WEBVTT 00:00:00.017 --> 00:00:07.037 This is all for sure with Allard Roebroek, who was one of the speakers at our summer school. 00:00:07.557 --> 00:00:13.097 And Allard gave this summary of an incredible development, 00:00:13.197 --> 00:00:19.337 actually, in the deployment of new and also pretty advanced tools in understanding 00:00:19.337 --> 00:00:25.137 these large-scale data sets that are becoming available in particular for different forms of imaging. 00:00:25.617 --> 00:00:29.717 So why is this important? Why should we be doing this? 00:00:30.877 --> 00:00:34.757 So there's two levels to that so I think why 00:00:34.757 --> 00:00:37.517 should we be doing neuroimaging is what we can say first 00:00:37.517 --> 00:00:43.597 is that because it gives us an unprecedented view in vivo on the brain with 00:00:43.597 --> 00:00:47.897 whole brain coverage and high spatial resolution especially with fMRI and with 00:00:47.897 --> 00:00:52.657 very high temporal resolution with EEG and MEG so neuroimaging I think might 00:00:52.657 --> 00:00:55.057 be clear to some of the listeners to the podcast. 00:00:55.817 --> 00:01:00.557 But what we've also been learning with neuroimaging is that this huge amount, 00:01:00.777 --> 00:01:04.717 this huge bandwidth of data that comes our way out of the MRI machines, 00:01:04.877 --> 00:01:07.997 like gigabytes per minute, basically, 00:01:08.557 --> 00:01:13.277 many, many voxels, tens, hundreds of thousands of voxel time series come out of it. 00:01:13.677 --> 00:01:16.757 We need a good way to understand that. We need a good way to model it. 00:01:17.157 --> 00:01:23.057 And just doing what some people have accusingly called the new phrenology, 00:01:23.097 --> 00:01:26.797 just saying like, oh, neuroimaging tells us that this region lights up when 00:01:26.797 --> 00:01:30.477 you see faces and this region lights up when you do that isn't good enough anymore. 00:01:30.637 --> 00:01:35.597 We need to understand this data in terms of network models, in terms of models that tell us. 00:01:36.300 --> 00:01:39.840 Which region does what, but also how it sends information on to other regions. 00:01:40.420 --> 00:01:46.040 And so that necessitates these complex models, that necessitates that you have 00:01:46.040 --> 00:01:49.580 a network model of what a task might do, how it's coupled to the neuroimaging 00:01:49.580 --> 00:01:53.840 data, because there is your neural network model at one level that might have 00:01:53.840 --> 00:01:55.380 an explanation of how the task works. 00:01:55.620 --> 00:02:01.500 You need to couple that to your data, which is not mere normal neural network activity. 00:02:01.700 --> 00:02:03.640 It's usually some kind of form of 00:02:03.640 --> 00:02:09.320 metabolics and fMRI and some form of synchronized activity in EEG and MEG. 00:02:09.500 --> 00:02:12.880 So you need to model all of that and then try to go in the reverse direction 00:02:12.880 --> 00:02:18.260 from the data to the neural activity to do useful things, even more useful things 00:02:18.260 --> 00:02:20.620 with neuroimaging data than we have been doing. Right. 00:02:21.040 --> 00:02:26.420 Now, in this, what you call modeling, you distinguish two approaches, right? 00:02:26.480 --> 00:02:30.720 So on the one, you have forward models or you have bottom-up models and you 00:02:30.720 --> 00:02:33.100 have these top-down models. So what's exactly the distinction there? 00:02:33.540 --> 00:02:38.540 Right. So it's important to say that this terminology is used and not used in many fields. 00:02:38.740 --> 00:02:42.200 So what I tried to do in the talk is to make clear that my definition is as follows. 00:02:42.680 --> 00:02:47.700 And it's used a bit wider in cognitive neuroscience and computational neuroscience. 00:02:47.700 --> 00:02:53.000 So bottom-up models are first principle models, often biophysically inspired, 00:02:53.200 --> 00:02:55.440 of how a task is performed, 00:02:55.940 --> 00:03:01.060 how neurons fire, for instance, or how whole groups of neurons summarized in 00:03:01.060 --> 00:03:06.480 neuronal mass models generate their activity and pass activity on to other neural masses. 00:03:07.232 --> 00:03:12.752 And then, for instance, how cerebral blood flow, cerebral blood volume, 00:03:13.052 --> 00:03:18.172 deoxygenation levels, hemoglobin levels change because of that. 00:03:18.272 --> 00:03:20.752 All of that you can simulate in the normal causal direction. 00:03:20.772 --> 00:03:23.492 So from neural activity to the 00:03:23.492 --> 00:03:27.192 changes in blood flow, blood volume and oxygenation to the fMRI signal. 00:03:27.252 --> 00:03:31.092 That would be bottom-up modeling or forward modeling. And that would be what 00:03:31.092 --> 00:03:35.292 normal computational neuroscience does a lot, except maybe for also simulating 00:03:35.292 --> 00:03:39.872 the fMRI model, the fMRI data, initially at least, but also lately, 00:03:40.012 --> 00:03:41.632 people got very interested in that. 00:03:41.872 --> 00:03:46.172 So the other type of modeling, top-down modeling, is inverse modeling, 00:03:46.252 --> 00:03:50.192 in the normal sense of the word, is that because we go anti-causal, 00:03:50.232 --> 00:03:51.252 we go in the reverse direction, 00:03:51.412 --> 00:03:55.312 we go from the data that we have to what we're actually interested in, 00:03:55.412 --> 00:03:57.232 which is the neural activity in the brain. 00:03:57.232 --> 00:04:00.432 And of course, for that, we need to go in the inverse direction, 00:04:00.512 --> 00:04:04.732 in the anti-causal direction, which means we have to invert all of those forward 00:04:04.732 --> 00:04:08.572 models that we have built up from how our signal is generated. 00:04:08.952 --> 00:04:12.292 So that's the two types of modeling that I distinguished in that way. 00:04:12.392 --> 00:04:16.852 And I do realize that top-down and bottom-up is used differently in different fields. 00:04:17.512 --> 00:04:21.752 Sure. But now, would you recommend either of these two methods? 00:04:21.992 --> 00:04:27.552 Would you say, look, top-down is more effective than bottom-up? No, no. 00:04:27.612 --> 00:04:32.452 What I would advocate in saying is that, whereas nowadays still, 00:04:32.552 --> 00:04:35.792 these are two largely disjoint fields. 00:04:35.932 --> 00:04:39.072 The computational neuroscientists do their bottom-up modeling, 00:04:39.172 --> 00:04:43.812 and the cognitive neuroscientists and the neuroimagers are doing their top-down modeling. 00:04:43.812 --> 00:04:48.212 Modeling, roughly independently, what we need to do is to join these fields, 00:04:48.332 --> 00:04:52.452 to break the barrier between them, and basically to work with models. 00:04:54.404 --> 00:04:58.624 To work with bottom-up models, which we can also invert. So the big challenge 00:04:58.624 --> 00:05:07.444 that we have ahead of us is to have basically models which can do a few things. 00:05:07.884 --> 00:05:10.524 One is that they're biophysically inspired, so that's number one. 00:05:10.924 --> 00:05:12.524 That was always true for bottom-up models. 00:05:13.704 --> 00:05:18.404 The second one is that they can actually perform the task in the sense that 00:05:18.404 --> 00:05:22.324 they have a real computational implementation, a real computational hypothesis 00:05:22.324 --> 00:05:25.824 about how a human or an animal performs a task. 00:05:26.644 --> 00:05:29.404 This was, both was often true for bottom-up models. 00:05:29.684 --> 00:05:35.964 On the other hand, you had the top-down models, which were largely able to invert 00:05:35.964 --> 00:05:39.524 observation models, like I've just explained, but they largely were aimed at 00:05:39.524 --> 00:05:40.824 explaining neuroimaging data. 00:05:40.984 --> 00:05:45.044 So for instance, when you do this top-down modeling, when you go from your fMRI 00:05:45.044 --> 00:05:52.464 data down to the hypothesized neuronal activity that you can see to decide which 00:05:52.464 --> 00:05:54.424 of several competing hypotheses, 00:05:54.804 --> 00:05:56.724 which of several competing models was the best. 00:05:56.904 --> 00:06:00.724 The ultimate measure for top-down modelers is often how well does it fit my data? 00:06:00.984 --> 00:06:05.924 And of course, fit is not everything. I mean, fit with a lot more parameters is not good. 00:06:06.004 --> 00:06:09.684 So often you would also add a complexity penalty there and look at something like model evidence. 00:06:09.884 --> 00:06:12.644 But still, in the end, what you're looking at is the fit to the neuroimaging data. 00:06:12.964 --> 00:06:16.464 So what I would like to see us do, and I think many people with me, 00:06:17.064 --> 00:06:18.924 is to join this and to work with models. 00:06:20.069 --> 00:06:27.789 Not only fit the data as top-down modelers do, but also perform the task as bottom-up modelers do. 00:06:27.889 --> 00:06:30.789 So I would like to work with bottom-up computational models of tasks, 00:06:30.989 --> 00:06:34.809 which we can also invert and therefore make accountable to neuroimaging data. 00:06:35.149 --> 00:06:39.369 And that's an ultimate combination of three things, which has not been achieved 00:06:39.369 --> 00:06:40.909 yet. That's a big challenge. 00:06:41.249 --> 00:06:45.449 But the way you describe this, doesn't this sound a bit like that you say, 00:06:45.489 --> 00:06:50.009 look, we have all these tools, and now we have these amazing methods to try 00:06:50.009 --> 00:06:52.869 to interpret the data in new ways. 00:06:53.449 --> 00:06:58.049 But now we need to find an interpretation. So now I have to find the bottom-up 00:06:58.049 --> 00:07:03.429 modelers who look more to biophysics, anatomy, and so on, to help me interpret 00:07:03.429 --> 00:07:06.189 this abstract description that I'm trying to disentangle. 00:07:06.929 --> 00:07:10.529 So it's clear. It will be clear what your motivation is to find that link because 00:07:10.529 --> 00:07:13.949 you need that translation step to work your way towards the neural substrate. 00:07:13.949 --> 00:07:18.749 But the modeler themselves might say, well, what's the leverage I get? 00:07:19.189 --> 00:07:25.269 Adding this layer of complexity, inventing some sort of approximation of an 00:07:25.269 --> 00:07:29.869 fMRI signal, so I have to model some hemodynamical system that is, 00:07:29.869 --> 00:07:33.509 again, linked to my neural simulation just to make Allard happy. 00:07:33.769 --> 00:07:38.049 So why would I do that? He's not doing it to make me happy. 00:07:38.129 --> 00:07:42.769 He's doing it to make himself happy because the basic problem that any bottom-up 00:07:42.769 --> 00:07:46.429 modeler has is that his approach is completely undetermined. 00:07:46.429 --> 00:07:53.329 So it's a well-known fact that given some type of high-level behavioral data, 00:07:53.509 --> 00:07:58.089 let's say change blindness, or even some high level of behavior, 00:07:58.469 --> 00:08:04.229 let's say exploration behavior, you can find an infinite set of models which 00:08:04.229 --> 00:08:05.609 could all generate that behavior. 00:08:05.609 --> 00:08:10.189 Behavior, how to then decide which of those infinite sets of models is the true model. 00:08:10.329 --> 00:08:13.809 And supposedly that would be our motivation here, to find how a human or how 00:08:13.809 --> 00:08:18.349 a rat or how a monkey performs a certain task and to understand therefore how 00:08:18.349 --> 00:08:20.909 the brain of the human or the animal works. 00:08:21.849 --> 00:08:26.529 So in other words, if the bottom-up modeler works with very high-level behavioral 00:08:26.529 --> 00:08:28.769 data, he will always be under constraint. 00:08:28.929 --> 00:08:31.889 He can come up with any type of model. I can always say, see, my model fits behavior. 00:08:32.089 --> 00:08:34.889 Yeah, sure, there's an infinite amount of other models that would also fit behavior, 00:08:34.989 --> 00:08:36.929 but mine happens to fit behavior. What else do you want from me? 00:08:37.249 --> 00:08:42.729 Well, that's not good enough, obviously. We need more data to constrain these 00:08:42.729 --> 00:08:47.829 models in order to decide not only does it fit the behavioral data or the task itself, 00:08:47.969 --> 00:08:52.209 can it perform the task, but also does it perform it in the way that the human 00:08:52.209 --> 00:08:54.069 brain or the mammal brain performs it. 00:08:54.149 --> 00:08:57.749 And there you need the extra data at the level of the actual neuronal activity. 00:08:57.869 --> 00:09:00.209 And that's the data that neuroimaging can provide us with. 00:09:01.030 --> 00:09:05.510 That's what the bottom-up modeler gets from it. Well, the promise is great. 00:09:05.890 --> 00:09:10.070 So I agree with that. And indeed, this problem of determinancy or the indeterminacy 00:09:10.070 --> 00:09:11.530 of many models is very fundamental. 00:09:11.770 --> 00:09:15.810 And indeed, we have to find ways to add constraints to models to make them meaningful. 00:09:15.930 --> 00:09:18.330 And the neuroimaging data is that constraint, the extra constraint. 00:09:18.530 --> 00:09:21.550 Well, this is the issue we should inspect, right? Because you are then, 00:09:21.590 --> 00:09:26.110 in some sense, promising that the neuroimaging data is of some special quality 00:09:26.110 --> 00:09:30.350 because it's not just yet another data source. It seems the way you describe 00:09:30.350 --> 00:09:32.650 it that you say it's a decisive data source. 00:09:32.810 --> 00:09:36.070 It's a data source of a very special quality. So what's that special quality? 00:09:36.270 --> 00:09:41.990 Why is it not yet another source of information like possibly gene expression 00:09:41.990 --> 00:09:45.870 or some second messenger system or behavior or what have you? 00:09:46.010 --> 00:09:49.470 What makes it special? Good question. It's more gradual. 00:09:49.630 --> 00:09:54.750 I understand your question, but it's more gradual. I am not promising that it's the decisive data set. 00:09:54.750 --> 00:09:59.670 What I am convinced of and I would like to convince the listeners of or my fellow 00:09:59.670 --> 00:10:05.750 modelers of is that it's a set of data which contains a lot of very important 00:10:05.750 --> 00:10:08.210 information that can constrain our model better. 00:10:08.530 --> 00:10:12.510 For the mere fact that what neuroimaging does, and that has always been its 00:10:12.510 --> 00:10:15.890 promise and that is actually true, it does allow us to look inside the brain, 00:10:16.030 --> 00:10:20.590 which we weren't previously able to do with in vivo functioning humans without 00:10:20.590 --> 00:10:22.370 actually intervening into the system. 00:10:22.370 --> 00:10:25.030 That has been the promise of neuroimaging and 00:10:25.030 --> 00:10:27.810 fMRI in particular for 15 years now and that is really 00:10:27.810 --> 00:10:30.590 true yes we can look into the brain at some level 00:10:30.590 --> 00:10:34.330 of this section of course it's it's hemodynamics it's metabolics but 00:10:34.330 --> 00:10:38.150 still it's the metabolics which is much closer to the brain activity than the 00:10:38.150 --> 00:10:41.090 eye movements or the behavior which we would otherwise be observing and we're 00:10:41.090 --> 00:10:45.530 still also observing that so it gives us a very important extra constraint which 00:10:45.530 --> 00:10:49.870 is much closer to the brain to the actual neural activity than what we had before 00:10:49.870 --> 00:10:52.370 and that's the privileged status of that data. 00:10:52.450 --> 00:10:54.690 It's not perfect, but it's a lot better than what we had. 00:10:55.665 --> 00:11:00.125 That isn't true under certain circumstances. No, if I, for instance, 00:11:00.245 --> 00:11:03.885 if you would restrict your argument to, let's say, the human brain, 00:11:04.045 --> 00:11:07.045 it's fairly clear. No, because we don't have many alternative methods. 00:11:07.425 --> 00:11:10.985 But however, if we go to macaque or rodent or cat and so on, 00:11:11.065 --> 00:11:15.945 I could argue, look, but then I have alternative physiological methods I can deploy. Such as? 00:11:16.385 --> 00:11:20.985 Let's name them. I can do my electrophysiology. Exactly. I can do much more 00:11:20.985 --> 00:11:22.525 detailed anatomical studies. 00:11:22.625 --> 00:11:26.865 I can do much more controlled behavioral studies. However, the paradigms I can 00:11:26.865 --> 00:11:28.645 deploy will, of course, be more restricted. 00:11:29.205 --> 00:11:32.545 There's another very important restriction there, Paul, and that is that none 00:11:32.545 --> 00:11:38.425 of those methods that you mentioned now have whole brain coverage in an in vivo 00:11:38.425 --> 00:11:41.585 animal without doing any damage and therefore have repeatability. 00:11:41.985 --> 00:11:46.865 Electrophysiology is brilliant. I'm not against it. But even having a very dense 00:11:46.865 --> 00:11:51.705 multi-electrode array, which you can do nowadays, you only have a very sparse 00:11:51.705 --> 00:11:55.305 sampling of a couple of populations of neurons. 00:11:55.425 --> 00:11:58.625 And those neurons you can observe very well with nice spike sorting algorithms. 00:11:58.745 --> 00:12:02.445 You can get at every electrode a couple of tens or maybe even hundreds of neurons, 00:12:02.625 --> 00:12:04.225 but only tens of hundreds of neurons. 00:12:05.005 --> 00:12:09.285 And then you have maybe 100 or 200 of those spots. But is that enough? 00:12:09.445 --> 00:12:12.445 I mean, we have millions, billions of neurons in the brain. 00:12:12.765 --> 00:12:18.605 So what electrophysiology is, what it's great at, is that it's one or even ten 00:12:18.605 --> 00:12:23.165 or a hundred huge magnifying glasses onto the activity of a few neurons that 00:12:23.165 --> 00:12:27.145 you can observe and you're ignorant about what happens in the rest of the brain. 00:12:27.305 --> 00:12:29.945 And the beautiful complementary aspect of something like fMRI, 00:12:30.145 --> 00:12:33.845 which I would also perform on a monkey, is that it's the whole brain and all of the brain. 00:12:34.445 --> 00:12:39.105 And of course then what we have is the whole brain and all of the brain at a 00:12:39.105 --> 00:12:43.045 certain spatial resolution, which is much coarser than electrophysiology Although 00:12:43.045 --> 00:12:47.065 with increasing field size and increasing MR technology, it's getting better and better and better. 00:12:47.165 --> 00:12:51.645 We're going towards the 100 micron level now for fMRI, but it's the whole brain. 00:12:52.424 --> 00:12:55.044 And that is something unique that we didn't have before in any of the methods 00:12:55.044 --> 00:12:55.844 that you just mentioned. 00:12:56.804 --> 00:13:02.424 Granted, this is all very acceptable. However, the weakness of one method is 00:13:02.424 --> 00:13:03.644 not the strength of another method. 00:13:03.984 --> 00:13:07.384 So, for instance, in your case, you can say, well, I have a whole brain. 00:13:08.124 --> 00:13:11.324 And granted, in principle, you extract data from the whole brain. 00:13:11.364 --> 00:13:15.124 But very often in the analysis, you see people suddenly starting to jump into 00:13:15.124 --> 00:13:18.184 what they call regions of interest because they will otherwise I cannot extract 00:13:18.184 --> 00:13:22.244 any meaningful correlations from my data. So then, to what extent do I actually 00:13:22.244 --> 00:13:26.224 really exploit still being this bottom-up modeler looking for constraints? 00:13:26.564 --> 00:13:29.524 To what extent am I then actually really exploiting this advantage? 00:13:29.664 --> 00:13:34.204 Because here I go, I have neat whole-brain data at a fairly coarse spatial temporal resolution. 00:13:34.564 --> 00:13:38.764 But now, in order to make sense of this very noisy signal, I start to impose 00:13:38.764 --> 00:13:41.964 more restrictive analysis windows, so I throw away the whole-brain information. 00:13:42.344 --> 00:13:45.384 So, what am I left with? What's my advantage? Very good question, Paul. 00:13:45.384 --> 00:13:48.924 I actually recognized some of my slides in my talk this morning where I actually 00:13:48.924 --> 00:13:52.824 implicitly complained about this approach where, although we have whole brain data, 00:13:53.204 --> 00:13:58.464 what top-down modelers tend to do with their neuroimaging data nowadays is to 00:13:58.464 --> 00:14:04.284 select a few selected regions within it and only model the interactions between those regions. 00:14:04.284 --> 00:14:08.084 I am strongly advocating that, yes, we should use the whole brain data. 00:14:08.884 --> 00:14:12.164 Not only should our models contain a computational model of the task, 00:14:12.444 --> 00:14:16.584 but also they should be accountable to the entire data set. 00:14:16.684 --> 00:14:19.024 They should preferably be models of the entire brain. 00:14:19.164 --> 00:14:25.784 So, yes, I fully agree there that sub-selecting data from this very rich whole 00:14:25.784 --> 00:14:27.864 brain data set that we have is not the right approach because it's actually 00:14:27.864 --> 00:14:29.944 not using the strength of the method. it. 00:14:30.724 --> 00:14:34.084 So top-down models of neuroimaging data, 00:14:34.284 --> 00:14:37.384 and this is something I've been advocating in the last couple of years and I'm 00:14:37.384 --> 00:14:40.864 currently working on, should be models of the whole brain, at least of the whole 00:14:40.864 --> 00:14:44.564 cortex, preferably of the whole brain, and take into account at, 00:14:44.664 --> 00:14:46.484 of course, the level where we are with. 00:14:47.392 --> 00:14:53.092 Particular modality, it should take into account all of the neuronal activity that it's sampling. 00:14:53.352 --> 00:14:59.592 Okay, but then if I would follow your advice, then I would impose this Allard-Rabruck 00:14:59.592 --> 00:15:00.612 filter of the literature. 00:15:00.832 --> 00:15:04.052 There are not very many papers I'm left with to constrain my model. Is that correct? 00:15:04.452 --> 00:15:07.552 No, but I'm giving you a vision onto a future. 00:15:08.432 --> 00:15:11.632 I'm giving you what I feel that we should be working on. 00:15:11.732 --> 00:15:15.172 And I think quite a few people agree with me that we are working on this and 00:15:15.172 --> 00:15:19.352 trying to get there. Okay, so that means that we actually agree that today, 00:15:19.532 --> 00:15:22.552 as the bottom-up modeler, I'm sort of a bit left alone. 00:15:22.732 --> 00:15:26.632 I will not be able to find these constraints that I actually would need. 00:15:27.352 --> 00:15:32.172 No, they are there, and I would invite any bottom-up modeler to join this research 00:15:32.172 --> 00:15:36.932 program and help us in achieving this and finding ways in which your bottom-up 00:15:36.932 --> 00:15:39.452 models can be made accountable to neuroimaging data. 00:15:40.312 --> 00:15:43.372 Yeah, but that would help you. It would still not necessarily help the bottom-up 00:15:43.372 --> 00:15:45.432 model, right? Yes, it would help you because of the reason I said earlier. 00:15:45.672 --> 00:15:48.732 Because without it, you would have less constraint on your bottom-up model. 00:15:48.912 --> 00:15:52.032 Sure. I got that. Yeah. I got that point, but I was making a different point. 00:15:52.112 --> 00:15:58.812 Because your whole brain requirement is excluding, as far as I understand this 00:15:58.812 --> 00:16:00.432 literature, a large chunk of literature. 00:16:00.712 --> 00:16:04.212 So now as bottom-up modeler of phenomenon X and structure Y, 00:16:04.572 --> 00:16:09.452 there are not that many studies I could look at today to help me constrain my model. 00:16:09.732 --> 00:16:12.452 Because it's not a model of the whole brain. Exactly, yeah. Right, 00:16:12.472 --> 00:16:14.432 yeah. You would agree with that contention? 00:16:15.472 --> 00:16:19.352 I would agree. I mean, yes, where, of course, the issue gets more subtle, 00:16:19.392 --> 00:16:23.172 and this is an issue you always have in modeling, and that will remain there 00:16:23.172 --> 00:16:28.032 also in this new program of doing things, of combining things, 00:16:28.132 --> 00:16:31.052 is that a model always needs to be a model at a certain level of abstraction. 00:16:32.974 --> 00:16:37.014 Often they say all models are wrong, but some are useful. It's a famous, famous thing to say. 00:16:37.954 --> 00:16:41.574 And some of the ones that are useful are the ones that take a suitable abstraction 00:16:41.574 --> 00:16:47.014 level and a suitable set of the data to model and to try and understand that 00:16:47.014 --> 00:16:48.474 part. And that will always remain true. 00:16:48.694 --> 00:16:53.854 A useful model is one that includes all the relevant areas for your task and 00:16:53.854 --> 00:16:56.214 has a nice computational model for those areas. 00:16:56.594 --> 00:17:00.714 And of course, although I would advocate that you would not forget Forget any 00:17:00.714 --> 00:17:05.074 of the important areas for your task that you would not constrain your model a priori to be, 00:17:05.254 --> 00:17:10.574 to contain two or three regions in the brain where you know or expect or are 00:17:10.574 --> 00:17:13.834 afraid that there might be two or three or four or five other regions in the 00:17:13.834 --> 00:17:18.014 brain that it contains only those two or three that you a priori started with. 00:17:18.454 --> 00:17:22.354 Once you do have all the regions which you reasonably believe to be involved 00:17:22.354 --> 00:17:25.874 in the task, the six or seven, let's say, then of course, sure, 00:17:26.014 --> 00:17:27.894 focus on the six or seven that are relevant. 00:17:27.894 --> 00:17:30.714 If the rest of the brain really isn't relevant, I won't force you to model it 00:17:30.714 --> 00:17:34.674 because that's not a useful abstraction. That's not a useful model then. Right. 00:17:35.234 --> 00:17:39.494 Okay, so here we go. So we're looking to this future where we have this whole 00:17:39.494 --> 00:17:47.354 brain fMRI data interpreted in some, let's say, connectivity scheme, I guess. 00:17:48.774 --> 00:17:53.074 But now from the modeling perspective, what I'm actually trying to replicate 00:17:53.074 --> 00:17:58.174 is the causal structure of the phenomenon. I really try to think about how neurons 00:17:58.174 --> 00:18:02.034 interact with each other and A causing B and so on, leading to some macroscopic 00:18:02.034 --> 00:18:03.554 phenomenon I could call behavior. 00:18:04.214 --> 00:18:09.654 But in some sense, if I look at the methods you're developing with your colleagues, 00:18:09.794 --> 00:18:14.534 we very quickly end up not talking about connectivity, but about a special kind 00:18:14.534 --> 00:18:16.554 of connectivity, which is functional connectivity, 00:18:16.834 --> 00:18:20.454 which in some sense is related to a correlation structure. 00:18:21.295 --> 00:18:26.855 And as we all know, going back to Hume and before, correlation is only a necessary 00:18:26.855 --> 00:18:27.995 condition of causation. 00:18:28.135 --> 00:18:32.735 And as a constraint, I would like to have access to causal information. 00:18:33.455 --> 00:18:36.475 So how are we going to solve this conundrum? So that means on the one hand, 00:18:36.535 --> 00:18:38.635 the whole brain data I still don't have. 00:18:38.735 --> 00:18:42.795 But now if we have it, imagine we have it, then I want to know I want to have causal information. 00:18:43.395 --> 00:18:47.115 But actually, if I understand what you were telling us, today, 00:18:47.235 --> 00:18:49.415 this is really difficult to get your hands on. 00:18:49.815 --> 00:18:52.835 So what's our problem there? But first, we're missing something. 00:18:53.035 --> 00:18:55.215 We don't only want causal information. 00:18:55.495 --> 00:18:58.795 We also want computational information. Let me explain what I mean here. 00:18:58.975 --> 00:19:06.055 We know that activity in V2 is caused by activity in V1. And that's an important fact to know. 00:19:06.175 --> 00:19:11.115 But that doesn't actually explain to us or give us any understanding about what 00:19:11.115 --> 00:19:13.855 V1 and V2 do exactly. Do we really know that? 00:19:14.435 --> 00:19:17.455 Let's assume for now that we know that. Okay. 00:19:17.995 --> 00:19:22.395 Right? I tried to give an example. Sure. Yes. I was not facetious. I was serious. 00:19:22.675 --> 00:19:26.975 Yeah. So actually, we actually believe that there might also be recurrent information. For instance. 00:19:27.235 --> 00:19:31.135 But that already goes towards a more computational model of what they're doing. 00:19:31.255 --> 00:19:36.035 So a causal model, which I agree many of the top-down models of neuroimaging 00:19:36.035 --> 00:19:39.875 data have focused on, are perfectly content in saying that, okay, 00:19:39.895 --> 00:19:44.535 if I can get with my top-down modeling, with my inversion of complex observation 00:19:44.535 --> 00:19:46.115 models to a model which says that 00:19:46.215 --> 00:19:52.275 information is going from V1 to V2, and also, by the way, from V2 back to V1, 00:19:52.335 --> 00:19:55.335 and then on to V3, V4, and so on, 00:19:55.435 --> 00:20:01.775 then they're happy to stop there and say, okay, I have just distinguished by 00:20:01.775 --> 00:20:07.095 fitting to my data the model where there's a recurrent information flow, V1 and V2, 00:20:07.175 --> 00:20:10.555 and just a feed-forward V1 to V2, but nothing else. 00:20:11.715 --> 00:20:14.715 That would be a causal inference that you just spoke about. On top of that, 00:20:14.795 --> 00:20:16.015 we also need a computational inference. 00:20:16.175 --> 00:20:19.455 We actually want to know what is it that V1 does with its input. 00:20:19.595 --> 00:20:21.315 What is it computing, right? 00:20:21.875 --> 00:20:27.155 And then what is it computing that it sends on to V2, and what would V2 supposedly 00:20:27.155 --> 00:20:31.915 be computing, which it either exclusively sent on to V3, V4, 00:20:32.095 --> 00:20:35.335 or which it would actually recurrently feed back to V1. 00:20:35.975 --> 00:20:40.255 So that's on top of the causal inference of saying that some type of information 00:20:40.255 --> 00:20:44.755 and we don't know what moves from V1 to V2 is actually having a computational 00:20:44.755 --> 00:20:50.115 model of direction-specific cells, of, you know, columns, orientation columns, 00:20:50.475 --> 00:20:53.595 color blobs, things like that, gives us a much finer understanding. 00:20:54.295 --> 00:20:57.695 So that's what I mean when I say that computational bottom-up models have to 00:20:57.695 --> 00:21:01.075 be integrated and made accountable to neuroimaging data. 00:21:02.012 --> 00:21:07.132 But now I'm confused in some sense because I feel that – So I think we had a 00:21:07.132 --> 00:21:09.672 good agreement on imposing constraints on these models. 00:21:09.852 --> 00:21:14.952 And that whatever set of constraints we can find, we should use them. 00:21:15.152 --> 00:21:19.012 In that sense, in the imaging data. So you now agree that you'll use newer imaging 00:21:19.012 --> 00:21:20.552 data not for Allard but actually for yourself? 00:21:20.932 --> 00:21:24.732 No, I don't agree to the imaging data yet. All right. I agree to the general 00:21:24.732 --> 00:21:28.252 principle. All right. Which we already adhere to for many years. 00:21:28.532 --> 00:21:34.712 All right. But, and then we, then your argument was to say, and fMRI data is 00:21:34.712 --> 00:21:38.872 one of the sources of information constraints that you can impose in your model. 00:21:39.112 --> 00:21:42.032 And then I made this argument about actually modeling causal structures, right? 00:21:42.252 --> 00:21:48.692 And I said, but the way you guys analyze this imaging data doesn't give me access to causal structure. 00:21:48.932 --> 00:21:53.132 So how then can it be a constraint that is in any way decisive for me? 00:21:53.212 --> 00:21:58.432 Because what I want to understand is the causal, the causal links in this network. 00:21:59.452 --> 00:22:03.732 But then in answering – so I had expected it would say something like, 00:22:03.752 --> 00:22:05.632 oh, but actually we do have access to causality. 00:22:05.692 --> 00:22:10.252 We have methods to try to infer causal structure so we can strengthen it. 00:22:10.312 --> 00:22:13.472 But then actually you made a jump and you went to something that's called computation, 00:22:13.852 --> 00:22:17.212 which is not necessarily well-defined. True. 00:22:18.032 --> 00:22:24.072 And then suddenly that was as if the imaging data would give me a sense of some 00:22:24.072 --> 00:22:28.892 functional property that I could then use as my constraint. So you seem to sidestep 00:22:28.892 --> 00:22:30.032 that issue of causality. 00:22:30.172 --> 00:22:33.552 All right. And now you seem to tell me that actually it's not about causality, 00:22:33.612 --> 00:22:37.532 but I can tell you something about the functional operations that happen in that structure. 00:22:37.852 --> 00:22:41.192 So now I'm confused because what are you offering to me, really? 00:22:41.192 --> 00:22:45.112 All right, right. So let's take it in a few steps. And it is about causality. It is about causality. 00:22:46.072 --> 00:22:49.412 And maybe I jumped over that too fast because I'm eager to move on, 00:22:49.452 --> 00:22:54.072 but causality is important. So it is important to know before we know anything 00:22:54.072 --> 00:22:58.652 that there might be causality, which is, by the way, also notoriously ill-defined 00:22:58.652 --> 00:22:59.772 just like computation is. 00:22:59.852 --> 00:23:03.632 But let's suppose we have some relevant intuitive understanding of what it means. 00:23:03.672 --> 00:23:07.672 It is relevant to know that there is causality from V1 to V2. 00:23:08.132 --> 00:23:11.092 Supposedly, we didn't know anything about the brain. Knowing that teaches us 00:23:11.092 --> 00:23:12.432 something, and it teaches us something useful. 00:23:13.132 --> 00:23:16.812 I then quickly moved on to say, but we want to know more. But let's actually, 00:23:17.172 --> 00:23:20.072 right, and that was my argument just now, but I understand your question like, 00:23:20.152 --> 00:23:22.172 oh, you're moving too fast. Let's look at causality first. 00:23:22.572 --> 00:23:28.832 So, yes, okay, since Hume and since, you know, what many first-year psychology 00:23:28.832 --> 00:23:33.112 students know is that from correlations alone, which is in principle what we 00:23:33.112 --> 00:23:36.232 have and what we can observe and compute from our neuroimaging data, 00:23:36.312 --> 00:23:37.292 we don't have causality. 00:23:37.772 --> 00:23:41.112 Causality is basically correlations plus additional assumptions. 00:23:42.592 --> 00:23:45.652 And the question is, of course, what are those assumptions? And if those assumptions 00:23:45.652 --> 00:23:50.512 are good assumptions, we might actually get to causality, to some form of causality. 00:23:51.021 --> 00:23:56.121 Causality derived from correlations to some observational causality rather than 00:23:56.121 --> 00:23:59.281 a perturbational causality which would be actually intervening to the system. 00:23:59.381 --> 00:24:03.361 I'm just leaving that aside for now but a useful form of causality but those 00:24:03.361 --> 00:24:04.521 assumptions are really important. 00:24:04.801 --> 00:24:09.621 Which assumptions do we put into our models which allow us to go from correlations 00:24:09.621 --> 00:24:13.621 and also higher forms of dependence to causality? 00:24:14.001 --> 00:24:18.161 Well that's where we actually also get back to computation in the end because 00:24:18.161 --> 00:24:21.621 computational models would be one of the forms of very fine, 00:24:21.781 --> 00:24:28.721 very detailed assumptions that would have us go from correlations to causations. 00:24:28.841 --> 00:24:30.961 But already before that, once again, that's the next step. 00:24:31.121 --> 00:24:33.961 I don't want to overstep the point of causality that you were making. 00:24:34.781 --> 00:24:39.081 Yes, already when we're doing pure top-down modeling as it's applied today, 00:24:39.221 --> 00:24:43.821 going from our neuroimaging data to causality, to influence at the neuronal 00:24:43.821 --> 00:24:47.081 level, it's all the assumptions which are in the model the assumptions which 00:24:47.081 --> 00:24:48.961 are in what we call the neurodynamic model. 00:24:49.081 --> 00:24:53.481 How do we actually model interacting neuronal populations? What does that look like? 00:24:54.061 --> 00:24:58.441 And in the observation model that takes our neuronal population activity to 00:24:58.441 --> 00:25:02.441 our observations which would be a model of hemodynamics for fMRI and the models 00:25:02.441 --> 00:25:07.721 of leak fields and gain matrices in EEG or MEG that contains all of those assumptions 00:25:07.721 --> 00:25:10.981 which combined with our data can tell us something about gazellon. 00:25:11.521 --> 00:25:15.161 Yes, so one reason why I like to stick a little bit to this causality issue 00:25:15.161 --> 00:25:20.321 is also if you jump to what we could call computation, it's also an excuse in 00:25:20.321 --> 00:25:22.981 some sense to start to discard constraints. 00:25:23.261 --> 00:25:25.821 Because if you say, oh, let's just worry about the computation, 00:25:26.021 --> 00:25:29.681 in some sense anything that happens in space and time you're going to discard. 00:25:29.841 --> 00:25:32.881 Because you say, no, now I move from a physical space to a functional space. 00:25:34.161 --> 00:25:38.521 And before we make that jump, I think it would be useful to understand And what 00:25:38.521 --> 00:25:43.421 possibilities do you have right now in your toolbox to make statements about 00:25:43.421 --> 00:25:48.641 causal interactions in this kind of imaging data that you have access to? How is this really done? 00:25:49.061 --> 00:25:51.161 All right. Okay. 00:25:52.327 --> 00:25:56.867 So, there's a couple of ways in which you can operationally define causality. 00:25:57.007 --> 00:25:59.767 And that's actually where we now have to press for a definition of causality, 00:25:59.907 --> 00:26:02.107 which, as I said, is notoriously ill-defined. 00:26:02.587 --> 00:26:05.807 So, there's a couple of possible definitions. I just mentioned the perturbational approach. 00:26:06.287 --> 00:26:11.347 Some philosophers would insist, actually, that the only way to get to causality 00:26:11.347 --> 00:26:15.587 is to perform experiments where you actively disturb the system and see what 00:26:15.587 --> 00:26:17.447 the consequences are. I can agree with that. 00:26:17.607 --> 00:26:20.007 It's the foundation of empirical science. Exactly, yes. Yes, 00:26:20.007 --> 00:26:26.127 but also immediately with an interactive functioning system and a complex system 00:26:26.127 --> 00:26:27.827 like the brain, it also presents you with problems, 00:26:28.027 --> 00:26:34.707 which you could actually, at a quite abstract, vague level, compare to the uncertainty principle. 00:26:35.047 --> 00:26:39.587 By observing the system and by perturbing the system, you are disturbing the 00:26:39.587 --> 00:26:44.687 system. So by perturbing the system, you're actually looking at an unnaturally functioning system. 00:26:44.747 --> 00:26:47.607 By using, for instance, something like transcranial magnetic stimulation, 00:26:47.907 --> 00:26:52.127 where you give quite a blast of magnetic pulse into the brain, 00:26:52.227 --> 00:26:56.807 which then induces electric currents in the brain tissue, 00:26:56.987 --> 00:26:59.687 you're doing something completely unnatural to the brain because that are not 00:26:59.687 --> 00:27:05.327 the usual microvolt currents, which are functionally normal to run in the brain. 00:27:05.327 --> 00:27:09.527 Basically, what you're doing is, a nice analogy is, if an alien would come from 00:27:09.527 --> 00:27:13.067 outer space and we're trying to understand what our cars are like, 00:27:13.147 --> 00:27:16.747 supposedly, if they came from outer space, they have much more advanced technology than that. 00:27:16.947 --> 00:27:20.527 But still, let's suppose that they see one of our cars and want to make sense of it. 00:27:20.767 --> 00:27:24.067 The correlational approach that I was advocating before would be to have the 00:27:24.067 --> 00:27:27.447 engine run and measure temperatures here and there and actually seeing that 00:27:27.447 --> 00:27:30.167 when the temperature goes up in one place, it also tends to go up in the other place. 00:27:30.227 --> 00:27:33.667 So actually, these things might have something to do with each other. but we 00:27:33.667 --> 00:27:36.647 need more assumptions to understand what it is that they do to each other which 00:27:36.647 --> 00:27:41.287 is what we were discussing before the perturbational approach here would be 00:27:41.287 --> 00:27:45.907 and to drive my point home taking a big sledgehammer giving a big knock onto 00:27:45.907 --> 00:27:50.807 one part of the engine seeing what shakes basically and deciding that what shakes 00:27:50.807 --> 00:27:53.587 must be functionally connected to where i knocked the engine. 00:27:54.327 --> 00:27:57.707 Now this point drives the point home that this might not actually be how an 00:27:57.707 --> 00:28:05.347 engine works and that if I knock an engine, I perturb the engine in such a way 00:28:05.347 --> 00:28:07.867 that the shaking that goes on then, 00:28:07.987 --> 00:28:10.447 the perturbations, the consequences of it, yes, 00:28:10.587 --> 00:28:13.047 they are most definitely and utterly causal. 00:28:13.367 --> 00:28:18.607 But are they actually the important causal things that make the engine work? I don't think so. 00:28:20.234 --> 00:28:24.934 I understand the distinction. I could argue, though, that in some ways you're 00:28:24.934 --> 00:28:26.034 setting up a straw man, no? 00:28:26.114 --> 00:28:28.334 Because I could argue, well, to understand how this engine works, 00:28:28.514 --> 00:28:31.374 I could do a more subtle perturbation. I could just, let's say, 00:28:31.414 --> 00:28:34.274 take out one of the spark plugs and see what happens to the power output. 00:28:34.594 --> 00:28:37.514 And once again, it wouldn't be the normal function of the engine to start with, 00:28:37.594 --> 00:28:39.514 in a subtle way, but it's still the same argument. 00:28:41.014 --> 00:28:45.414 But I could have more subtle perturbations on the system. Even more subtle. 00:28:45.594 --> 00:28:48.914 And then how subtle would they have to be? I change the air mixture and the 00:28:48.914 --> 00:28:52.234 carburent I'm injecting and so on. 00:28:52.354 --> 00:28:59.414 But I'm not sure if this is completely fair because effectively on your human 00:28:59.414 --> 00:29:03.974 subject that you try to decompose functionally, you also perform small perturbations. 00:29:03.974 --> 00:29:04.934 That's your experimental protocol. 00:29:05.234 --> 00:29:08.734 Exactly. Right? And importantly, these are natural perturbations. 00:29:08.734 --> 00:29:13.114 These are the perturbations that I would advocate using are to provide you as 00:29:13.114 --> 00:29:15.634 a subject with a perturbation into your visual field, 00:29:15.774 --> 00:29:19.474 for instance, flashing a face picture at you, which you would then have to perceive 00:29:19.474 --> 00:29:22.974 as being a face and to decide whether this is the face of your neighbor or not. 00:29:23.114 --> 00:29:26.894 Of course, that then causes things to happen in your brain, but natural things 00:29:26.894 --> 00:29:29.814 as you would naturally function. And that's the important distinction. 00:29:30.494 --> 00:29:34.334 Indeed. But as you know, also the first year psychology student will be told 00:29:34.334 --> 00:29:36.314 about the problem of ecological validity. 00:29:36.774 --> 00:29:41.534 And your subject in the scanner looking at all sorts of pictures is also not 00:29:41.534 --> 00:29:44.754 necessarily exposed to a natural situation. So it seems that you're saying, 00:29:44.874 --> 00:29:49.034 well, some perturbations are more natural than others. 00:29:49.294 --> 00:29:51.974 And that's what I'm saying. I'm saying it's a continuity. It's a continuity. 00:29:52.974 --> 00:29:59.614 So I would say that an active perturbation in the form of TMS is highly non-natural. 00:30:00.312 --> 00:30:04.572 I would also say that flashing a face picture for 100 milliseconds to one of 00:30:04.572 --> 00:30:09.592 my subjects is more natural because it's the natural channel of input into my 00:30:09.592 --> 00:30:11.692 brain. That's at least better than what I did with TMS. 00:30:11.872 --> 00:30:15.412 By the way, I'm sounding negative about TMS. I think TMS is a great method. 00:30:15.532 --> 00:30:18.232 I think the podcast listeners should know that. 00:30:18.572 --> 00:30:23.712 I'm just trying to contrast two concepts here, and that's why I'm beating at TMS here. 00:30:23.852 --> 00:30:27.112 No, I don't understand. So at least the perturbation is coming in through the 00:30:27.112 --> 00:30:30.912 natural channels that my brain is used to receive information through. 00:30:31.252 --> 00:30:36.632 But of course, in the natural world, I'm not having faces flash to me. 00:30:36.672 --> 00:30:39.992 Rather, I have a dynamic stimulus array that is continually changing. 00:30:40.212 --> 00:30:44.252 So an already more natural stimulus would be to look at a movie and being able to fixate around. 00:30:44.632 --> 00:30:49.172 And an even more natural stimulus after that might be actually to actively explore my environment. 00:30:49.172 --> 00:30:54.532 And there's a whole continuous array which indeed explores the trade-off between 00:30:54.532 --> 00:31:00.832 having ecologically valid and natural stimuli on one side and having controllability upon the other. 00:31:00.992 --> 00:31:04.912 We're a controllable stimulus where we keep everything constant and have no 00:31:04.912 --> 00:31:08.732 confounds and vary one thing so that we know what we're actually doing is often 00:31:08.732 --> 00:31:10.732 highly unnatural and highly non-ecological. 00:31:10.732 --> 00:31:15.692 But very ecological experiments showing movies active exploration are highly 00:31:15.692 --> 00:31:19.132 non-controlled, so there are many things changing at the same time. 00:31:19.212 --> 00:31:21.332 So anything we observe, what is it really due to? 00:31:21.532 --> 00:31:26.712 And in designing experiments on this continuum and finding places in the middle, 00:31:26.812 --> 00:31:29.652 useful places in the middle, is actually where we might learn a lot. 00:31:30.452 --> 00:31:35.132 So, if I got it right, that means that this contrast you seem to sketch between 00:31:35.132 --> 00:31:39.132 associative methods and perturbation-based methods is less of a contrast. 00:31:39.232 --> 00:31:41.672 Because in some sense, you're saying we might want to do both, 00:31:41.712 --> 00:31:44.552 but you just want to be careful with the perturbations we consider. 00:31:44.852 --> 00:31:48.532 Absolutely. So, this is the bottom line of it. It's not necessarily one or the other. 00:31:48.692 --> 00:31:51.192 That's a beautiful conclusion, Paul, because that shows that I'm definitely 00:31:51.192 --> 00:31:54.472 not against TMS. I'm just seeing that… You just didn't know it yet. 00:31:55.592 --> 00:31:59.792 Oh, I definitely… I'm getting a new view onto it within our discussion. 00:31:59.792 --> 00:32:03.152 But no, that's exactly what I would like to stress, is that these are very complementary 00:32:03.152 --> 00:32:08.032 methods, which are not only black and white, but they are a whole gradation 00:32:08.032 --> 00:32:12.312 from white to gray to black that we should explore and use in a complementary 00:32:12.312 --> 00:32:14.752 way, and also especially in the in-between levels. Right. 00:32:15.092 --> 00:32:19.332 Very good. So there's one thing that worries me a bit. 00:32:19.692 --> 00:32:26.072 So it's clear that you master these methods and techniques, right? 00:32:26.172 --> 00:32:30.572 So you know what you're doing. But in some sense, a lot of these tools are also 00:32:30.572 --> 00:32:34.012 used by people who have not spent all this time training on the details of them. 00:32:34.712 --> 00:32:39.632 And, you know, you provide it with a data set and some numbers will come out. 00:32:39.732 --> 00:32:46.152 How can we make sure that the tool is not overpowering the users, that we can really... 00:32:46.832 --> 00:32:50.892 Of course, it's important that the large community of users can make use of 00:32:50.892 --> 00:32:55.012 these tools, but you don't want them all to call you all the time to ask you 00:32:55.012 --> 00:32:56.472 for advice what they really should be doing. 00:32:56.592 --> 00:33:00.472 So how can we make sure that while the complexity of the tools is increasing, 00:33:00.752 --> 00:33:03.492 we can still, as a community, make effective use of them? 00:33:04.012 --> 00:33:10.372 That's a complex question because essentially, yes, so with increasing power 00:33:10.372 --> 00:33:13.972 of the tools, often they need to become increasingly complex and they are increasingly 00:33:13.972 --> 00:33:15.332 more difficult to understand. 00:33:15.412 --> 00:33:19.972 They become often increasingly more statistically and mathematically technical. 00:33:20.352 --> 00:33:25.912 And a cognitive psychologist or a cognitive neuroscientist, of course, 00:33:25.912 --> 00:33:29.112 has a legitimate claim to want to use these methods because they were actually designed for her. 00:33:29.966 --> 00:33:36.646 But then we're, of course, presented with the conundrum that you just sketched, 00:33:36.646 --> 00:33:42.646 that a researcher has to understand things about mathematics and things about 00:33:42.646 --> 00:33:47.486 statistics and things about dynamic systems and system identification that she wasn't trained to do. 00:33:47.486 --> 00:33:51.766 So yes, the conundrum here is that obviously the responsibility always lies with the researcher. 00:33:51.926 --> 00:33:55.486 You as a researcher are always responsible for the experiments that you do. 00:33:55.506 --> 00:34:00.586 And you as a researcher are always responsible for understanding the methods 00:34:00.586 --> 00:34:03.586 that you use to such a degree that you wouldn't do anything foolish. 00:34:04.046 --> 00:34:06.146 And to such a degree is where it gets subtle. 00:34:06.846 --> 00:34:11.686 It's your responsibility that if you feel kind of iffy about some part of your 00:34:11.686 --> 00:34:16.346 analysis where you weren't actually sure that this actually is the right thing to do. 00:34:16.346 --> 00:34:19.006 But it gave you beautiful data that you would actually like to 00:34:19.006 --> 00:34:21.746 submit to nurture neuroscience that you would call me then 00:34:21.746 --> 00:34:24.866 at that moment and ask me or anyone else that's that's 00:34:24.866 --> 00:34:27.646 much smarter still and ask them like i use your method can 00:34:27.646 --> 00:34:30.586 i really do this because i'm getting beautiful data but is this actually allowed 00:34:30.586 --> 00:34:33.586 should i do it this way that's your own responsibility that's what you should 00:34:33.586 --> 00:34:41.226 do um and then of course even though uh you're responsible for understanding 00:34:41.226 --> 00:34:44.786 the methods that you use it's also perfectly natural that you can only understand 00:34:44.786 --> 00:34:46.886 them to such a level With the increasing complexity, 00:34:47.146 --> 00:34:51.786 especially in neuroimaging, where physics is involved for the MRI machine, 00:34:51.906 --> 00:34:55.886 which is a highly physically theoretic machine, has to do with static fields, 00:34:55.986 --> 00:35:00.126 gradient fields, RF fields, where then statistics is involved, 00:35:00.446 --> 00:35:01.246 mathematics is involved, 00:35:01.586 --> 00:35:03.586 neuroanatomy is involved, psychology is involved. 00:35:03.926 --> 00:35:09.186 It's a very multidimensional field. 00:35:09.646 --> 00:35:11.446 There's many expertises involved. 00:35:11.446 --> 00:35:13.866 And naturally, you would only be an expert in one or two of those. 00:35:14.126 --> 00:35:17.726 And you would only want to know the necessary thing about the other fields. 00:35:18.066 --> 00:35:20.026 Yes, that's definitely true. And that's an increasing challenge, 00:35:20.066 --> 00:35:24.826 actually, for neuroimaging, neuroscience, and also neuromodelers to try and 00:35:24.826 --> 00:35:27.486 figure that out, to form consortiums, in fact, of researchers, 00:35:27.606 --> 00:35:30.086 which can maybe help each other doing that in the right way. 00:35:30.406 --> 00:35:34.946 Correct. So, but then would you claim that 100% of the users of these tools 00:35:34.946 --> 00:35:37.366 satisfy these constraints you just spelled out? 00:35:38.446 --> 00:35:42.006 I wouldn't know. I wouldn't claim to have insight into who is using these tools. 00:35:42.106 --> 00:35:43.146 I don't have control over that. 00:35:43.546 --> 00:35:46.086 But you see the papers that are being written with them, no? 00:35:46.206 --> 00:35:49.146 Yeah, and there are many fantastic papers, which I admire and love. 00:35:49.246 --> 00:35:52.626 And there are some papers that I'm highly skeptical about and that I think shouldn't 00:35:52.626 --> 00:35:53.726 have been performed in that way. 00:35:54.186 --> 00:35:56.306 So should we improve our standards, do you think? 00:35:57.522 --> 00:36:00.002 Or you feel that things are... Improving standards is always better. 00:36:00.062 --> 00:36:01.042 The question is how to do that. 00:36:02.082 --> 00:36:06.882 I'm completely for improving standards. So the question then would become one, 00:36:06.922 --> 00:36:12.082 which is highly interesting, by the way, with the exploding amounts of journals 00:36:12.082 --> 00:36:14.402 in fields of neuroscience, 00:36:14.742 --> 00:36:17.302 computational neuroscience, cognitive neuroscience, neuroimaging, 00:36:18.002 --> 00:36:22.222 and the exponential explosion of papers appearing in these journals every week. 00:36:22.342 --> 00:36:27.222 Actually, Partha Mithra was discussing this today in his talk. How daunting that is. 00:36:27.522 --> 00:36:30.462 How do we control that? Yeah, how do we keep up a standard like that? 00:36:31.782 --> 00:36:35.842 That's something that it's a nice side issue that we could discuss for three 00:36:35.842 --> 00:36:38.222 quarters of an hour more. Maybe we do that in another podcast. 00:36:38.982 --> 00:36:41.802 But yeah, that's one of the challenges of science in general. 00:36:42.202 --> 00:36:48.102 Certainly a field like neuroscience that has exploded in volume over the last 20 to 30 years. 00:36:48.282 --> 00:36:54.402 Correct. So now if you had to convince people to use your method, 00:36:54.782 --> 00:37:01.002 what is this convincing test case that you would put forward? 00:37:01.082 --> 00:37:05.662 Where you say, look, there was this problem, this question that we could not answer before. 00:37:05.942 --> 00:37:10.302 But now we really gained fundamental insight in how the brain operates because 00:37:10.302 --> 00:37:15.042 of the data that we have access to, the imaging data, and the tools that we 00:37:15.042 --> 00:37:16.262 have developed to analyze it. 00:37:16.262 --> 00:37:19.802 What would be the... This would be hypothetical or already happened? 00:37:19.982 --> 00:37:21.342 No, no. It should have been published now. 00:37:21.622 --> 00:37:24.122 It should have been published now. Yeah, it should have already been in the 00:37:24.122 --> 00:37:27.382 literature. What's the key thing you would point to? All right. Good question. 00:37:28.102 --> 00:37:31.822 There's a few very good studies. Just give me one. I'm going to try not to name 00:37:31.822 --> 00:37:34.862 my own. That's very self-indulgent. So let's not do that. 00:37:35.402 --> 00:37:37.702 Let's not do that. Okay. 00:37:40.580 --> 00:37:43.420 There were a couple of interesting things discussed, so let's actually take 00:37:43.420 --> 00:37:45.760 another talk in the symposium over here. 00:37:45.820 --> 00:37:52.020 Maybe you've done a podcast with Melanie Bolli, who showed very interestingly, 00:37:52.040 --> 00:37:53.820 using the dynamic causal modeling framework, 00:37:54.180 --> 00:37:56.300 how she could distinguish between 00:37:56.300 --> 00:38:04.100 models where certain merely feed-forward processing was taking place, 00:38:04.180 --> 00:38:08.280 or fully recurrent processing was taking place, which supposedly can be one 00:38:08.280 --> 00:38:10.080 of the substrates of consciousness in the brain, 00:38:10.580 --> 00:38:13.140 and therefore could test these models against each other. 00:38:13.220 --> 00:38:17.640 One model where there's merely feed-forward processing going on and leading 00:38:17.640 --> 00:38:21.860 to some fit to the MEG data, in that case, EEG data. 00:38:22.380 --> 00:38:25.700 And the one where there is recurrent processing taking place. 00:38:26.780 --> 00:38:30.920 And the models that I advocate, the top-down models alone, could already distinguish 00:38:30.920 --> 00:38:34.560 between these two models and tell you that one fits the data much better than the other. 00:38:34.660 --> 00:38:37.860 And that therefore, supposedly, we can answer questions about consciousness. 00:38:37.860 --> 00:38:41.000 Is one of the most, you know, the biggest challenges that we have to face. 00:38:41.660 --> 00:38:43.640 So that already, I think, is convincing. 00:38:44.560 --> 00:38:47.200 And when we can take these models even further, like we've been discussing here, 00:38:47.260 --> 00:38:50.360 then maybe we can do even more than that. It's going to be fantastic. 00:38:50.660 --> 00:38:51.500 It's going to be fantastic. 00:38:52.400 --> 00:38:54.480 Very good. So now to conclude. 00:38:56.360 --> 00:38:59.940 So you're sort of coming up in this field. You're one of the rising stars, 00:39:00.220 --> 00:39:04.080 clearly in charge of what you do. 00:39:04.120 --> 00:39:07.840 You know what you're doing. So what's the law of Allard Raubruch that we should 00:39:07.840 --> 00:39:10.860 follow in order to make advances in our science? 00:39:11.160 --> 00:39:15.780 What's the law that you want to put out there for us to adhere to? Yeah, the law of Allard. 00:39:16.600 --> 00:39:19.660 Allard's law. I need a law. I didn't have a law yet. What will it be? 00:39:19.820 --> 00:39:22.360 So I have to make up a law now? Yep. Oh, wow. 00:39:23.350 --> 00:39:27.330 Oh, wow. But it can be one about the brain. 00:39:27.390 --> 00:39:33.550 It can be one of the methods we follow, the rules of behavior we adhere to as 00:39:33.550 --> 00:39:34.810 scientists. You're free. 00:39:35.630 --> 00:39:40.230 All right. There's one law, which in my own studies, my master science studies 00:39:40.230 --> 00:39:43.410 and my PhD studies, I've always loved. 00:39:43.890 --> 00:39:46.890 And it's not a law about the brain, but let me state that law and see if we 00:39:46.890 --> 00:39:50.090 can make it into another. The law that I've always loved is Hofstadter's law. 00:39:50.770 --> 00:39:55.530 And Hofstadter's law is about how painful it can be to program something, 00:39:55.750 --> 00:40:00.630 how precise you have to be, how meticulous you have to be, and how you should 00:40:00.630 --> 00:40:04.370 not think at any point that you are actually already done. So Hofstadter's law 00:40:04.370 --> 00:40:05.310 is actually a recurrent law. 00:40:06.510 --> 00:40:12.450 You know, Douglas Hofstadter always had a fond thing for recurrence. And it goes thus. 00:40:13.990 --> 00:40:17.850 A program is never done. 00:40:19.030 --> 00:40:23.110 Even when you take into account Hofstadter's law and your expectations about 00:40:23.110 --> 00:40:24.270 when the program will be done. 00:40:24.750 --> 00:40:29.610 So that kind of tells you that, okay, so if you think you're going to be done, 00:40:29.630 --> 00:40:30.590 you're not going to be done yet. 00:40:31.110 --> 00:40:35.990 So maybe I could paraphrase that into a law about. 00:40:37.170 --> 00:40:40.890 Neuronal modeling and modeling of the brain, which is that even though we think 00:40:40.890 --> 00:40:42.770 we'd be done with our models today, 00:40:42.950 --> 00:40:45.830 because for instance, we had causality from top-down models, 00:40:45.970 --> 00:40:51.030 we're not nearly done yet because there's always a higher level and also a higher 00:40:51.030 --> 00:40:54.930 level of abstraction and a lower level of biophysical detail that we can put 00:40:54.930 --> 00:40:56.690 into our models and learn to understand more. 00:40:56.890 --> 00:41:01.650 So we're never done with modeling the brain, even if you take into account Hofstadter's 00:41:01.650 --> 00:41:05.030 law for the brain, let's call it that. I couldn't honestly take credit for that myself. 00:41:05.290 --> 00:41:10.470 Okay. So, so Ehlers' law would be something like the analysis of effective connectivity is never done. 00:41:11.512 --> 00:41:17.832 Could always be better, yes. Very good. So then, if I would go and visit you 00:41:17.832 --> 00:41:22.552 five years from now, and say, look, Allard, in this podcast interview, 00:41:22.772 --> 00:41:25.952 you gave me this prediction, and now I want to know whether you were right or wrong. 00:41:26.032 --> 00:41:29.112 What's this one prediction you would really want to stick your neck out today, 00:41:29.852 --> 00:41:31.652 so that I can check five years from now? 00:41:32.052 --> 00:41:34.472 Oh, that's a nice one, especially when you've got it on a record. 00:41:35.252 --> 00:41:39.192 Are you doing stuff like this? Are you going to make a prediction that I can 00:41:39.192 --> 00:41:42.152 come and validate in five years in your lab? Sure. Yeah? 00:41:42.932 --> 00:41:44.512 Okay, you make yours only, Mark. 00:41:45.492 --> 00:41:50.712 My prediction is that we would have an integrated biophysically and anatomically 00:41:50.712 --> 00:41:57.872 constrained model of the core structures of the brain, including cerebellum, 00:41:57.872 --> 00:42:00.012 amygdala, hippocampus, thalamus, cortex, 00:42:00.352 --> 00:42:05.032 and components of the basal ganglia. At which level of detail? 00:42:06.852 --> 00:42:11.212 Including biophysical detail. So, there will be multi-compartmental models, 00:42:11.492 --> 00:42:14.832 we will have anatomical detail, connectivity, 00:42:15.232 --> 00:42:19.972 and we will be able to account for all aspects of classical and operant conditioning 00:42:19.972 --> 00:42:23.012 with these models. All right, perfect. So, now mine. 00:42:23.552 --> 00:42:28.652 My prediction would be that in five years, we've been able to push fMRI, 00:42:28.852 --> 00:42:32.432 Functional Magnetic Resonance Imaging, especially at ultra-high fields, 00:42:32.572 --> 00:42:39.292 7 Tesla and beyond, to such a degree that we can validate models such as which 00:42:39.292 --> 00:42:41.532 you are talking about in your prediction 00:42:41.532 --> 00:42:47.052 at the level of cortical columnar structures and cortical layers, 00:42:47.292 --> 00:42:51.092 which is something that we can only dream about today, even with modern 3D machines. 00:42:51.492 --> 00:42:54.592 My prediction would be in five years, we're going to be very close to that. 00:42:54.712 --> 00:42:59.572 And that is actually going to be the convincing reason why to to use this type 00:42:59.572 --> 00:43:02.832 of data as your added constraints onto those bottom-up models. 00:43:03.112 --> 00:43:06.252 Great. Alan Raubrook, thank you very much. Pleasure.