Joel Ye: Examining Neural Data More Efficiently and Holistically

Sarah Webb  00:03

This is Science in Parallel, a podcast about people and projects in computational science. I’m your host, Sarah Webb. In season six, we’ve been looking at artificial intelligence applied to engineering and science problems. I’ve been talking with researchers working on foundation models, a type of AI trained on large data sets and then applied to general problems, ChatGPT and other similar tools are built on large language models, which are one type of foundation model. In this episode, I’m talking with Joel Ye, a Ph.D. student at Carnegie Mellon University who has been working on foundation models for neuroscience. He’s particularly interested in how researchers study brain data and is examining brain computer interfaces. These tools could expand the ways paralyzed patients interact with the world.

Sarah Webb  01:03

Joel is a third year Department of Energy Computational Science Graduate fellow, and that program supports this podcast. Joel talked about his research and the differences between language data and neural data, and how that could shape where foundation models are useful in this field.

Sarah Webb  01:26

Joel, it is great to have you on the podcast.

Joel Ye  01:29

It’s lovely to be here. Thanks for having me.

Sarah Webb  01:31

Joel, how did you get interested in computing, in neuroscience? What makes this interface of science interesting for you?

Joel Ye  01:39

I always interacted the world through, like a lot of my days spent on the computer growing up. I almost struggle to imagine any other sort of major interface to the world. For neuroscience, though I got into that in maybe the last five years, and there I was mostly driven by the potential use cases of really understanding the brain. Do that this divides sort of into, like a clinical use case because there’s lots of neurodegenerative disorders that impact many people as they age. And also there’s the kind of neurotech angle, which is, what can we do with the kind of brain data and our understanding of the brain for healthy populations through their life? And you know, for me, both of these are really interesting because these types of applications are like health needs. One of our main, like, sort of frameworks for understanding the brain is that the activity inside our brains is this real, distributed computational process where different neurons will interact with each other, and different brains interact with each other. And so it’s a network of activity. And now AI nowadays is also all about networks of sort of pseudoneurons interacting with each other. It’s kind of a like, not quite equivalence class between the two things, but there’s a lot of interesting overlaps in how we study the two. So there’s that.

Sarah Webb  03:00

So talk to me about kind of the core scientific question that you’re most interested in answering here at this neuroscience-AI interface.

Joel Ye  03:11

I would probably reframe this in the sense that most of what I’m interested in I see as almost kind of like systems engineering endpoints rather than scientific questions. And I think this probably applies a little bit pretty broadly for foundation modeling. So there’s two sort of takes on this. First, if we focus more on foundation models for neuroscience, which I’m here to talk about today, I think maybe the major question is, how do we create models that provide us effective interfaces to neural data. And the salient thing about neural data is often the fact that the data that we can record and observe in any of our experiments is very partial. It’s very limited. We can only interact with parts of the brain at once, and then, like after an hour, two hours, no matter what we’re recording from, we kind of need to, like, take a break.

Joel Ye  04:03

And so we’ll come back, and all of a sudden the brain has changed, or we’re looking at the brain from, like, an entirely new animal or a new person. And so the foundation model, hopefully, is an interface that will help us consider all of this data holistically. That’s like, sort of one vision for me, personally. I study foundation models, but I also am mostly focused on brain-computer interfaces. And there again, so for brain-computer interfaces, is this application where we are interested in working with neural activity that we record in people, and we want to translate that to motor intention. If they’re sort of paralyzed and can’t move, for example. And there, the question is, how do we create similar systems that are able to reliably, and perhaps even adaptively, look at their neural data and translate that to their intention? So how can we help users interact with the world?

Sarah Webb  04:55

So it sounds like in part, we’re talking about the available data, and the types of data that you have have changed dramatically, in addition to the computational tools and all of these other things. And of course, this is true in so many fields of science right now.

Joel Ye  05:12

That’s right. Previously, we would look at neural data and we’d say, okay, for example, I work with single-neuron spiking, and we’d say, okay, we can sort of make sense of this, and it relates to this external variable, like movement we want to study really clearly. But now there’s, I guess, a little bit of anxiety that now we’re recording thousands of neurons at once, and we can look at all of them. And it’s a little bit intimidatingly complex at this point.

Sarah Webb  05:35

How did you get into specifically thinking about foundation models for neural data?

Joel Ye  05:40

I have been inspired by the direction that we could but it was almost just a qualitative inspiration, rather than like a statistically or theoretically motivated inspiration. The idea that different data sets not in neuroscience, but in the sort of domains where foundation models first took hold: different language or vision data sets. You’ll train a network to do one task, and they’ll perform surprisingly well on lots of related tasks. And so when we first began, collectively as a field, to think about, well, what is the promise of this method or these ideas in neuroscience, we would look at different data sets related on these axes I alluded to earlier, like different animals, and say, Well, what is the if we train a network in one of these dimensions, does it apply or benefit our modeling of other related data sets? This initial like proof of concept, is not really what I would call foundation modeling work, but I have been interested in that, and this is all sort of recent thinking in the field. In the last two, three years, certainly as the machine learning ideas have also matured and become available for researchers outside natural language processing or vision.

Sarah Webb  06:51

Let’s dive straight into that in terms of how you think about foundation models, and what particularly makes this an avenue that you’re pursuing for the questions that you’re interested in.

Joel Ye  07:00

I think the strongest use case for foundation models in neuroscience is data efficiency. Because every time we want to analyze neural data, we’re sort of interested in sort of asking, like, what is the specific relationship of this data that I’m looking at? And how does it help me answer this neuroscientific question? Or, how does it analyze, like, how can I draw a relationship to these, like other covariates or other variables that I collected at the time? Often times, this will take a decent amount of data to characterize, because there’s lots of variables affecting the specific neural data that we see, and isolating sort of a single relationship within that will sort of require, let’s collect this trial a bunch of times average out the noise. This can be expensive. Experimental time is very valuable, and so as we try to study more complicated behaviors, for example, we’ll need to collect more and more data, and this will be intractable.

Joel Ye  07:53

So one of the hopes is that if foundation models can precondition how we analyze a new dataset. We can try to accomplish more complicated things within a fixed timeframe. That’s the case for data efficiency. I’ll rewind there and just say for brain computer interfaces, there’s less of a scientific question, but really, like concrete metrics around can we infer what the participant or BCI user is trying to do by having them only collect a few trials of data where they’re trying to say, move left, move right. So previously, the user would need to say, I’m going to move left, like eight times in a row. And okay, now we’re sort of starting to get it we say, can we just get that in one or two tries? And what this equates to in these practical applications is we can get higher performance with less effort. These types of applications are also relevant in the hospital settings, where clinicians are trying to characterize, like some we have like these neural data based diagnostics, and they’ll achieve some performance based on the amount of time you have in the hospital, and we would hope to achieve higher performance with a fixed amount of time.

Sarah Webb  08:58

Can you kind of talk to me about the kind of data that you’re working with? Particularly for people who are not neuroscientists, what does that look like, and then kind of, how does that feed into a model?

Joel Ye  09:08

So I have spent most of my time working with neuronal spiking activity. How this data originates is we have electrodes that will insert into the brain, and when neurons in the brain fire, they’ll sort of distort the like the electric field around them. And we sort of listen to these electrical field distortions, and when it’s like a particularly sharp distortion, we know, okay, there’s a neuron really nearby our electrode, and it has spiked. So there are patterns in, like, the collective spiking of lots of neurons together and meshed in a web. So we have a bunch of electrodes sitting nearby a bunch of neurons. And so we can come up with what are called these pictures of how a neural population is firing in a coordinated pattern. This is what I’m used to processing for these neural foundation models, and how we would feel. These into into any specific model, I would say, like where foundation models have most could have been studied in neuroscience has been for like this, functional activity of neurons processing things evolving over time. So other common data modalities that are analyzed in foundation models include functional magnetic resonance imaging or fMRI.

Sarah Webb  10:21

FMRI is built on the same technology as medical MRI. It measures brain activity by detecting changes in blood flow in the tissue.

Joel Ye  10:31

There are different types of electrodes that people put on depending on like how small the actual electrical contact is with the brain, and this will give you qualitatively different reads on the neural activity. So for example, outside the brain, a fairly popular neural data modality is called electroencephalography. So EEG, which is just people put electrical contacts, actually just outside, and put contacts on your head, or they’ll sometimes have a little preparations, but it’s essentially just electrode sitting outside your head. And even then, you can still capture some traces of what the brain may be trying to do, and these same techniques and questions come up. So it’s an exciting time, in terms of like having a same, like computational framework to try to study all of these different modalities, even though they occur and have a signal over very different timescales and spaces spatial scales.

Sarah Webb  11:16

So talk a little bit about those timescales, because, I mean, I’m imagining we’re talking about anywhere from fractions of seconds to minutes or longer. Is that a reasonable span, or is the span longer?

Joel Ye  11:28

I think neuroscience broadly does have to deal with a lot of spatial and temporal skills. I think for most intensive purposes, the actual quickest time scale that most non cellular neuroscientists might be concerned with is essentially the unit, the firing of a single neuronal unit, and that occurs over the timescale of about one millisecond. Okay, so, yes, we’re talking about like, millisecond timescales of modeling and for brain-computer interfaces, we’re more concerned with like, sort of like, what is the perceptually noticeable timescale? So that would be on the order of tens to hundreds of milliseconds of how often do we need to process the neural data and do something with that neural data? I think the slowest timescale neural data modalities, in terms of our recording technology, is maybe something like fMRI, which operates on the multiple seconds timescale.

Sarah Webb  12:21

What are you working on right now?

Joel Ye  12:23

One of my latest projects has been building a big foundation model for, specifically motor cortical data from non human primates and humans and to study if we can actually achieve some of the benefits of foundation models, such as data efficiency for motor decoding.

Sarah Webb  12:40

A quick explanation: Motor cortical data comes from the motor cortex, a part of the brain that controls voluntary movements.

Joel Ye  12:49

That’s the main thing. So we’ll have prerecorded data, variable, motor variables, and we’ll have motor cortical activity. We try to drive the relationship between them. We built a big foundation model for this, and now all the diagnostics in that work pointed relatively positively, but I’m now trying to take this and move it into the main application I’m interested in, which is for brain-computer interfaces. So it actually turns out that there’s a big evaluation gap between what we can easily study in the context of, like, sort of more machine learning centric, like you have a dataset you’re interested in, sort of, like look, our performance and modeling the relationship between the data and this dataset went up and saying that that model is useful in something like closed-loop control.

Sarah Webb  13:38

Closed-loop control is a system with three parts. For example, a human interacting with a computer screen or a robot arm. The third element is the physical world and the way that the computer or robot changes it. That third piece provides feedback to the human, closing the loop and allowing the human to adjust. By comparison, an open-loop system doesn’t provide the direct feedback to the user.

Joel Ye  14:07

So you may notice, may have noticed that the adoption of foundation models in robotics has been slower than it has been in the sort of more static domains of language and vision, because there’s lots of subtle interactions that are not captured between like the controller. Or in robotics, it’d be like the robot that’s moving around and its environment. For us, it’s sort of the way the human is modulating their neural activity to move, oftentimes, still like a robot arm around, and the neural activity essentially that we see during control, when they’re sort of like moving it back and forth, is out of distribution for basically these deep network models. A their behavior is not really even constrained. So it’s not clear that all the benefits we see in sort of like the well-controlled offline setting of prerecorded static data sets will translate easily to closed-loop control. So in summary, what I’m working on is evaluating the benefits of these pretrained models for closed-loop control.

Sarah Webb  15:12

So what have you learned so far and what’s next?

Joel Ye  15:16

Well, what I’ve learned so far is that these are pretty different playing fields, the open-loop and closed-loop problems. There’s a few other like directions I can briefly talk about. For example, there’s a thing outside of motor control that is also relevant for brain computer interfaces is sensory stimulation. So in this case, we want to provide these tiny electrical inputs into the neural population that we’re seeing so that we can provide the neural population like change the neural population state. What this feels like is, for example, when a person is moving a robot arm, and the robot arm touches a block, they can feel like their finger had a little tap or a little buzz.

Joel Ye  15:59

And this is actually useful for control. It’s not sort of just like a phenomena that occurs and it’s like pleasant and surprising. It’s like people will make use of this sensory feedback to improve how quickly and confidently they will pick up an object when they’re moving over about arm. So there’s a lot of challenges related to how we model the data that occurs when we stimulate the brain region, and there’s a potential application of applying basically the same large scale modeling techniques there. As for motor control, it’s like a wide open space, if we can resolve the issues around motor control well enough. And I think that’s a pretty challenging space, because when we’re talking about sensory stimulation, we will have kind of an exponential design space in like what we should do.

Sarah Webb  16:47

What do you see as the limitations of foundation models, either in your work or overall?

Joel Ye  16:54

 Ultimately, the currency for a lot of neuroscience is this insight, and right now, it requires creativity and familiarity with foundation models in order to like, understand something about neural data through the interface that foundation models provide us. So extracting insight from like the outside of the model is difficult. Now there’s been a lot of exciting work in natural language processing and computer vision of directly looking at how the model internals are configured to try to extract insight from there. And I think that drives a lot of people’s excitement for saying, like, maybe we can understand something about how the model is understanding the data if we investigate which model layers are connected to each other and how they’re wired together.

Joel Ye  17:42

However, I think that this is likely to be, I think it’s at least as early days for that as it is for the other sort of like black box usage of these models. Now, from a model-building perspective, for neuroscience specifically, I think that the idea that different data sets have individual, unique variability, and that’s something that needs to be acknowledged. And I think for language, where you’re trying to come up with a unified picture of language, underlying laws of how language are generated, essentially, like how people speak, however different neural data sets, like you’re looking at one part of the brain, another part of the brain, and there is maybe a commonality, like spikes can’t fire like infinite times in a row, right? So there is, like, a basic statistical constraint there. But I think that different neural data sets may have a lot of unique variability, and that can make it actually challenging to sort of imagine what a fully unified foundation model for all neural data would look like? In fact, I think it’s possible that we may arrive in a world where actually there are maybe not one huge model for all neural data but maybe dozens of models for different little areas of neuroscience or things like that.

Sarah Webb  18:55

What advice would you pass along to somebody who is interested in foundation models for neuroscience or systems engineering or other areas of science. What do you think is important to pass along to others?

Joel Ye  19:14

Yeah, so this, I’m actually quite excited to answer. It’s an interesting time to be in AI for science. What we see is all of these different fields already very computationally savvy, you know, like very like technical fields. But ultimately we’re seeing the arrival of, like, a somewhat new culture, like a machine-learning culture. And it takes the adoption of this culture, I think, for foundation modeling overall, to succeed. And what do I mean by this culture? I mean a field that has open data and open code and like clear evaluations, or like what we are trying to do, like these three pillars are really fundamentally prerequisites for new people to contribute. And. Right? You need to have some idea of like, what the methods are, and some idea of like what you’re trying to do right now, I think lots of fields, including neuroscience, like this culture and this tooling, therefore, is not really available yet, but it’s spinning up over, like, the last four or five years.

Joel Ye  20:16

 Like, people are recognizing the importance that like, it needs to be really easy for new people, not even, like, new to neuroscience. But say, like, I don’t do machine learning now: What am I even supposed to do to sort of try to start to engage traditionally? People try to do this by, like, going to, like workshops, or, like taking classes. But this, this is, like, not quite maybe the right strategy for machine learning. So basically, if it’s not clear that there’s like open resources for you to contribute to for a particular scientific problem of interest, keep checking back. I think over the course of these recent years, these tools are starting to come online. And if you have a machine-learning background, I think if you read through some of these scientific papers, you can focus on the methodological shortcomings, which I’m sure there are many, because as an interdisciplinary domain, we don’t keep up with the latest and greatest technical methods.

Joel Ye  21:06

And so you can either try to apply your expertise there to improve these models, or, for example, try to improve code infrastructure, try to formalize datasets and repackage them and define evaluations with them, and basically like try to translate your expertise for the domain. If you are a domain scientist, like a neuroscientist, but you don’t do machine learning, we need that expertise to define what is important. I know something like translating the idea that we want to understand something about neural data, or that we like the visual circuit should operate in this manner, into like a metric. It’s difficult, but even sort of evaluating the performance, like developing a way to evaluate models on essentially how well they capture your insights or your data, that’s really useful.

Sarah Webb  21:56

Is there anything that we haven’t talked about that you think is important to mention?

Joel Ye  22:00

I think in neuroscience, a major shift or acknowledgement that maybe foundation models are useful has not yet happened, and I think that’s justified. I don’t think that’s saying, like, oh, people are totally sleeping on this idea. I think right now, with the sort of subfield of neural foundation models, we’re really trying to figure out, like, where we’re useful, whether the models we’re building actually work. And there’s a few like, narrative beats that need to happen, like some landmark new insight that is developed, or some like, really incontrovertible, like, utility of these models, one or two of these wins, I think, will be necessary for the field to grow more and get established. That’s the status point.

Joel Ye  22:40

And then a second idea that I heard recently is that there is an interesting tension between data that is collected for neuroscience and maybe data that is important for neural foundation models. And this is not unique to neuroscience. There’s an opinion piece by Michael Bronstein, who works in synthetic biology. So basically, how it translates to neuroscience is the traditional neuroscientific data sets are collected for controlling as many variables as possible in order to make sense of a simple relationship, whereas oftentimes this can be maybe a little bit stifling for foundation models who can reckon with actually, like, maybe, like, there are 10 knobs that are changing at once on our data, and actually, that’s maybe sort of the type of data we need to study. Complicated. naturalistic is sort of the buzzword in neuroscience, where we’re not constraining the mouse to just shake its head left, or shake its head left, but let it interact with other mice, for example, or let a monkey sort of move around its environment. That type of data might be much richer. It’s also entirely complementary to what traditional methods are able to approach. So that sort of new area of data might be what is necessary for neural foundation models to flourish. And there’s a little bit of a chicken-and-egg problem in terms of funding it or incentivizing it, because you’re like, I don’t want to collect that data because I don’t know that foundation models will work. But maybe that’s the type of data where foundation models will really flourish.

Sarah Webb  24:12

On that note, Joel, thank you so much for your time. It’s a pleasure talking with you.

Joel Ye  24:13

Of course. Thanks for having me.

Sarah Webb  24:19

To learn more about Joel Ye’s research at Carnegie Mellon University and to explore more episodes and information about foundation models in science, please check out our show’s website at science in parallel.org.

Sarah Webb  24:34

Science in Parallel is produced by the Krell Institute and is a media project of the Department of Energy Computational Science Graduate Fellowship program. Any opinions expressed are those of the speaker and not those of their employers, the Krell Institute or the U.S. Department of Energy. Our music is by Steve O’Reilly. This episode was written and produced by Sarah Webb and edited by Susan Valot.

Transcript prepared using otter.ai.