I would have thought I'd yeah having to learn at some point. So it's just, you know, learning and discovering are ah intrinsically motivational for me.
00:21:05
Speaker
Yeah, um that's really like always having that
Advice for Aspiring Cognitive Scientists
00:21:08
Speaker
curiosity. And for people that do kind of have that curiosity or interested in this, of course, philosophy and cognitive science isn't the easiest field to kind of explore in undergrad. There still are plenty of opportunities, though.
00:21:19
Speaker
And you kind of spoke on how you go to talks, you um and and classes are obviously a great way to learn that. but um And how what kind of advice do you have for students and younger individuals like trying to explore more about cognitive science and this intersection of where we're understanding mind the mind and cognition?
00:21:36
Speaker
Well, I guess I would encourage people to become cognitive science majors if they're sort of your life stage. um And the maybe the most important thing to do that that you might not know to do would be to work inside a laboratory ah because it's there's only so much you can know from regular courses, but if when you work in a lab, you you really get inside the way people are thinking and you also really understand the limitations and ah um potential um benefits of the methods that they're using.
00:22:13
Speaker
So that's the other piece of advice. So go to go to talks, take take classes, but also try to volunteer in a lab. Of course. Yeah. And we really appreciate that.
Closing Thoughts and Podcast Info
00:22:23
Speaker
And really appreciate your time today and kind of walking us through like your, what you've kind of learned, what you're kind of really passionate about.
00:22:29
Speaker
It's been super exciting to have you on. So thank you so much, Dr. Morrison, for coming on today and really sharing your story. i Really appreciate it. Absolutely. Thanks. Of course.
00:22:39
Speaker
Thanks for listening to The Healthcare Theory. Every Tuesday, expect a new episode on the platform of your choice. You can find us on Spotify, Apple Music, YouTube, any streaming platform you can imagine.
00:22:51
Speaker
We'll also be posting more short-form educational content on Instagram and TikTok. And if you really want to learn more about what's gone wrong with healthcare care and how you can help, check out our blog at thehealthcaretheory.org. Repeat, thehealthcaretheory.org. Again, i appreciate you tuning in and I hope to see you again soon.