WEBVTT 00:00:03.517 --> 00:00:10.497 This is the Convergent Science Network podcast. Leading researchers in the domain 00:00:10.497 --> 00:00:16.777 of neuroscience, brain theory and technology are interviewed by Paul Vershoor and Tony Prescott. 00:00:20.297 --> 00:00:25.617 So this is Paul Vershoor for the Convergent Science Network podcast and we're 00:00:25.617 --> 00:00:29.797 here at the Barcelona Cooperation Brain and Technology Summer School of 2018, 00:00:30.857 --> 00:00:35.077 and as one of our speakers today we have Ancordi Marquos Hello, 00:00:36.437 --> 00:00:42.897 Ancordi so you were talking about two things like you emphasized very much this 00:00:42.897 --> 00:00:52.197 whole idea of how prefrontal cortex represents the memory of goals and initiates actions right so, 00:00:54.457 --> 00:00:59.437 So what do you see as sort of the key feature of these neurons that needs to be explained? 00:01:00.894 --> 00:01:05.434 I think the most important result or the most important thing that we have seen 00:01:05.434 --> 00:01:11.734 with data is that the neural network in the frontal cortex is really heterogeneous. 00:01:11.874 --> 00:01:16.654 So you cannot explain anything just by looking at the average finding rate of 00:01:16.654 --> 00:01:18.534 depopulation because you have nothing there. 00:01:18.754 --> 00:01:24.554 You have really to go inside the individual dynamics to be able to see something 00:01:24.554 --> 00:01:26.794 and to really understand what is going on there. 00:01:27.194 --> 00:01:35.614 Right, so what are the key observations that you then start out from in this exploration? 00:01:36.314 --> 00:01:41.554 So one of the key results is that we found that there are some neurons or some 00:01:41.554 --> 00:01:45.354 group of neurons in the prefrontal cortex which really play an important role 00:01:45.354 --> 00:01:47.694 in shaping the whole dynamics of the network. 00:01:48.094 --> 00:01:52.134 And these neurons are so important because they are very susceptible to the 00:01:52.134 --> 00:01:55.054 input so that they change state very easily. 00:01:55.314 --> 00:02:01.714 And that, but consequently, they can shape the whole network activity and we 00:02:01.714 --> 00:02:06.674 can be flexible about what we do or how we interact with the world. 00:02:07.114 --> 00:02:10.834 While in the same time, we have neurons which are much more stable so that they 00:02:10.834 --> 00:02:16.734 keep track of what we did and which is our internal state or all the information 00:02:16.734 --> 00:02:19.974 that could be relevant for your way of doing it. Right. 00:02:20.574 --> 00:02:26.654 But one of the first experiments that you discussed was actually an interesting one, 00:02:26.714 --> 00:02:32.394 where you looked at how neurons responded to the feedback that the animal received 00:02:32.394 --> 00:02:36.474 after it performed a task, correctly or incorrectly. 00:02:37.334 --> 00:02:41.134 And you showed that if you measure from dorsolateral prefrontal cortex. 00:02:42.274 --> 00:02:45.494 Neurons came actually two flavors, right? 00:02:45.494 --> 00:02:52.114 That some showed an elevated response and others actually did not change the 00:02:52.114 --> 00:02:55.254 response level after the feedback was received. 00:02:55.714 --> 00:03:02.354 But what was curious there is that you showed that these neurons increased their response. 00:03:02.794 --> 00:03:06.874 So I initiate a movement, I get my feedback, I make an error, 00:03:07.014 --> 00:03:11.854 and the neurons, there were neurons that actually elevated their response when 00:03:11.854 --> 00:03:13.994 an error was made, right? 00:03:13.994 --> 00:03:19.634 As opposed to neurons that actually did not change at all in case the response was correct. 00:03:20.114 --> 00:03:23.314 So this was a little bit the starting point to show that, okay, 00:03:23.394 --> 00:03:29.554 there is sort of an internal memory system at work, but it elevates the response when I make a mistake. 00:03:31.634 --> 00:03:33.594 So what's the significance of that? 00:03:35.490 --> 00:03:42.290 Well, in the task in which we did the analysis, actually, we saw that it was 00:03:42.290 --> 00:03:46.570 not really very much there because they won't use that information at all in 00:03:46.570 --> 00:03:49.270 their subsequent behavior. 00:03:49.870 --> 00:03:54.010 But we think that this might be very important in dynamic environments or in 00:03:54.010 --> 00:03:59.890 situations in which you have to be flexible, actually, to adapt to changes. 00:04:00.130 --> 00:04:04.390 So you need to know if you did well or not, because that could have an influence 00:04:04.390 --> 00:04:05.710 in your subsequent behavior. 00:04:06.370 --> 00:04:11.930 So we think that the different neurons are anyway monitoring this information, 00:04:12.750 --> 00:04:18.130 in case there are some changes or in case this could be relevant for something else later on. 00:04:18.750 --> 00:04:25.790 In this case, this elevated response in the firing rate to an error trap. 00:04:26.830 --> 00:04:28.970 It might not be of immediate relevance 00:04:28.970 --> 00:04:35.090 in that task because you're not controlling in any way error, right? 00:04:35.730 --> 00:04:40.590 But could you still argue that maybe for other mnemonic functions this error 00:04:40.590 --> 00:04:46.410 function might play a role in some other process that is not directly translated to task performance? 00:04:47.290 --> 00:04:53.150 Or would you really say irrelevant? No, no. I say irrelevant for what we looked at. 00:04:53.350 --> 00:04:56.130 I don't think it's irrelevant for everything. 00:04:56.470 --> 00:05:01.550 So I think if the prefrontal cortex or the brain in general is monitoring and 00:05:01.550 --> 00:05:07.830 monitoring this information or keeping track of that is because it might be useful at some point. 00:05:08.670 --> 00:05:14.230 For this task in particular, we didn't find any correlation between incorrect 00:05:14.230 --> 00:05:20.350 and the migration of behavior, but it could be relevant if in the same task 00:05:20.350 --> 00:05:22.310 we change some, unexpectedly, 00:05:22.990 --> 00:05:24.910 we change something of the task. 00:05:25.070 --> 00:05:29.070 The more people use this information, the more adaptation. 00:05:29.910 --> 00:05:36.790 So, it's not irrelevant in general, but it's, yeah, maybe the word was not well used. 00:05:36.970 --> 00:05:41.570 It's not irrelevant for everything, but it's not relevant in the way the monk 00:05:41.570 --> 00:05:43.530 is using that information in that task. 00:05:43.630 --> 00:05:45.690 I think it doesn't mean that it's irrelevant. 00:05:48.318 --> 00:05:53.518 So what's interesting as well is that usually there's an interpretation that 00:05:53.518 --> 00:05:59.238 that feedback to the NNL in error is translated to some sort of modulatory signal 00:05:59.238 --> 00:06:00.838 to prefrontal cortex, right? 00:06:00.838 --> 00:06:05.598 Where you would say, well, I got an unexpected reward, dopamine goes up transiently, 00:06:05.598 --> 00:06:07.298 and I should see some change in the response. 00:06:07.558 --> 00:06:12.698 I have an unexpected error, and I should see a transient downregulation of dopamine, 00:06:12.838 --> 00:06:17.858 which should also be translating to a transient response change in the neurons. 00:06:18.518 --> 00:06:23.618 But in some sense, you see something very different now, because we see here 00:06:23.618 --> 00:06:30.538 now that an error translates into an increased response, which would be associated with, let's say, 00:06:31.658 --> 00:06:35.958 release of dopamine, but that doesn't make sense given this standard model. 00:06:36.078 --> 00:06:40.718 The standard model says, no, the not expected error should be a reduction in 00:06:40.718 --> 00:06:41.838 dopamine response, right? 00:06:41.878 --> 00:06:47.578 So, from that perspective of how I've been thinking about how error regulates 00:06:47.578 --> 00:06:49.558 prefrontal cortical responses, 00:06:49.898 --> 00:06:54.798 is this a surprising outcome or you think it's really sort of fixed in the standard 00:06:54.798 --> 00:06:57.538 interpretation of how this system should work? 00:06:58.738 --> 00:07:02.318 For us, it was at first, it was surprising. 00:07:02.698 --> 00:07:05.758 We were not expecting that the activity would increase for the employers, 00:07:06.138 --> 00:07:11.398 but as you said, we would expect it to increase for the, if there is a double 00:07:11.398 --> 00:07:14.078 increase, we were expecting it to increase for the reward delivery. 00:07:14.518 --> 00:07:16.318 But we found actually the opposite. 00:07:21.978 --> 00:07:26.778 We found it, we tried to find some kind of correlation actually between the 00:07:26.778 --> 00:07:31.378 error and and the difficulty of the trial, to say, okay, do we have some correlation 00:07:31.378 --> 00:07:34.298 where that's maybe for more difficult trials? 00:07:35.338 --> 00:07:40.078 You don't have this activation in the neurons because you were just expected 00:07:40.078 --> 00:07:42.378 to be wrong, so you just have nothing. 00:07:42.898 --> 00:07:48.038 But for the easiest trials, for some reason that we cannot really explain yet, 00:07:48.958 --> 00:07:52.138 you have this activation because in some places it's like, hey, 00:07:52.218 --> 00:07:54.898 look, this was wrong, and I was not expecting that. 00:07:54.898 --> 00:08:00.718 But we don't speculate yet because we cannot answer that in which are the mechanisms 00:08:00.718 --> 00:08:05.318 causing this but actually we didn't find either this correlation with difficulty 00:08:05.318 --> 00:08:08.918 so we could not explain that in that way so, 00:08:09.678 --> 00:08:12.298 yeah it was a surprising result and, 00:08:13.138 --> 00:08:19.578 from what we have now we cannot go more into the details of that Is that fair 00:08:19.578 --> 00:08:24.918 to say that the question is a standard model or do you Do you believe it will 00:08:24.918 --> 00:08:30.898 be incorporated in the standard model if you just have a bit more time to look at the details of this? 00:08:31.934 --> 00:08:34.514 I think this should be incorporated into the standard model. 00:08:34.654 --> 00:08:37.034 I don't think it's against the standard model. 00:08:37.214 --> 00:08:40.714 I think this is some feature that should be taken into account. Okay. 00:08:41.874 --> 00:08:47.274 So now we know that we have responses at different timescales. 00:08:47.534 --> 00:08:51.554 We have very systematic responses even after the action has been emitted, 00:08:52.254 --> 00:08:56.174 after we've received our feedback, in this case, that relates to this error, the error component. 00:08:56.614 --> 00:08:59.954 But then sometimes you were delving much more in detail in that, 00:08:59.954 --> 00:09:05.794 And you started to look at, okay, how does the response of these prefrontal cortical neurons, 00:09:06.094 --> 00:09:11.134 if a simple discrimination test, whether to say whether something is closer or further away, 00:09:11.414 --> 00:09:16.034 or whether a time interval is shorter or longer, with the simple discrimination 00:09:16.034 --> 00:09:21.434 test, you started to see that these memory-based responses of these neurons 00:09:21.434 --> 00:09:25.234 were actually not univariate. 00:09:25.234 --> 00:09:28.034 It was not that they were just reflecting one property of the task, 00:09:28.214 --> 00:09:32.054 right? They could also switch, if you want, their response properties. 00:09:33.114 --> 00:09:38.534 So what's really the implication of that? 00:09:38.654 --> 00:09:43.234 So how dramatic is that, this switching of the response properties? 00:09:43.834 --> 00:09:47.734 Do you mean from calling something memory to calling something else afterwards? 00:09:48.134 --> 00:09:55.034 Yeah. How dramatic? Yeah, so this was the experiment that you published in Psychedelic 00:09:55.034 --> 00:09:57.474 Reports in 2016, if I'm correct. 00:09:58.094 --> 00:10:03.894 Yes. So where you show these persistent activities in prefrontal cortex are not committed. 00:10:04.054 --> 00:10:08.094 If you look at the single cell, it's not that the single cell has high fidelity 00:10:08.094 --> 00:10:09.774 to one property of the task. 00:10:10.454 --> 00:10:13.994 It might initially maybe code for what the goal is of the task, 00:10:14.194 --> 00:10:18.114 but at some point it just switches its allegiance and starts to code something else. 00:10:18.134 --> 00:10:20.514 It starts to go in the action you want to transmit. 00:10:22.194 --> 00:10:25.794 So how fundamental is that property, this ability to switch? 00:10:25.794 --> 00:10:31.294 Do you see that in many of these neurons, or is it something that is sort of very, very rare? 00:10:31.914 --> 00:10:35.034 No, actually, this is quite common in different types of neurons. 00:10:35.134 --> 00:10:39.454 So they are not just tuned to one of the properties of the task, but normally they code. 00:10:39.794 --> 00:10:44.934 The normal thing is that they represent more than one feature of the task. 00:10:45.074 --> 00:10:47.974 So it's quite common to find that in the prefrontal code. 00:10:48.334 --> 00:10:51.994 But they code one feature in sequence, right? So at one period of time, 00:10:52.234 --> 00:10:55.434 they code one thing, and at another period of time, they code the other thing? No. 00:10:55.534 --> 00:10:59.174 Or are they more multiplexing? No, it could be overlapped, too. 00:10:59.234 --> 00:11:04.354 So you could have, for instance, in the result that I presented about the transformation 00:11:04.354 --> 00:11:09.094 from goal in memory to the action, there you see that some neurons score the 00:11:09.094 --> 00:11:13.854 goal at the same time as the action. So they overlap. 00:11:14.174 --> 00:11:19.074 They have a bit of delay, but during some period, both things are going by the 00:11:19.074 --> 00:11:26.774 same level. If they switch their representational connotation, 00:11:27.938 --> 00:11:31.378 do you see that also as reflecting the progression of the decision-making? 00:11:32.558 --> 00:11:36.118 Like, initially I have to make up my mind of what are my options, 00:11:36.338 --> 00:11:39.118 what's the relevant evidence, what's memory telling me? 00:11:39.378 --> 00:11:43.798 So at this point in time, you want to dedicate more representation or resource 00:11:43.798 --> 00:11:45.558 to what's the goal I'm pursuing. 00:11:46.078 --> 00:11:50.318 But maybe at some point when you have pruned down your options and you say, 00:11:50.398 --> 00:11:54.858 okay, to achieve the goal, this is now my action, or at least a subset of actions 00:11:54.858 --> 00:11:58.938 I should consider, maybe that in the progression of the decision process, 00:11:59.558 --> 00:12:03.318 I might want to shift my representational resource. 00:12:03.778 --> 00:12:07.298 Is that also how you see that? Yeah, I find this to be correctly, 00:12:07.578 --> 00:12:09.658 yeah, we found it like a sequence of states. 00:12:10.058 --> 00:12:15.498 So first you have something in memory and because of the, because you are iterating 00:12:15.498 --> 00:12:19.398 more and more information now, I should see which action to perform, 00:12:19.698 --> 00:12:25.578 so neurons are switching the tuning or the faring towards that feature because 00:12:25.578 --> 00:12:27.698 it's the one which is most relevant at that point. 00:12:27.898 --> 00:12:30.258 So it's more or less a sequentiality of states. 00:12:30.798 --> 00:12:37.558 Actually, there are some works that are only with Hilleman of Mode that what 00:12:37.558 --> 00:12:38.718 they saw is actually that, 00:12:38.798 --> 00:12:42.798 that we have sequences of states through which the neural population in the 00:12:42.798 --> 00:12:47.798 prefrontal cortex and also in other cortical areas Yes, also through sequences of states, 00:12:48.198 --> 00:12:53.238 so that they are adapting to the new events or new features coming in. 00:12:53.498 --> 00:12:58.898 But does it reflect only that new information comes in, or does it also reflect 00:12:58.898 --> 00:13:00.958 the progression of the decision-making process? 00:13:01.638 --> 00:13:04.858 Both. Also the progression of the decision-making process. 00:13:04.938 --> 00:13:10.278 You can have changes in the states as the decision is being made, 00:13:10.438 --> 00:13:13.018 and also the causes are intervening on some new information. 00:13:13.138 --> 00:13:14.238 So both cases are different. 00:13:15.164 --> 00:13:21.084 Right. So, most of your neurons are in dorsolateral prefrontal cortex, right? 00:13:22.524 --> 00:13:27.344 Do you think that this is a generic feature also if you go to medial frontal cortex? 00:13:27.604 --> 00:13:33.404 Or would you feel it's really a more specialized property of dorsolateral prefrontal cortex? 00:13:34.324 --> 00:13:44.064 Well, this I cannot respond with a clear answer because we didn't know about that. 00:13:44.064 --> 00:13:49.784 But my guess would be that you might find it also in a way in the asymptotic 00:13:49.784 --> 00:13:51.844 or the saloprilage frontal cortex. 00:13:52.184 --> 00:13:57.944 I think it's more likely that they call it a ticking voice in the brain of some changes. Right. 00:13:58.484 --> 00:14:03.084 But so, then in this first experiment, you started to focus more also on how 00:14:03.084 --> 00:14:08.084 populations of neurons are actually then tuning themselves to the task, right? 00:14:08.084 --> 00:14:15.824 And from that came also this idea that some populations are showing this progression 00:14:15.824 --> 00:14:20.704 in sort of a task-dependent way, and other populations do not show this progression. 00:14:23.504 --> 00:14:29.444 When you show these groups of neurons that you were measuring from, 00:14:29.824 --> 00:14:35.064 then you have the pre-go and post-go goal neurons. 00:14:35.064 --> 00:14:41.224 Neurons, but some of those are actually switching their representational state, and others do not. 00:14:42.364 --> 00:14:47.184 So, how should I think about that? How should I interpret that? What does that mean? 00:14:47.424 --> 00:14:51.944 Also, in the face of what we just discussed, that there is this sort of progression of decision-making, 00:14:52.284 --> 00:14:58.424 the resource is reallocated in some sense, but should I look at this result 00:14:58.424 --> 00:15:03.464 as telling us that, okay, I have a bunch of neurons in the prefrontal cortex? 00:15:04.595 --> 00:15:08.975 In the end, they're multiplexing because across all these neurons you will find 00:15:08.975 --> 00:15:14.975 that they represent all possible aspects of the task and all possible combinations of these aspects. 00:15:15.215 --> 00:15:17.295 This is roughly what it means, right? 00:15:17.835 --> 00:15:24.055 Some of these will then again switch what they represent, and others will not 00:15:24.055 --> 00:15:25.695 switch what they represent, right? 00:15:25.755 --> 00:15:29.555 So there's a fixed representational substrate, and others will more dynamically 00:15:29.555 --> 00:15:31.595 allocate themselves to that. 00:15:32.035 --> 00:15:34.875 So how should How should we interpret that? That there are these neurons that 00:15:34.875 --> 00:15:37.735 are stupid, they're dynamic, and others are smarter and they're dynamic? 00:15:38.035 --> 00:15:46.555 Or how do you interpret that? How we interpret it is, we have two main key features 00:15:46.555 --> 00:15:47.435 in the prefrontal cortex. 00:15:47.935 --> 00:15:51.775 One are composed by neurons, which are very stable in different ways, 00:15:51.835 --> 00:15:57.035 so they are keeping track of what is going on, of the information which is important 00:15:57.035 --> 00:16:00.435 for something, for the decision, for attention, for whatever. 00:16:00.435 --> 00:16:02.375 There, so they are very stable and robust. 00:16:02.715 --> 00:16:07.515 And then we have a second group of neurons, which are flexible because in the 00:16:07.515 --> 00:16:10.535 end, you cannot just stick to something and that's it, 00:16:10.635 --> 00:16:16.575 you need to be flexible to be able to perform an action or to list or to get 00:16:16.575 --> 00:16:18.355 more information or whatever. 00:16:18.555 --> 00:16:22.695 So these two dynamics are very important in the brain because one is telling 00:16:22.695 --> 00:16:25.075 us, okay, this is the information we have. 00:16:25.155 --> 00:16:30.015 And the other is saying, okay, but now we need to move from that information. We need to do something. 00:16:30.435 --> 00:16:34.775 So the other ones, the ones which are shaping the whole nerve or any different 00:16:34.775 --> 00:16:38.595 dark cortex to say, okay, now is the time to perform an action. Right. 00:16:38.995 --> 00:16:42.835 But wouldn't it be fair then to say that there is one set of neurons, 00:16:43.175 --> 00:16:47.395 let's say 50% of the population, gives you like a ground truth and says, 00:16:47.495 --> 00:16:49.455 this is what is out there in the world. 00:16:50.035 --> 00:16:52.655 This is what my needs are and my goals. 00:16:53.315 --> 00:16:56.615 And I'm not going to, this is the blackboard in which I'm going to operate now. happen. 00:16:57.035 --> 00:17:00.355 And they have another representational substrate or another set of neurons that 00:17:00.355 --> 00:17:04.155 then say, okay, now if I want to come to a solution of this, 00:17:04.255 --> 00:17:05.815 the other things I should be doing. 00:17:06.774 --> 00:17:10.594 So it sort of moves away from that ground truth and now starts to sort of prune, 00:17:10.734 --> 00:17:13.194 you know, okay, but this isn't the specific cue that matters. 00:17:13.534 --> 00:17:16.514 And this isn't the specific thing I should be doing and I should ignore the 00:17:16.514 --> 00:17:18.674 other stuff. Is this roughly how you would think about that? 00:17:19.354 --> 00:17:22.774 Not exactly. Because we are talking always about two loop of neurons, 00:17:22.974 --> 00:17:24.134 but these are two extremes. 00:17:24.594 --> 00:17:28.734 We believe that in the prefrontal cortex we have a continuum of neurons. 00:17:28.994 --> 00:17:33.394 So that means that it's not that we have only a loop of neurons that are very 00:17:33.394 --> 00:17:35.594 stable and then one which is very flexible. 00:17:36.154 --> 00:17:41.374 But that we have also something in the middle that goes from stability to flexibility. 00:17:41.934 --> 00:17:45.894 So, when I talk about two groups, I'm always talking about the same cases, 00:17:46.094 --> 00:17:50.954 but we believe that we have neurons doing things also in the middle range. 00:17:51.454 --> 00:17:54.774 Okay, that is true, because indeed, in your data, you filtered quite a bit, 00:17:54.834 --> 00:17:58.354 right? In the end, we only looked at the no-switchers or the switchers. 00:17:58.474 --> 00:18:01.914 Exactly, because these were the strengths of the model. All right, 00:18:01.954 --> 00:18:06.214 so it's interesting, because it means in all cases you represent all possible 00:18:06.214 --> 00:18:08.174 aspects of the task and all its combinations, 00:18:08.574 --> 00:18:13.874 but then also you represent all possible transformations of that over time. 00:18:14.054 --> 00:18:15.434 Is that what you're saying? 00:18:15.874 --> 00:18:19.754 So you're expanding the whole search space, actually. Yeah, yeah. 00:18:21.934 --> 00:18:28.254 So then… For instance, with our model, we actually fitted the data that we were 00:18:28.254 --> 00:18:32.454 looking at, but remember that we were always looking at 1,000 neurons, 00:18:32.734 --> 00:18:37.554 and we are selecting a class 2 population of them because they are doing what we are looking at. 00:18:37.934 --> 00:18:41.394 But of course, we have many different groups of neurons, and many neurons doing 00:18:41.394 --> 00:18:42.714 all of this stuff. Right. 00:18:43.994 --> 00:18:47.974 Exactly. No, this is, I understand. Okay. So it is, so we have a hyper, 00:18:48.114 --> 00:18:53.514 high dimensional representation of your task and its changes, right? 00:18:53.554 --> 00:18:56.234 So that you would believe there's a bit of a problem. 00:18:56.314 --> 00:18:58.734 How do I now select among all these options? 00:18:58.914 --> 00:19:01.414 How do you see the selection taking place? 00:19:02.054 --> 00:19:06.894 And this also brings us a little bit to the model that you built of this, right? 00:19:06.974 --> 00:19:13.554 So now I have this high dimensional space with all sorts of combinations of features of the task. 00:19:15.318 --> 00:19:22.658 One is sort of constant, and the other is dynamically now, if you want, 00:19:22.818 --> 00:19:26.098 trying options and weighing them, right? 00:19:26.218 --> 00:19:33.318 So how do you see that process play out? 00:19:35.338 --> 00:19:39.218 How should I imagine that this actually can lead to even a decision? 00:19:40.298 --> 00:19:45.678 That it can converge to a decision on an action I should perform? All these neurons. 00:19:45.778 --> 00:19:49.758 Yeah. This whole process, this whole dynamo process that you mentioned, Probe. 00:19:49.858 --> 00:19:55.318 Well, they are, I mean, I could see that as, I mean, do you imagine that the 00:19:55.318 --> 00:19:58.558 neurons which are stable and not falling apart are hitting back on the information 00:19:58.558 --> 00:20:00.718 which might be important at that point? 00:20:00.878 --> 00:20:05.498 And you see the other group of the flexible group getting more information from 00:20:05.498 --> 00:20:09.598 outside, but also from our internal motivation, attention, and everything. 00:20:09.598 --> 00:20:17.698 And you see this is a couple of connected areas and neurons inside the frontal corpus. 00:20:18.038 --> 00:20:24.818 As I show in the presentation, just by connecting it with some excitation and emission, 00:20:25.098 --> 00:20:30.578 you would get this selection at the end, because with the flexibility of the 00:20:30.578 --> 00:20:36.278 neurons, what you get is that they decide in a way which part or which neurons 00:20:36.278 --> 00:20:37.898 will be responding next, 00:20:38.198 --> 00:20:41.558 which in the end will lead to the other bit action to perform. 00:20:42.778 --> 00:20:49.018 So you propose a model where you say, well, you can think about this as a system 00:20:49.018 --> 00:20:55.778 of competing pools of neurons that are dedicated to certain aspects of a task, 00:20:55.878 --> 00:20:57.278 let's say, a goal that I pursue. 00:20:59.518 --> 00:21:03.158 Then these neurons have the capability to maintain the state, 00:21:03.238 --> 00:21:05.458 like a memory, like it's It's a mango leaf system. 00:21:06.078 --> 00:21:11.238 And they have to calculate to rapidly switch. They have to calculate because they're by state. 00:21:11.898 --> 00:21:16.298 By virtue of how they're wired, they can maintain multiple states and they can 00:21:16.298 --> 00:21:17.258 switch from one to the other. 00:21:18.622 --> 00:21:26.862 But now every pool is dedicated to one feature of the task, like the goal, or a cue, or an action. 00:21:28.122 --> 00:21:35.122 And then you show that you could then get this switching behavior by wiring 00:21:35.122 --> 00:21:38.202 these pools of neurons up in a sort of a smart way. 00:21:38.202 --> 00:21:44.302 That if I get a non-specific excitation across all these pools of neurons that 00:21:44.302 --> 00:21:49.782 I can now flip, that I can sort of go for QA or QB or response A or response B. 00:21:49.822 --> 00:21:52.442 You're correct? Yeah. Okay, cool. 00:21:53.242 --> 00:21:56.162 And you demonstrated to us that this can work. 00:21:56.542 --> 00:22:02.382 But you demonstrated that in a very low-dimensional space because now we just looked at one queue. 00:22:03.982 --> 00:22:09.282 But what we looked at earlier or as you discussed it earlier in your physiology, 00:22:10.082 --> 00:22:13.762 you see that actually you code all the features of the desk and all their combinations 00:22:13.762 --> 00:22:15.862 and it's changing over time. Thank you. 00:22:17.190 --> 00:22:21.910 So how do you see that model then scale? Because every pool of neurons in your 00:22:21.910 --> 00:22:24.750 model is one of these, right? 00:22:24.830 --> 00:22:28.750 One of these cues or one of these cue combinations, right? 00:22:29.270 --> 00:22:34.550 So do you see that scaling being feasible for your model? 00:22:34.810 --> 00:22:39.070 Okay. So each pool is not exactly only one feature of the task. 00:22:39.830 --> 00:22:46.750 Because the pools that I showed in the presentation are actually modules of a network. 00:22:47.190 --> 00:22:53.710 And each model is composed by eight groups of excitatory neurons and one group of inhibitory neurons. 00:22:54.230 --> 00:22:59.250 So each of this group is responding to one, for instance, if they are coding the goal. 00:22:59.410 --> 00:23:04.030 So if it's a model dedicated to the goal, they would be coding eight different goals. 00:23:04.770 --> 00:23:08.910 So each pool would be responsible for being selected for a different goal. 00:23:09.130 --> 00:23:12.330 And the same happens with the other model. So inside each model, 00:23:12.570 --> 00:23:18.490 we have a brain selectivity for eight different variables of the same feature. 00:23:18.970 --> 00:23:25.250 So I think it's very realistic to scale it up because it will be taken with 00:23:25.250 --> 00:23:26.890 the model that we presented today. 00:23:26.970 --> 00:23:30.550 We already have many features covered. 00:23:30.950 --> 00:23:39.610 Like if each model is actually coding the eight different variables of a goal. 00:23:39.790 --> 00:23:45.890 So it's not only red and blue, like in our task, they could also code three, yellow, till eight. 00:23:46.190 --> 00:23:50.210 So actually the model is already capable of explaining quite well. 00:23:50.210 --> 00:23:53.630 So the scaling would become problematic when I start to add goals then. 00:23:53.990 --> 00:23:58.610 So you have the grouping around goals that would be a critical feature. 00:24:00.610 --> 00:24:06.750 As long as this population is linked to that goal, I can capture the whole set. 00:24:07.350 --> 00:24:12.670 If there are no more than eight goals, we can explain the data. Okay. 00:24:13.290 --> 00:24:15.190 That would be unbelievable, you say. Yeah. 00:24:15.990 --> 00:24:22.830 But the other thing that you can think of is that your model is a labeled line kind of model, right? 00:24:22.950 --> 00:24:30.510 So the synapses I get must be uniquely linked to some feature, right? 00:24:30.570 --> 00:24:33.990 Otherwise, the dynamics cannot work, right? 00:24:34.390 --> 00:24:39.550 So it would also mean every neuron is dedicated to just a sort of permanently 00:24:39.550 --> 00:24:41.830 dedicated to some aspect of a task. 00:24:43.210 --> 00:24:48.210 But isn't it an important feature of a working memory system of prefrontal cortex 00:24:48.210 --> 00:24:51.350 that you can dynamically be allocated to any task? 00:24:52.270 --> 00:24:59.250 So how would that work? I have my pools of neurons and now I'm doing a discrimination task, but maybe. 00:25:00.906 --> 00:25:04.506 The next hour, I have to do a navigation task, or I have to do an operative 00:25:04.506 --> 00:25:07.306 conditioning task, or whatever, I have to deal with other kinds of problems. 00:25:07.666 --> 00:25:11.926 So how can you have this flexible allocation of the content, 00:25:12.086 --> 00:25:13.166 if you want, the semantics? 00:25:13.546 --> 00:25:18.506 Well, I think this model could be pretty general, because you just change your length with, 00:25:19.286 --> 00:25:23.286 time, the new task, and then you change your semantic activity between the modules 00:25:23.286 --> 00:25:28.246 or from the input that you get, and then you will have four different kinds 00:25:28.246 --> 00:25:30.146 of behavior with the same model. Really? What? 00:25:30.906 --> 00:25:36.186 So I think if we... Yeah, monkeys were really well trained in these distance discrimination tasks, 00:25:36.426 --> 00:25:39.946 but if they would switch to a different task, I think this would be learned 00:25:39.946 --> 00:25:44.766 in the synaptic connections of the module, and then you would have the same 00:25:44.766 --> 00:25:46.906 data index that would be... 00:25:48.446 --> 00:25:53.326 But actually, what we wanted to do was... I mean, the whole view that I presented 00:25:53.326 --> 00:25:58.446 there was more like a truth of concept to say, okay, we have this... 00:25:58.446 --> 00:26:04.926 If we have these heterogeneous neurons in the brain, as we think we have, 00:26:05.146 --> 00:26:08.646 we can't really explain it as we observe. 00:26:09.006 --> 00:26:14.006 So it's more like a proof of concept to say we have a brain and we have an area 00:26:14.006 --> 00:26:18.586 which is full of heterogeneity in the excitability of the neuron. 00:26:18.946 --> 00:26:26.046 Right. So, but now the scaling along goals, if I have goals that are competing 00:26:26.046 --> 00:26:30.206 with each other, how would I account for that in this system? 00:26:31.441 --> 00:26:35.041 You have opposing goals, opposing goals that are contradictory. 00:26:35.801 --> 00:26:39.201 Okay, so there you would have a competition. So this is in the end, 00:26:39.241 --> 00:26:41.081 each model is an attractor model. 00:26:41.721 --> 00:26:45.501 So you would have within a model, you would have two pools which are. 00:26:48.941 --> 00:26:52.501 Stimulated, and then they would compete. And in the end, you will have one winning 00:26:52.501 --> 00:26:56.281 goal and one that leads to the final execution. Okay. 00:26:57.761 --> 00:27:01.041 So in this case, indeed, so it's like an attractor model? 00:27:02.681 --> 00:27:07.401 We're looking at firing rates of the model, and in some sense you're interpreting 00:27:07.401 --> 00:27:12.701 or trying to predict performance and firing rate, but in the earlier work, 00:27:12.841 --> 00:27:17.561 a very influential paper that you, which you were the first author, 00:27:18.021 --> 00:27:22.921 you actually made the point that firing rate is not that informative about performance, 00:27:23.081 --> 00:27:29.321 it's much more the variability of the responses that really matters if tasks are more complex. 00:27:30.481 --> 00:27:38.121 Did you step away from that view? Are you back now more into the rate-coding view of things? 00:27:38.501 --> 00:27:42.501 No, I'm very interested in studying the variability or which information we 00:27:42.501 --> 00:27:44.021 can extract from the variability. 00:27:44.961 --> 00:27:48.741 In this case, I didn't look… well, actually, we did something with variability, 00:27:49.121 --> 00:27:53.301 but we calculated it by looking at the bars and pulses in our data. 00:27:53.301 --> 00:27:59.881 Data because in the models that we presented, we have the stable models are 00:27:59.881 --> 00:28:03.361 the ones which are very stable and they don't have variability or almost nothing, 00:28:03.541 --> 00:28:04.941 not variability in the framework, 00:28:05.121 --> 00:28:09.541 whereas the ones which are more flexible have much more variability. 00:28:10.601 --> 00:28:14.201 And we tested that in the data, looking at the bursts and pauses, 00:28:14.401 --> 00:28:18.761 and we found that actually we could, yeah, this prediction was right. Right. 00:28:19.001 --> 00:28:23.261 What we don't have is anything with the behavior there because we would not 00:28:23.261 --> 00:28:27.281 actually, I tried to look at that, but we didn't have any good. 00:28:27.481 --> 00:28:30.101 In this step, the planning is that we don't have reaction time. 00:28:31.541 --> 00:28:35.961 So that's very limited, because whenever the targets appear, 00:28:36.681 --> 00:28:41.101 the monkey can perform the action, because he already knows what to choose. 00:28:41.561 --> 00:28:45.941 So we didn't have reaction time, so we could not correlate it with variability 00:28:45.941 --> 00:28:48.521 or with any other measure of difficulty. 00:28:49.301 --> 00:28:55.101 So that's why this is time we didn't look at that, because I didn't have the possibilities. 00:28:55.761 --> 00:29:01.161 But I think variability, and we proved that also in MOLLE, is very informative. 00:29:01.541 --> 00:29:04.781 It's one of the key features of this flexible model. 00:29:05.261 --> 00:29:11.161 So are you saying that are you moving away from the standard model of distribution 00:29:11.161 --> 00:29:12.581 models of decision-making? 00:29:12.741 --> 00:29:17.201 Because the standard view would be, look, I just integrate whatever information 00:29:17.201 --> 00:29:21.041 I get, usually it's perceptionally evident, until I hit threshold, 00:29:21.181 --> 00:29:22.921 or I hit threshold first, it's the winner. 00:29:23.261 --> 00:29:28.961 So how are you moving away from that standard view on the decision-making process? 00:29:29.361 --> 00:29:34.341 I think this standard model is sad, very nice and they are very useful to actually 00:29:34.341 --> 00:29:37.181 to have some predictions about which behavior you might encounter, 00:29:37.681 --> 00:29:44.221 but I think they are not so good to understand to really spending dynamics of the of a network, 00:29:44.821 --> 00:29:49.641 So you there you have an intuition of what you would expect as an average activity, 00:29:50.201 --> 00:29:54.061 But with the diffusion model, you cannot know what would be the individual dynamics 00:29:54.061 --> 00:29:56.001 because they don't do predictions on the. 00:29:56.792 --> 00:30:00.092 So, I think the models are okay, but they have their limitations, 00:30:00.512 --> 00:30:02.952 so we must know how to use them. 00:30:03.352 --> 00:30:07.672 So, because I'm more interested in the individual dynamics, so to understand 00:30:07.672 --> 00:30:14.232 how neurons respond individually, I'm more interested in the cycling ground 00:30:14.232 --> 00:30:17.792 models, because there you can really account and explain what you observe. 00:30:18.092 --> 00:30:22.452 You know, it's still the proponents of the rate of the rate-coded diffusion 00:30:22.452 --> 00:30:29.452 models is also the memory potential refining rate of individual cells that are 00:30:29.452 --> 00:30:31.912 predictive of performance, right? 00:30:31.952 --> 00:30:38.792 Because I can see that the flow that reflects integration is correlated with, 00:30:38.972 --> 00:30:43.232 let's say, the direction times I get or with the accuracy that I get, right? 00:30:43.272 --> 00:30:46.132 So the claim would be made. 00:30:46.332 --> 00:30:51.192 Yeah, yeah, I know. But I think it's one of the, my point of view is more the 00:30:51.192 --> 00:30:52.232 average of the population. 00:30:52.632 --> 00:30:56.152 I don't think you can explain that with a single neuron. I know that there are 00:30:56.152 --> 00:30:59.152 some people that say that you can and you see that in the individual neurons, 00:30:59.352 --> 00:31:02.472 but there are also some recent papers that show that actually, 00:31:03.012 --> 00:31:09.792 the ramping activity is an artifact of averaging different state activity with many neurons. 00:31:10.452 --> 00:31:14.052 So I don't think you can explain the individual dynamics of that. 00:31:14.192 --> 00:31:19.812 I think you can have an intuition of the main activity the data must be what it is. 00:31:20.072 --> 00:31:24.552 But I could then still say, well, maybe what matters is the population response, 00:31:24.892 --> 00:31:29.112 and as long as I can predict performance from the population response, I'm happy. 00:31:29.252 --> 00:31:32.112 I don't care for these single cells, because they're all averaging out. Just noise. 00:31:32.412 --> 00:31:36.172 Yeah, but then I would say, why do we have so many neurons? With one neuron, we wouldn't be enough. 00:31:38.472 --> 00:31:44.312 I often feel like that. Yes, okay. So, yeah, I think if we have so many neurons, 00:31:44.452 --> 00:31:50.272 according to so many different ways, because of something that I must be able to... 00:31:50.272 --> 00:31:52.752 No, wait, don't you say I set the problem because I can say, 00:31:52.792 --> 00:31:58.792 well, maybe I have all this variability in neurons so the average has some stability 00:31:58.792 --> 00:32:04.732 or follows some distribution and that average is then what really drives the behavior. 00:32:07.172 --> 00:32:11.592 But, yeah, you could say that but I think, for instance, in the paper that we 00:32:11.592 --> 00:32:13.112 have together much more so influenced. 00:32:14.152 --> 00:32:19.492 We could prove that that's not always the case. so that the inactivity is not turning everything. 00:32:20.232 --> 00:32:23.712 Well, that's the other point, of course, right? That maybe the point is, 00:32:23.792 --> 00:32:25.772 if you have a simple task, you have a simple code. 00:32:26.312 --> 00:32:31.912 And rates are great. But if a task gets more complex, like discrimination, 00:32:33.072 --> 00:32:38.772 if I ask, right, then you have to switch to a different coding model. 00:32:39.512 --> 00:32:43.392 Maybe, but what is that simple? What is a simple task? Because in a way, 00:32:43.432 --> 00:32:46.672 everything is complex. I mean, discrimination they should tax. 00:32:48.285 --> 00:32:55.125 Well, if I have a random motion display with predominant motion in one direction, 00:32:55.365 --> 00:32:58.705 I just integrate over this whole field. I just need the Monmic Integrator. 00:32:59.305 --> 00:33:04.065 And okay, we'll go one direction or the other. I don't need to compare anything, I just integrate. 00:33:04.405 --> 00:33:07.305 And just do reintegrating motion. So that's a simple class. 00:33:07.985 --> 00:33:13.865 If I need to compare two stimuli, or I have to compare two time intervals, 00:33:14.365 --> 00:33:18.465 I already demand more for my memory. you have to keep something in memory, 00:33:18.545 --> 00:33:19.425 you have to compare things. 00:33:20.245 --> 00:33:29.885 So this might be another argument. Yeah, but still I think we are much more optimal than that. 00:33:30.125 --> 00:33:35.745 So why would you have so many neurons doing the same if you could have just one doing that? 00:33:36.665 --> 00:33:39.685 Because in the end it's energy that we should not... 00:33:40.605 --> 00:33:48.045 That's true redundancy or you can also argue well another perspective is of 00:33:48.045 --> 00:33:52.365 course it goes into the simple task but if it's a simple task you have simple cues one cue. 00:33:53.765 --> 00:33:58.085 But if you look at the world in which his brains evolve everything is ambiguous, 00:33:59.085 --> 00:34:04.685 predictability is an issue the Markovian assumption doesn't hold the future 00:34:04.685 --> 00:34:08.705 might be somewhat different from the past you have right so, 00:34:10.005 --> 00:34:15.425 maybe if we have these highly controlled and reduced paradigms the force also 00:34:15.425 --> 00:34:19.665 leads to a very reduced perspective on the complex of the system that we are 00:34:19.665 --> 00:34:25.225 in so maybe just and then you are in something already moving in a different 00:34:25.225 --> 00:34:28.385 direction but there's always a question do you want to still. 00:34:30.465 --> 00:34:35.245 Park it inside the standard model do you want to move away from this idea that 00:34:35.245 --> 00:34:38.605 firing rates integration integration, explains everything. 00:34:39.345 --> 00:34:43.905 Yeah, no, I actually would like to move to, I mean, it's not that I don't agree 00:34:43.905 --> 00:34:46.565 with these models, as I say, they can be useful for some things, 00:34:46.645 --> 00:34:47.865 but I think they have their limitations. 00:34:48.445 --> 00:34:51.985 And it's better that we open to different options and that we really try to 00:34:51.985 --> 00:34:53.665 understand the digital dynamic. 00:34:55.725 --> 00:35:03.085 Right. That's great. So, Ingrid, you, of course, also then delayed full disclosure. closure. 00:35:03.605 --> 00:35:07.705 You've been a member of SPECS here, right? We've been working together for quite a while. 00:35:09.025 --> 00:35:12.685 We mainly looked at models of the brain. 00:35:13.365 --> 00:35:19.365 Then you went into neurophysiology of the monolingual primate and mycotics with all the genovesial. 00:35:20.145 --> 00:35:24.965 And now you start to stand on your own legs at the Neuroscience Institute in Alicante, right? 00:35:25.045 --> 00:35:28.685 So you have to make, you've made this tour now, you're sort of a young researcher, 00:35:28.845 --> 00:35:29.945 you're building your career. 00:35:30.405 --> 00:35:35.945 But But given your experience, what would be a Carni's law that we should follow 00:35:35.945 --> 00:35:38.945 to understand the brain? Carni's law? Mm-hmm. 00:35:40.976 --> 00:35:44.516 That I mean that it's fundamental 00:35:44.516 --> 00:35:50.696 to use a combined experimental and theoretical approach because I think with 00:35:50.696 --> 00:35:55.456 only experience we cannot understand the brain because we have only observations 00:35:55.456 --> 00:36:01.036 and with observation it's okay but yeah we don't know about the insights, 00:36:02.016 --> 00:36:07.016 and with the theory we can really stay that but theory alone would be also not 00:36:07.016 --> 00:36:10.756 so good because because then we don't have anything to prove that what we are 00:36:10.756 --> 00:36:12.056 saying is actually correct. 00:36:12.756 --> 00:36:17.116 So for me and in my future life, I would like to continue combining both, 00:36:17.796 --> 00:36:21.216 because I think that both are fundamental for us to advance in understanding. 00:36:22.296 --> 00:36:26.776 Great. Then I will come down to Elegant a few years from now, 00:36:27.816 --> 00:36:33.916 to check the state of the art in your lab and to see whether you managed to 00:36:33.916 --> 00:36:38.256 really falsify or verify the key hypothesis. 00:36:39.156 --> 00:36:44.176 So what's the key hypothesis that you would like to see tested in this time frame of four years? 00:36:45.476 --> 00:36:48.556 One that I'm still really willing to, 00:36:49.416 --> 00:36:57.356 show since I ended my PhD actually is that the variability in the firing rate 00:36:57.356 --> 00:37:01.636 of the neurons are causing the uncertainty on the differences in modalities 00:37:01.636 --> 00:37:05.096 and then this is read out by a second state, 00:37:05.976 --> 00:37:11.956 process, which is actually computing the confidence based on this diverse uncertainty 00:37:11.956 --> 00:37:14.516 that is called in the five minutes. 00:37:17.136 --> 00:37:22.316 So this is the key thing where I would like to continue and to be advised. 00:37:22.956 --> 00:37:27.816 And I hope that in some years from now I will have an answer to that and I can 00:37:27.816 --> 00:37:31.696 really prove that it is very, very important. 00:37:31.816 --> 00:37:36.896 It's not only about the filing rate and that we should start working on different things now. 00:37:37.336 --> 00:37:41.316 Fantastic. Connie Marcos, thank you very much for this conversation. Thank you. 00:37:45.976 --> 00:37:51.776 The CSN Podcast was produced by the Convergent Science Network of Biometrics 00:37:51.776 --> 00:37:58.196 and Biohybrid Systems, a project funded by the European Sevens Research Framework Program. 00:37:59.696 --> 00:38:05.036 For more interviews, recorded lectures, or upcoming conferences in the field 00:38:05.036 --> 00:38:11.296 of biometrics and biohybrid systems, go to csnnetwork.eu. 00:38:12.076 --> 00:38:13.456 And thank you for listening.