Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Episode 22: The Neuroscience of Free Will: Guest Aaron Schurger
23 Plays4 years ago
Guest Dr. Aaron Schurger talks to us about his research on the meaning of the "readiness potential", which has been referred to as "the brain signature of the will". Although this neural signal was already famous from research in the 1960s, it was Benjamin Libet's infamous experiments in the 1980s that proportedly showed that the readiness potential preceded an act of free will by a few hundred milliseconds. More recently (in press), Dr. Schurger and his colleagues have convincingly demonstrated that the readiness potential is not in fact predictive of an act of free will, but instead comes from a lack of a proper experimental control.
Resources:
Here is what a classifier is (a topic that comes up that may be unfamiliar to some).
For advanced readers, check out AdaBoost, a tool that increases performance in classifiers and other types of machine learning.
Papers
"The Time Course of Neural Activity Predictive of Impending Movement" (Basbug, Schapire, & Schurger, TO BE PUBLISHED SOON)
An accumulator model for spontaneous neural activity prior to self-initiated movement (Schurger, Sitt, & Dehaene, 2012)
Unconscious cerebral initiative and the role of conscious will in voluntary action (Libet's 1985 experiments)
Special Guest: Aaron Schurger.
Recommended
Transcript
Introduction and Guest Presentation
00:00:10
Speaker
Hello and welcome to Cog Nation, the podcast about neuroscience, technology, and other stuff we like. With your hosts, me, Joe Hardy. And I'm Rolf Nelson.
Research on Conscious Will and Neuroscience
00:00:20
Speaker
On this episode, we speak with Dr. Aaron Shurger, assistant professor at Chapman University, where he is a member of the Institute for Interdisciplinary Brain and Behavior Sciences. Dr. Shurger is doing some really influential work on the role of conscious will and behavior and what neuroscience can tell us about that.
00:00:40
Speaker
Yeah, so we talk about a couple of Dr. Sherger's recent papers and the relation that they have to philosophical ideas of free will and some old experiments from the 1980s that he's shown some evidence against. So I think you'll find this a really interesting conversation.
Intersection of Consciousness and Free Will
00:01:00
Speaker
Yeah, it's a great show. Hope you enjoy. Dr. Sherger, thank you so much for being on the show. Really appreciate your taking the time to be with us today.
00:01:09
Speaker
Yeah, thanks for inviting me. It's, it's really a pleasure to be here. Great to have you. Yeah. We wanted to talk about your research. Uh, you've done some really interesting stuff lately in the area that sort of is at this intersection of, you know, brain activity, free will, understanding how the conscious mind controls or, you know, or does not control, uh, you know, behavior.
00:01:35
Speaker
What got you interested in this area, this topic matter? And, you know, give us maybe a little bit of background that sort of motivates your research. Yeah, sure. I mean, initially what I was really fascinated and passionate about was consciousness research. And for whatever reason, the research, pretty well-known research by Benjamin Libet that was considered relevant for free will.
00:02:04
Speaker
was something that floated around in those circles. Even though you might think that free will and consciousness are sort of different topics, I think they have a pretty strong intersection because we like to think of free will as acts that emerge from consciousness.
00:02:25
Speaker
We have a conscious intention and at least the way it feels to us is that that conscious intention manifests as
Critiques and Limitations of Libet's Experiments
00:02:33
Speaker
an action. And so there is this sense in which research on free will
00:02:40
Speaker
just fits right in that context. And one thing, so on a show a few episodes ago, we treated this topic too, and we talked about Dan Wagner's work on this, which you may be familiar with. So his book and his article that discuss the origins of free intent.
00:03:06
Speaker
Yeah, and the feeling, in his case, the feeling of conscious will. Right, that what needed to be explained was the reason why we felt that we were in control of this action. Right, and is that feeling justified? And of course, from Wegener's point of view, it's maybe kind of not so justified. His whole thing was that this is maybe an illusion,
00:03:35
Speaker
In fact, that was the title of his book. That's right, The Illusion of Conscious Will. Yeah, that's right. That's right. But that's how I got first introduced to the topic was, in fact, a conference that I had gone to very early in my career that was about consciousness. But there were several talks about conscious will, and those talks all referred back to this work by Libit.
00:04:05
Speaker
And that's what first got me interested in it. And so I got to know that work. And as I thought more about it, and I think this sort of evolved over a period of a few years, in fact, I wondered less and less about conscious free will, that aspect of it, and more about this readiness potential, which was the brain signal that was used as a sort of timeline
00:04:34
Speaker
uh, for the experiments that Libit did. Um, it's this, uh, you know, for, for listeners who might not be familiar with it, um, it's an electrical potential that comes from activity in the motor cortex. And it builds up very slowly over the period of maybe about a second, uh, in advance of a spontaneous voluntary movement. So movement that you make without a cue or without a stimulus.
00:05:04
Speaker
Okay, so I can take a stab at describing what I understand about the Libet experiments, and then maybe you can correct me if I'm making any mistakes or not describing it right. So in Libet's experiments, he was looking at
00:05:20
Speaker
the origins of an act of what he would call free will, I guess. So the act that he took was to make a small motor movement, just flicking your wrist or moving your finger, was it? Yeah, he had a couple versions of this. Yeah, I think it was a flick of the wrist. Yeah.
00:05:39
Speaker
So you would sit there, and at the moment where you felt the intent to move your rest, or you felt the urge to move your rest, you would move it. So this would be presumably an act that wasn't cued by something outside, and it originated entirely with the person. It was an act of free will, spontaneous will.
00:05:59
Speaker
And what, okay, so the other thing that Libet has them do is they're watching a clock. So it's rotating quickly. And what they're to do is to note the exact time at which they felt the intention to move. Right. So, uh, Time zero would be the time that they moved and time, you know, minus whatever would be the, the time that they first felt that intention to move and then
00:06:26
Speaker
Yeah, you're sort of working back from there. That's right. And so if I'm correct here, so he finds about maybe 200 to 150 milliseconds before the action, people report the intent to move or report noticing the intent to move. However, this is preceded by a few hundred milliseconds
Readiness Potential and Prediction Challenges
00:06:48
Speaker
by the readiness potential so that you can predict when someone is going to move
00:06:57
Speaker
prior to when they notice the intent to move. In other words, there's something that precedes the intent to move. And the idea is that this indicates that an act of free will is always preceded by some preparatory neural activity. Right. And that's about where my understanding is at. So everything you said was correct, except the predict. OK. You cannot really
00:07:26
Speaker
predict. I mean, in the strictest sense of the word. So predict, I'm a real stickler about that. You know, predict means tell the future. And so if you, you know, people have tried to do this, and the reason I can say that with with some confidence is that those attempts haven't really worked very well.
00:07:45
Speaker
So I guess this is particularly relevant to brain computer interfaces too. So if you're trying to use this readiness potential for any kind of predictive power, you're saying it's really not going to. Exactly. It works slightly better than a coin toss. Works better than a coin toss. But to say that you could predict, that I don't think is justified. I mean, again, having something that works better than a coin toss, for me, that doesn't count as predicting.
00:08:15
Speaker
So if you, if you try to lock on to that, to the readiness potential and you try to detect it in real time and use that to predict when the movement is going to happen in advance, that just doesn't work very well. So in other words, Libet is looking at these and he's, he's making a post hoc interpretation of what these readiness potentials are based on when the movement starts.
00:08:43
Speaker
That's right, and the readiness potential by definition is something that you recover in the time-locked average. So you have to take X number of instances of one of these spontaneous movements, take the EEG traces from those and average them together in order to see anything. If you look at a single instance, it's a very, very weak signal compared to the noise, and you pretty much can't see it.
00:09:13
Speaker
Okay. So the idea here is that, you know, in the experimental condition before you do the analysis, you know, already when the movement in fact took place. Yes. And so there's this post hoc bias where you, you know, you're time-locking after the fact to that moment where the movement happened. And if you average backwards from that, you in fact see a pattern that looks
00:09:39
Speaker
really nice. It's a very smooth pattern when averaged across a large number of trials. That's absolutely correct, yeah.
00:09:47
Speaker
And what you're suggesting and what your research shows is that that is kind of, well, is rather misleading. Yes. Yeah. It's a biased sample. It's a highly biased sample. You only looked at instances where there actually was a movement. Most of the time during that experiment was spent not moving. And only a small fraction of the time was actually spent moving.
00:10:13
Speaker
So a couple of analogies I give for this to just help people kind of wrap their heads around it is one of them is with weather forecasting. So imagine that you're trying to learn how to predict rainfall based on weather data. And you do that with a sample of data that is
00:10:43
Speaker
in which it always rains, right? So you have a sample of data of weather during the day of time periods where that time period always culminates in rainfall. You're really missing the big picture, right? Because whatever you learn from that exercise, it's going to be difficult to use that to generalize to the reality because much of the time it doesn't rain.
00:11:11
Speaker
In order to be able to predict the onset of rainfall, we can say, well, maybe that's analogous to the onset of a movement. You need to learn about what's happening when it's going to rain and learn about what's happening when it doesn't rain. One of the consequences of ignoring
00:11:38
Speaker
the negative examples is that you may make lots of false alarms. That's right. That's right. Yeah. And that was in fact, at least in the research to date that I know of trying to build a BCI, a brain computer interface that works off of the readiness potential, one of the main problems they've had is keeping that false alarm rate down.
Neural Noise and Experimental Bias
00:12:08
Speaker
Yeah, the research that you've done on this topic, I think, along with your colleagues, has been very, very interesting from the perspective of shedding some interesting light on both the methodology that you should use, but also then, of course, what that implies for the readiness potential itself. So maybe we could dive into a little bit of this work that you've done recently and
00:12:38
Speaker
talk about the time course of neural activity, predictive of impending movement, and maybe a little bit about that experimental paradigm and what you think is important there. Yeah. So this was really an offshoot of work we had done earlier back in 2012.
00:13:01
Speaker
where we argued that the readiness potential reflects ongoing, you might say, random or stochastic fluctuations in neural activity that get sort of caught in the flash photo of time-locked averaging. So because you're locking to the onset of movement,
00:13:26
Speaker
if those ongoing fluctuations tend to favor movements at certain times rather than others. So if, in fact, at crests, right, at crests in those fluctuations, if movement is more likely then, at those times, then by time locking to the onset of the movement, you'll recover this slow
00:13:51
Speaker
fluctuation which is in fact stochastic and ongoing and does not reflect something preparatory or intentional.
00:14:02
Speaker
Now in Libet's experiments too, so this, I mean the task that he chose is one where, so if you're just kind of sitting there and either moving your wrist or not, you're kind of right on the borderline or it would be, it seems as though it would be a particular situation where it's easy to
00:14:26
Speaker
introduced very little to get your hand to actually move. So it seems like a situation that's particularly susceptible to any random noise that might be floating around in the system, right?
00:14:40
Speaker
Right. In fact, that's spot on. That's exactly what we were thinking when we came up with these experiments was, well, hold on a minute. For the moment, at least, let's set aside the whole debate about free will and compartmentalize this a little bit and look just at the facts.
00:15:04
Speaker
As much as people have complained about this task of Libet saying, oh, well, this is such a weird task. Nobody ever does that in real life. Nobody sits there and performs a movement spontaneously for no good reason, just for the heck of it, which is what you're asking people to do. So, okay, fine. Let's set that aside though, because the fact is you bring people in the lab and you ask them to do this weird, awkward, unusual kind of task. Fact is that they do it.
00:15:32
Speaker
They do it. They can do it. At first they kind of say, well, gee, that's a little weird. What do you mean? And then after a few tries, they're like, okay, I get it. And they do it. So, okay, the brain has a solution for that problem. And the question we posed was what could that solution be? What is the brain doing that enables it to produce a movement spontaneously without any good reason? And it's sort of a randomish time. And
00:16:01
Speaker
what came to me right away was, well, let's say I was trying to build a robotic arm that could do this task. How might I do that? And immediately I thought, well, that's easy, right? I just bring up the noise close to the threshold and I just wait.
00:16:20
Speaker
So if you're, and if you're thinking about how, I don't, I don't know if it's exactly how participants are experiencing it, but it is, you're right. It is a, I mean, it is a problem for participants in the experience experiment where they want to perform as well as possible. And so they have to understand the task and that you don't, you don't necessarily
00:16:43
Speaker
need to have any reference to free will to perform this task, it seems. Like you say, if you can program this sort of behavior in a simple robot, then it's probably true that it's not something that absolutely requires that sort of free will that we're talking about. Yeah, that's right. And that's been argued in the past. And I
00:17:06
Speaker
I largely agree that if there is an act of free will involved in this task, it's the act of accepting to follow the instructions of the experiment in the first place. The rest is, you could say, mechanical. That's not to say that there aren't
00:17:26
Speaker
other factors in this experiment. I've thought a lot about
Neural Noise in Nature and Experiments
00:17:29
Speaker
it. And of course, there's no doubt, like I do that. I mean, I've sat there and done the task myself many, many, many times. And of course, you quickly realize, okay, there's more layers to this than just random fluctuations in a threshold. You're thinking things like, oh, well, I moved kind of a little bit early on the last one. So maybe I should wait a little bit longer on this one.
00:17:55
Speaker
There's some strategizing that goes on. We found, in fact, a very slight but very consistent preference for the bottom of the clock. Yeah, I had wondered about what sorts of regularities you might find because if people are adopting some strategy, it's likely that they're going to perform in a regular kind of way, in a biased kind of way.
00:18:22
Speaker
Yeah, I think there are multiple levels of biases and strategies and influences on the eventual outcome in that kind of experiment. We just latched on to one of them, which is that if background noise in the motor system is somehow involved, then this makes perfect sense. It explains a lot of things.
00:18:49
Speaker
That background noise gets caught in the act in a way because you're time locking to the movement. Can you say anything about what we might know about background noise in this sort of situation?
00:19:07
Speaker
Yeah, or in any situation in fact. Or in any situation. In any situation. What we know is that, and this is a little bit hard to explain without a visual, but I'll do my best. We know that noise in the brain, and in fact noise in nature in general, if you just look in time series in nature, you don't see what's called white noise. White noise is noise where every sample is independent of the one before it.
00:19:38
Speaker
So totally random. Totally random. Well, random. And yes, totally random. But the key is that each sample is drawn from a certain distribution, and the next sample is drawn from that same distribution, but completely independent of the one just before it. So all frequencies are equally present in white noise.
00:20:02
Speaker
The kind of noise that you see in nature and in neural systems at all levels of spatial scale is pink noise or autocorrelated noise. Lots of different ways to say it. Some people say 1 over F noise.
00:20:20
Speaker
So if you look at it by eye, it's noise that has a kind of slow, drifty character to it. It fluctuates all around, but you see this, it's like waves on the surface of the ocean. It's slow, it's drifty, and it's dominated more by the lower frequencies, and there's ever less and less energy in the higher and higher frequencies.
00:20:51
Speaker
So hopefully that makes some sense about it. When you see it by eye, you can just see the difference. See, that's white noise. That pink noise shows up in a lot of different contexts. So for example, if you look at the distribution of energy, you know, visual energy in a natural scene, you see that same sort of pattern. But I think in the
00:21:14
Speaker
In the world of EEG, there's a lot made of the different slow wave oscillations and their meaning, right? There's a lot of research and activity trying to understand what those different frequencies are doing. Yes. Yeah, that's right. And this idea of 1 over f noise or pink noise, this is the component that's not periodic.
00:21:44
Speaker
So you may have oscillations going on. Imagine if you could factor those out, you still have this slow random drifting going on that may not necessarily be oscillatory. And that's at least the factor that we think is playing into this.
00:22:06
Speaker
is this non-oscillatory random drift in neural firing in the motor cortex. So would you describe the drift as the reason why it might be hovering around the threshold and sort of poke over the threshold more often? Then say white, then if it was say pure white noise.
00:22:34
Speaker
Well, one difference is, I mean, hopefully that noise, that drift is well below the threshold most of the time. Most of the time, right. Which is why the situation, yes, is odd for the Libet experiments. Yeah, right.
00:22:51
Speaker
So one way to think of that, one way to model that is to say, well, in the context of that experiment and given the instructions that you're given by the experimenter, you get into this mode where you're just right at the threshold of moving. At any moment you could tip the scales. You're in a state of heightened readiness to move, let's say.
00:23:15
Speaker
One way to model that is to say, well, I'm just going to raise the whole noise floor up closer to the threshold and kind of play around right at that level. And in that context,
00:23:33
Speaker
that random drift can sometimes tip the scales, can sometimes cross the threshold and favor you moving at that moment at a crest or say more so than at a trough in those fluctuations.
00:23:48
Speaker
So random thought here, I wonder if this is something that would play into offsides in a football game. So jumping the gun a little bit, or maybe even in a track race, too, jumping the gun. So as you're right at that absolutely ready to start the race, anything will set it off, or a very small fluctuation will set it off.
00:24:10
Speaker
Right. It's a beautiful example, actually. I hadn't thought of the offsides one, but yeah, I think that's the same sort of context. So yes, absolutely. I mean, if you could, those are both of those contexts, it would be difficult to get in there and look at what's happening in the brain. But imagine that you could, and you could sort of take a microscope to what's going on. You may well see that, right, some ongoing
00:24:37
Speaker
random fluctuations might contribute to that happening when it does happen. So this 1 over F noise is causing some artifacts when you time lock your analysis to a movement event. Are there other artifacts that might also play into this if you time lock to that event?
00:25:02
Speaker
There are other potential artifacts. One of the main ones that we've thought about, and actually that's so coming back, we've kind of digressed from the time course study that you mentioned, but yeah, one of them is that that noise can sort of at any moment
00:25:29
Speaker
as long as you're kind of close to the threshold could potentially tip those scales. And if that's the case, then one of the things that predicts is that your ability to predict the movement based on that signal is actually going to be poor.
00:25:57
Speaker
And it's in fact going to not look like the signal that you see. You're going to see this slow ramping that goes back quite far in time. And that might lead you to think that you can predict quite far ahead in time.
Machine Learning and Movement Prediction
00:26:14
Speaker
In fact, one of the things that our model predicted is that that's a mistake, that in fact, that's a mistaken assumption. Then in fact, you won't be able to predict very well. That's what we did in that study, the time course of neural activity predictive of impending movement.
00:26:39
Speaker
is we tried. So we took a situation where we had little windows of time, little epochs that ended with a movement, and pretty well-matched epochs that ended without a movement. And we used machine learning to try and tell them apart.
00:27:02
Speaker
So in this one, okay, so a participant would advance the slide and it would either need to be advanced by them clicking to advance it, or it would advance on its own at a time based on their history of pressing the button. Is that correct?
00:27:24
Speaker
Right. The distribution of delays on the trials when it advanced automatically was drawn from their own distribution of waiting times on the trials where they advanced the slide manual. So that by the end of the experiment, on average, the window of time was roughly of equal length, whether they moved or didn't move.
00:27:49
Speaker
So then you can get a clear signal that compares the same situation when they actually make the movement and when they don't actually make any motor movement. And what do we see in the comparison between these? Well, when we try to predict those, let's not use the word predict. When we try to classify those,
00:28:16
Speaker
using machine learning at each position of a sliding window. Surprisingly, what you might think is that, oh, well, we should basically recover the shape of the readiness potential, this slow ramp. But in fact, what we find is that you don't get a slow ramp. You get something that is pretty much near chance until almost the onset of movement and then suddenly lurches up to near perfect.
00:28:50
Speaker
So just trying to understand this a little bit better. In this case, so again, back to the task, the participants are looking at a slideshow. It goes from a gray screen to a picture. It's like a, looks like they're natural scenes or something like that. Nature photos. Okay. So you go between a gray screen and a nature photo and.
00:29:15
Speaker
And on the manual trials, you press a button to advance the slide. And on the automatic trials, the slide is advanced at a time that's selected from the distribution of your button presses from past trials. Cool. And so then in that case, you're talking about the ones where you press the button are what you call active trials. And the ones where you don't press the button are called passive trials. Is that, is that right too? Okay.
00:29:45
Speaker
Correct. So if you look at the time course of just the potential that you could, in the active trials, you see this pattern that, you know, if you average across all these trials, across the participants, you do see this readiness potential quite clearly in the active trials, but less so in the passive trials, just visually looking at the graphs, right?
00:30:10
Speaker
Right. I mean, you'd think you might see nothing at all, in fact, in the passive trials. But one of the things you mentioned before was, you asked before, are there any other potential artifacts? And one of them might be anticipation.
00:30:30
Speaker
Um, and that's a tricky one because it's always going to be there. So when you're, whether you move yourself or move, or let the slide advance by itself, you are anticipating a visual. Right. And you have some sense of how long it's going to take because while you're not necessarily in control of this trial, you know, basically the distribution of the timings that it will advance.
00:30:57
Speaker
Yeah, you probably have a rough, or your brain has a rough handle on the distribution. You don't know the exact time on any given trial, but you're anticipating that something is gonna happen soonish. That being a visual event, a slide transition. So the experiment,
00:31:25
Speaker
in a way, controls for that, controls for anticipation, controls to some degree for the issue of autocorrelation. But there's still some anticipation that is tough to get around, which is you're anticipating
00:31:50
Speaker
on the trials when you move yourself, your brain is anticipating some proprioceptive feedback that then you don't have on the automatic trials, on the passive trials. But all that really does is makes it even more likely, not less, but more likely that we should be able to classify these trials, that we should be able to tell them apart
00:32:21
Speaker
And so I think that makes it more salient, a result that we really don't tell them apart until quite close to the onset of movement, which is one of the things we were predicting. So what's the interpretation of this then? How do you make sense of what is a readiness potential? I mean, is it something that exists and what does it signify?
00:32:51
Speaker
Well, I mean, our best guess now, based on the research we've been doing, is that the early part of the readiness potential, so going back as, you know, as far back as it goes, one of the things about the readiness potential that's stood out for me is how variable it is from one individual to another. Compared to other neural phenomena, it's insanely variable.
00:33:17
Speaker
Which is which is strange already, right? Yeah, it is strange. It's it's suspicious, right? So for some people it goes back maybe a half a second for some people it goes back a second or even a second and a half That's huge Yeah, that's much longer than I had I had assumed it could be yeah. Yeah, I mean I've seen it drift back really far in some subjects Far back in time So there yeah, there is that
00:33:47
Speaker
So I guess you may only have intuitions or educated guesses about what it might be reflecting, but if you could guess, how would you describe it?
Decision-Making and Neural Activity
00:33:58
Speaker
Well, I think I like to keep explanations as simple as possible until I'm forced to do otherwise. Right, sure. And so here I think what I would say you can describe the readiness potential as a composite of two things.
00:34:15
Speaker
One are spontaneous fluctuations, spontaneous 1 over f noise, which crests in that noise happen to coincide with the onset of movement. That accounts for the early part of the readiness potential up to about
00:34:42
Speaker
two-tenths of a second before the movement, up to about the time when Libet's subjects reported feeling the urge to move, by the way.
00:34:51
Speaker
Right about there, two tenths of a second, 150 milliseconds, somewhere around there. And then the latter part of the readiness potential, which is where it's really more peaky, it has a kind of a sharp peak. That's actually a motor potential coming from primary motor cortex that's literally sending a signal to your muscles to move.
00:35:13
Speaker
And that accounts for the latter part, the readiness potential. So this early part, I would say I call it pre-decisional. You haven't technically decided yet or your brain hasn't technically decided to move yet. You're just collecting evidence, I guess. Yeah, collecting evidence weighing whether or not to move.
00:35:36
Speaker
but not having decided yet. And then at about 150, 200 milliseconds before the movement, that's actually when the threshold is crossed and that leads to a burst of activity in the primary motor cortex and then a subsequent movement. And what's nice about that explanation is that it jives really well with a lot of research. So
00:36:03
Speaker
For one thing I just mentioned, that just happens to be the time when subjects very consistently over many, many studies, not just libits, when subjects report that that was when they felt they had decided to move. That was when they felt the decision had been made. And there's an abrupt increase in cortical spinal excitability at about 150 milliseconds before movement.
00:36:30
Speaker
before a spontaneous movement. And a more recent study from John Dylan Haynes' group showed that it looks like the sort of point of no return, the point at which you can no longer withhold a movement if you're asked to, is around 200 milliseconds before the movement onset as well. Why don't we take a short break and then we'll get back with some of the implications of this.
00:37:05
Speaker
Quick plug, if you like the show, please share with a friend. Write us on iTunes and like us on Facebook. You can also get more details about our episodes at cognation.fireside.fm. All right, we're back.
00:37:31
Speaker
So Aaron, thank you again for being on the show. Appreciate it. Super enjoying the conversation. Wanted to dive in a little bit more on some of the specifics of your classifiers that you're using in this particular study, because I think it's interesting from a practical perspective, but I think it also might shed some light on sort of the theoretical elements as well.
00:37:57
Speaker
So in this research, you showed that you could build a classifier based on the biased data that is the way that this research has typically been done and get very good predictions far in advance of the movement in most cases. But that was not the case when you did this more controlled approach. Can you talk about that a little bit?
00:38:21
Speaker
Yeah, that was really one of the things that we wanted to highlight with this work. Because if we just showed that you're unable to classify these two different kinds of episodes that end in a movement and episodes that end without a movement.
00:38:41
Speaker
If you're unable to classify those before the time of movement, one could argue, and they'd be justified, that that's just a null result. And a null result is just a null result, right? Right, maybe it's just not a good model, right?
00:38:58
Speaker
Maybe it's just not a good model, although there are reasons to doubt that. I think it is an excellent model, in fact. But let's just say, for the sake of argument. And so what we wanted to do was to say, well, let's look at what happens when we only use the data with movement. If we're right, then maybe we should regain the ability, well, apparently regain the ability to predict.
00:39:28
Speaker
when we, or to classify, when we work only with the data that ends in a movement. And that's in fact what happened. So without going into detail about exactly how we did that, we worked only with
00:39:44
Speaker
the data where there was a movement. And when you do that, when you're doing this kind of class, this is a binary classification. When you're doing a binary classification, you need both positive and negative exemplars. And you train your classifier on a subset of those, and then you test it on another subset. You do that in a sort of round robin fashion. That's called cross-validation. It's to make sure that you're actually generalizing that your classifier isn't just memorizing the data.
00:40:15
Speaker
And so when we do that with only the data that ends in a movement, what do we use for the negative exemplars? Because we only have positive ones. So what we did, and this is just something that has been done in the past in other studies, is we said, well, let's take a little window of time that's really far back in time from the movement.
Challenges in Predicting Spontaneous Actions
00:40:43
Speaker
And we'll just imagine that it's far back enough in time from the movement that probably it's got nothing to do with the movement, and we'll call that our negative example. And then we'll position the sliding window somewhere else closer to the movement, and we'll call that our positive example. And we'll just ask the classifier to tell those apart. And what we predicted, if the onset of movement tends to coincide with crests in background fluctuations,
00:41:13
Speaker
then that method of classifying, comparing some early time window to time windows at other positions, that you're guaranteed to have better and better and better accuracy as you get closer to the movement. And that's what happened. So we could tell this apart exceedingly well, almost too well.
00:41:40
Speaker
And that's in fact, just the point we were trying to make. So this is almost certainly artifactual. We've convinced ourselves that we can, we can quote unquote, predict. That's just because we only have some of these subjects, you know, you see some really great predictions area under the curve approaching like point nine, a second and a half before the movement, which is very suspicious.
00:42:05
Speaker
very suspicious. And that was the goal of that whole exercise. So we thought that that might arouse suspicion. And when we looked at the results, we were right. I mean, we were like, yeah, this does arouse. If I saw this, I would be really suspicious. And I think that suspicion is totally justified because when we then
00:42:31
Speaker
repeat the whole exercise, but now we reintroduced the data where you don't move. And now I think that's a genuine comparison. This is a genuine task for a classifier to do. You have positive and negative examples, both. And you have to learn to discriminate between them. And that ability to classify early on, early with respect to the onset of the movement, just completely vanishes.
00:43:02
Speaker
And you can't tell them apart until basically right before the movement. And then at the time of the movement, your classification accuracy just lurches right up to near ceiling, almost one. Yeah, so that's really, yeah, that's very, very compelling to me in the sense that it also tends to correspond much better as we were talking about before with the person's subjective experience of when they feel like they're
00:43:32
Speaker
planning to make a movement. If one were to want to be argumentative here, one might just say, again, you still have the null result issue, the sense that maybe there is some signal that's happening earlier, but it's just not being picked up well by the model, for example.
00:43:58
Speaker
That's always possible in this kind of approach to studying the problem. That's just the reality. Yes, someone could always say that. Really, the only answer to that is to say, look, in addition to doing this sort of control analysis, we've done the
00:44:21
Speaker
very best we possibly can with the technology we have and been extremely careful and rigorous and persistent in trying to find any difference we could before the onset of movement. We used a classifier called Ataboost
00:44:44
Speaker
uh, which in fact, one of my co-authors, uh, co-developed Rob Shapiro, he was the co-inventor of Adaboost, um, that is
00:44:58
Speaker
With sufficient data, and I'll wave my hands a little bit as to exactly how much data that means, but with sufficient data, Adaboost actually offers some provable guarantees that if there is a difference between the positive and negative examples, with sufficient data it will find it.
00:45:22
Speaker
And that at least helps to reassure that combined with the fact that right at and after the onset of movement, we can tell them apart at close to 99% correct. I think those two things combined at least is compelling.
00:45:43
Speaker
Um, it, it, it never erases the argument that, well, this could just be a no result. There's no way I find it very compelling. And certainly the modeling work seems, you know, very, very robust. The one thing that I found a little challenging in understanding it was just in more, you know, purely from a intuition perspective is comparing, you know, the, the visual representation of the graphs.
00:46:12
Speaker
of the readiness potential when you're looking at the time course in like figure two, where you're comparing active trials versus passive trials on average. And you can see that on average, it does look like there is a difference in that time course between the active and passive trials. Then it begs the question, well, what is causing that
00:46:38
Speaker
time course that is then not being picked up by the model. Right.
00:46:48
Speaker
Right. And I think some of the things that may qualify as potential artifacts in our design or potential problems, I think they actually go in our favor because someone's going to say, oh, well, your conditions aren't that well matched. In fact, they're not as well matched as you think they are, something like that. Well, if they're not so well matched, that's all the more reason for the classifier to find a difference and just further begs the question of why it didn't.
00:47:15
Speaker
Yeah, for sure. Yeah. No, I think it's, I think it's super, super interesting and super compelling from that perspective. The fact that, you know, when you, there's this result that has been, you know, discussed in the neuroscience literature for what over five decades now. And yeah, we're talking about the readiness potential, something that is taught in introductory neuroscience courses and what you're suggesting. And I think, you know, rightly so.
00:47:44
Speaker
This is something that could be really an artifact of a particular experimental design rather than a true underlying reality of the way the brain works. Yes, it's actually both.
00:48:06
Speaker
So it's both. It's the interaction between those two. It's an artifact of the approach of time-locking to movement on set and the fact that fluctuations in neural activity are autocorrelated, have that pink noise character. So it's telling you something about both. If brain noise, if brain fluctuations were white in character, then we wouldn't get this either.
00:48:36
Speaker
All right, so I'm wondering if we can zoom out a little bit with the time that we have remaining.
Neural Noise and Decision-Making in Daily Life
00:48:41
Speaker
One of the takeaways that we're getting here is that some decision making processes are susceptible to random noise, especially if it's a precarious decision, maybe that it could go one way or the other easily, could be influenced by noise.
00:49:02
Speaker
How does this extend to other kinds of decisions or are there other situations in kind of everyday life that we might be seeing the same kind of influence? Yeah, I think so. I think potentially it has relevance in all kinds of situations. So wherever you have a decision that's really close, let's just say simply difficult decision that's really like a tie.
00:49:32
Speaker
So you're agonizing about it and you really feel the pull of both sides almost perfectly equally. Yeah, if you've heard the story of Bourdan's ass, Bourdan's donkey, he's caught between two piles of food and he dies because he dies of hunger. Because he's right on that fence and he just can't move himself up that fence.
00:49:54
Speaker
And we always manage to break the symmetry. We don't die of hunger, right? So something happens that breaks the symmetry and that could be it, right? We may make use of our own internal noise, if you will. We almost like a roll of the die that we can turn to if it's right in the eyes.
00:50:15
Speaker
It could be a strategy to direct a decision in this way to make it so close that you kind of throw it up to chance. Right. Yeah, those would be the experimental circumstances that you'd want to cook up in order to expose it, which is what Libit did, I think, or what Libit was aiming for.
00:50:39
Speaker
I think one of the other areas where this could be relevant is in game theory. Because there are circumstances, there are situations where it's adaptive, it's useful to be unpredictable. Rock, paper, scissors, right? Rock, paper, scissors, matching pennies.
00:51:00
Speaker
sports, all kinds of sports, maybe this is a mechanism that helps to generate unpredictable behavior. What's interesting is that if this is such a mechanism, if it's used in that way, then in fact, that behavior really is pretty unpredictable. It's unpredictable even to you.
00:51:30
Speaker
Which could be, it could be advantageous in some circumstances, but yeah. Yeah. Yeah. You could imagine, uh, you know, I've thought, I don't know. I've thought about sports like, uh, like soccer, European football, uh, you know, where that, where that could be a factor. Um, you know, you're head to head with someone and are you going to, are you going to cut to the left or cut to the right? Being unpredictable is a, is an advantage.
00:51:57
Speaker
is an advantage. So there are contexts where it's an advantage to be unpredictable and that has come up in game theory. And another question that I feel like I'm bound to ask is, I mean, certainly not everyone is convinced by Libet's experiments that there's some other precursor to free will, or that his experiments showed anything about the causal nature of free will.
00:52:25
Speaker
And I think a lot of neuroscientists might say, well, of course, there's something that precedes your intention. There always has to be a cause of your intention to move or of anything you do for that matter. So how would you think about this issue? Do you think that Libet's experiments really don't have anything to do with free will? You said that maybe you're moving away from thinking about these experiments like that. How does this sit?
00:52:54
Speaker
Right.
00:52:56
Speaker
Right, I think that the implications for free will that we originally attributed to Libet's experiments, I think
Discussion on Free Will and Neuroscience
00:53:06
Speaker
that those may have been misconstrued and I'm not sure that it in fact has those implications for free will given the way that we've explained the readiness potential. In other words, the implications for free will in Libet's experiment depend crucially on your interpretation of what the readiness potential means.
00:53:26
Speaker
And I think no one really stopped to carefully ask that question, well, what really does this readiness potential mean? Before we get onto this further question of about free will, what does this temporal marker mean that we're using to make, to draw conclusions about free will? So yeah, I think that Libet's experiment may not have the implications for free will that we have long thought it did.
00:53:54
Speaker
Now, even if it were not for... So just imagine in another reality that the readiness potential was perfectly predictive and had none of the random properties that you talk about. Would you take this, would you take Libet's experiment as evidence against free will? Yeah, that's an excellent question. And it's one that I've thought a lot about because I think
00:54:19
Speaker
Given the importance of the subject matter, free will, it's something that's very important to us as humans. I think we should, as scientists, think about, well, what criteria do we wanna have before we start making declarations about the existence or non-existence of free will? And I think that
00:54:48
Speaker
You know, we expect, for sure you expect a certain amount of purchase on the movement based on what's happening in advance of the movement. There's no doubt the movement from a scientific point of view does not just emerge out of thin air. It comes from the brain and the decision to move now is not completely divorced from what was happening in the brain a few milliseconds or tens or hundreds of milliseconds or even a few seconds earlier. So there is a causal chain of some sort.
00:55:17
Speaker
There's some sort of causal, let's call it a causal milieu, and that if I use machine learning or other techniques, I'm going to pick up on that causal milieu, but that's going to give me a certain level of predictive power that's maybe better than chance, but not perfect by any means, not even close, better than chance we expect.
00:55:42
Speaker
given that interpretation. So what would count as evidence, what would be newsworthy from the point of view of free will is if, yeah, if a second or two or even more in advance of the movement, I could consistently predict what you're going to do and when you're going to do it at 99% correct, we'll allow for something less than perfect because of, let's say, measurement error.
00:56:11
Speaker
then that would require us to take a step back and ask the hard questions about, well, could this possibly be compatible with something like conscious free will? So yes, let's say in theory, it would be possible to come up with evidence that posed a really serious challenge to the idea of conscious free will, but we don't have that kind of evidence yet.
Conclusion and Thanks to Dr. Shurger
00:56:38
Speaker
Well, I think that's a great place to leave things and just being respectful of your time. Aaron, I really appreciate you coming on the show. I think it's a great conversation and thanks a lot. Thanks, Aaron. Thanks very much for inviting me. It's been a pleasure.