WEBVTT 00:00:00.017 --> 00:00:07.937 This is Paul for sure. I'm talking to Chao-Jing Wang, one of our speakers in the summer school. 00:00:08.417 --> 00:00:14.377 And in your talk, you focus very much on the role of working memory, 00:00:14.617 --> 00:00:16.317 sustained activity and cognition. 00:00:16.317 --> 00:00:18.797 You started with a fairly strong 00:00:18.797 --> 00:00:24.817 claim where you said that actually delay activity is the key to cognition. 00:00:25.297 --> 00:00:30.717 So why do you think this is the key ingredient that we should worry about? 00:00:33.097 --> 00:00:39.857 Right. I guess the general idea is that imagine that you cannot hold in your 00:00:39.857 --> 00:00:44.177 mind anything in absence of direct sensory stimulation. 00:00:44.177 --> 00:00:49.757 Then it seemed to me that most of the repertoire of behaviors is going to be 00:00:49.757 --> 00:00:53.317 reduced to reflex all you can do is you know just, 00:00:54.157 --> 00:01:01.537 reactively respond to stimuli right away otherwise you wait you forgot what 00:01:01.537 --> 00:01:06.757 was the stimulus so you cannot really act accordingly and you become enslaved 00:01:06.757 --> 00:01:11.597 to the external world whereas if you have this ability to hold something in 00:01:11.597 --> 00:01:13.777 your mind even when the input is gone on, 00:01:13.797 --> 00:01:20.717 then you are freed from the immediate stimulation and so that you can become 00:01:20.717 --> 00:01:22.317 a lot more flexible, right? 00:01:22.697 --> 00:01:28.857 You can, for example, wait and decide what you do about the stimulus and still 00:01:28.857 --> 00:01:30.757 remembering what was the stimulus. 00:01:31.197 --> 00:01:38.057 Right. Yeah. And then as an implementation of this, you were pointing very much to cortical circuits. 00:01:38.617 --> 00:01:42.157 So what's so special then about cortical circuits that they can actually can 00:01:42.157 --> 00:01:43.757 give you this kind of memory functions. 00:01:45.383 --> 00:01:51.723 Right. So, in fact, persistent activity itself is probably more widespread than 00:01:51.723 --> 00:01:56.303 working memory-related persistent activity. So, I want to distinguish those two things. 00:01:56.703 --> 00:02:04.363 For example, even when you try to hold a gaze, your eyes are fixating on something. 00:02:04.903 --> 00:02:10.463 That's maintained by persistent activity in the ocular motor system outside of the cortex. 00:02:11.443 --> 00:02:19.423 But there's a very long history of studies pointing to the prefrontal cortex, 00:02:20.143 --> 00:02:22.723 as a very important structure for working memory. 00:02:23.003 --> 00:02:30.263 So that dates back to, I guess, 1920s and 30s, where people showed that if you 00:02:30.263 --> 00:02:35.603 do lesion of the prefrontal cortex on an animal like a monkey, 00:02:35.763 --> 00:02:39.743 we're not able to do delayed response tasks, which depend on working memory. 00:02:39.743 --> 00:02:44.863 So from there on, I think, you know, there is a lot of studies showing that 00:02:44.863 --> 00:02:49.703 prefrontal cortex is really important for, you know, working memory. 00:02:50.583 --> 00:02:55.943 And that's why there's a lot of focus on cortex. 00:02:56.963 --> 00:03:02.963 But then you were saying that it is the special characteristic of this dense 00:03:02.963 --> 00:03:07.743 recurrence in these circuits that could sustain this memory function. 00:03:08.083 --> 00:03:13.763 Is that correct? That's the idea, yeah. Okay, so then I could argue that actually 00:03:13.763 --> 00:03:16.663 this kind of dense recurrence I would find throughout the cortex. 00:03:17.043 --> 00:03:21.583 So why is then not my occipital cortex, that's usually more dedicated to vision, 00:03:21.763 --> 00:03:24.683 actually a working memory system showing sustained activity? 00:03:25.443 --> 00:03:27.803 What's then so special about these frontal areas? 00:03:30.070 --> 00:03:32.190 That they can support working memory. 00:03:34.430 --> 00:03:42.610 So I guess, empirically, we do not have enough data to answer your question explicitly. 00:03:43.650 --> 00:03:47.610 But generally, it probably is a matter of degree, right? 00:03:47.710 --> 00:03:53.130 So we know even in the primary sensory areas, the majority of synaptic inputs 00:03:53.130 --> 00:03:55.210 are coming from within the local circuits. 00:03:55.210 --> 00:04:04.130 But maybe there's just more in the brief on the cortex compared to sensory areas. 00:04:04.330 --> 00:04:08.470 And what's interesting from the computational point of view is that you can 00:04:08.470 --> 00:04:13.230 show that because of these feedback systems, your dynamics is nonlinear. 00:04:13.490 --> 00:04:19.730 And in nonlinear systems, I think it's very important that just some graded 00:04:19.730 --> 00:04:24.950 differences, some quantitative differences, for example, in the amount of recurrent connections. 00:04:25.210 --> 00:04:28.630 Can lead to qualitatively different behaviors. 00:04:29.630 --> 00:04:34.430 Like, you know, if your recurrent connections is below a threshold level, 00:04:34.590 --> 00:04:35.670 you don't see persistent activity. 00:04:35.970 --> 00:04:38.850 If it's above a threshold level, you start to see persistent activity. 00:04:39.150 --> 00:04:44.250 So that may be my take about that. Just to add that it could also be there's 00:04:44.250 --> 00:04:48.790 something else, such as neuromodulation, which may be somewhat different. 00:04:48.950 --> 00:04:53.330 So, for example, dopamine modulation ventilation may be, say, 00:04:53.570 --> 00:04:59.350 somewhat more prominent in the peripheral cortex than in early sensory systems, 00:04:59.590 --> 00:05:02.570 for example, that could make a difference as well. 00:05:02.790 --> 00:05:07.910 But I would say, again, it's a matter of quantitative differences leading to 00:05:07.910 --> 00:05:08.790 qualitative differences. 00:05:09.590 --> 00:05:14.830 Okay. But that's still a very much cortex-centered view. 00:05:15.110 --> 00:05:19.930 So does it mean that you would be making the strong statement that okay working 00:05:19.930 --> 00:05:26.070 memory can be fully realized by cortical circuits and does not depend on subcortical 00:05:26.070 --> 00:05:30.130 activity beyond possibly some forms of neuromodulation so. 00:05:31.510 --> 00:05:37.550 Yeah we don't know much about that i guess there's one thing certainly people think, 00:05:38.630 --> 00:05:44.470 um for why subcortical structures might be important that is skating so you 00:05:44.470 --> 00:05:50.570 don't want not any sensory stimulus coming into your brain to be stored in working memory, right? 00:05:50.670 --> 00:05:54.830 So basically, you somehow have to gauge what really is behaviorally important 00:05:54.830 --> 00:05:59.550 that you have to maintain internally, let's say, in the prefrontal cortex. 00:06:00.530 --> 00:06:04.550 And that seems to depend on basal ganglia. There's this idea that, 00:06:04.550 --> 00:06:10.330 you know, basal ganglia, for example, through the thalamus can gauge what signal 00:06:10.330 --> 00:06:11.530 is important and what is not. 00:06:11.530 --> 00:06:19.690 At least you know part of the gating function is dependent on basal ganglia okay yeah so, 00:06:20.879 --> 00:06:25.519 It could also be that persistent activity itself may, in part, 00:06:25.739 --> 00:06:32.299 depend on reverberation in a bigger loop involving the subcortical system. 00:06:33.559 --> 00:06:36.739 But for that, people have speculated about that for a long time. 00:06:37.019 --> 00:06:42.979 As far as I know, we do not yet have very strong evidence for that. Yeah. 00:06:43.419 --> 00:06:48.859 So then the core ingredients of your model will be, let's say, 00:06:48.959 --> 00:06:57.919 recurrently coupled excitatory circuits with, let's say, some inhibitory add-ons to get selectivity. 00:06:57.919 --> 00:07:01.779 And this is, I think, some summary of a cortical circuit. 00:07:02.019 --> 00:07:08.419 And this would then project downstream to, let's say, other subcortical structures 00:07:08.419 --> 00:07:15.499 to trigger action, as in, for instance, your standard two-choice saccade-based tasks. 00:07:15.659 --> 00:07:18.399 It would be like the choice aspects would happen in this cortical model, 00:07:18.579 --> 00:07:22.799 and then the saccade would be triggered by, let's say, superior colliculus to 00:07:22.799 --> 00:07:25.599 which this decision-making system would project. 00:07:25.599 --> 00:07:30.119 So is that an assumption you really would like to insist on, 00:07:30.179 --> 00:07:35.699 that these action elements are extra cortical, and that more the sensory-based 00:07:35.699 --> 00:07:39.399 and the rule-based aspects are processed at the cortical level? 00:07:44.699 --> 00:07:51.819 Well, that, I guess, depends. So in a way, saccade is special, right? 00:07:51.879 --> 00:07:58.719 So if you talk about manual responses, then maybe parts of the motor cortex 00:07:58.719 --> 00:08:00.379 are the command centers. 00:08:01.199 --> 00:08:10.099 And so that would be kind of downstream system from decision circuits where 00:08:10.099 --> 00:08:12.879 information are integrated, maybe a choice is produced. 00:08:13.359 --> 00:08:20.439 So in that case, then even the, I think it's the part of the response generation 00:08:20.439 --> 00:08:23.399 is occurring inside the cortex, 00:08:23.699 --> 00:08:28.359 you know, always together with basal ganglia and some others like thalamus, 00:08:28.519 --> 00:08:33.879 but in that case, you know, a lot of things may happen in the cortex as well. Okay. 00:08:34.499 --> 00:08:39.259 So your, the approach, I mean, also given your background in physics, 00:08:39.539 --> 00:08:44.359 you have used one of these tools that has been important to theoretical neuroscience 00:08:44.359 --> 00:08:48.759 from physics of, let's say, attractor networks to analyze this model. 00:08:50.099 --> 00:08:55.019 So why is the notion of an attractive network helping you to understand this 00:08:55.019 --> 00:08:57.679 pre-protocol, these cortical dynamics? 00:09:02.020 --> 00:09:05.000 You know sometimes i like to say especially to 00:09:05.000 --> 00:09:07.960 experimentalists that the the uh the 00:09:07.960 --> 00:09:10.720 word attractor networks seem to 00:09:10.720 --> 00:09:15.920 have certain seem to provoke certain reactions in some people right so i would 00:09:15.920 --> 00:09:23.380 sometimes start out uh by saying that uh attractor dynamics is not like a black 00:09:23.380 --> 00:09:28.460 hole so it's not something that's very rigid that if you are in a tractor state 00:09:28.460 --> 00:09:30.700 then it's like if you're You're sucked in. 00:09:30.780 --> 00:09:33.820 You cannot get out. And it's a very rigid kind of thing. 00:09:35.060 --> 00:09:38.900 Attracted states are simply relatively stable states. 00:09:39.960 --> 00:09:46.060 That's all. It could even be something, you know, more than just a steady state. 00:09:46.300 --> 00:09:51.080 For example, you could have chaotic attractors where you have a lot of temporal 00:09:51.080 --> 00:09:56.720 dynamics going on inside that, you know, chaotic attracted state, right? 00:09:56.880 --> 00:10:02.280 That's one thing. The other thing is that any inputs or neural modulation can 00:10:02.280 --> 00:10:06.780 easily change, if you like, the landscape of attractive states. 00:10:07.060 --> 00:10:09.220 So you can easily actually control, 00:10:09.580 --> 00:10:14.300 manipulate, you know, the landscape of multiple attractive states. 00:10:14.700 --> 00:10:19.840 So, you know, simply, as I was saying, you know, attractive states are simply 00:10:19.840 --> 00:10:21.120 relatively stable states. 00:10:21.420 --> 00:10:27.640 And if you want to describe mathematically what is a persistent activity, right? 00:10:29.280 --> 00:10:36.080 Then it's natural to think about that as a relatively stable state that's all, right? 00:10:36.140 --> 00:10:41.360 So as a result the concept of attractive states is very natural people still 00:10:41.360 --> 00:10:44.740 debate about the actual dynamical structures, 00:10:45.500 --> 00:10:51.720 of persistent activity that should be described well by attractive networks, 00:10:52.040 --> 00:10:57.880 but that's I think a separate issue from a more general question whether, 00:10:58.660 --> 00:11:03.480 attractor networks is the right framework or not to describe persistivity. 00:11:03.880 --> 00:11:11.500 Because I could argue that in some sense any system that will show some persistence 00:11:11.500 --> 00:11:16.360 of whatever spatial temporal scale you want to express this I can re-describe 00:11:16.360 --> 00:11:17.720 in terms of an attractor network. 00:11:18.120 --> 00:11:23.360 So the risk might then be that the framework of an attractor is so, 00:11:23.420 --> 00:11:26.660 let's say, super powerful powerful, that it might not give you much leverage 00:11:26.660 --> 00:11:29.300 with respect to a specific phenomenon. 00:11:30.300 --> 00:11:34.260 So what's the leverage it has given you to really understand prefrontal cortex? 00:11:36.184 --> 00:11:40.384 Sorry, I'm not sure I really understand the... Well, it's very simple. 00:11:40.624 --> 00:11:44.224 In some sense, you could argue that if you look at the nervous system, 00:11:44.304 --> 00:11:50.824 you will find at many different levels, you will find persistent states in some temporal window. 00:11:51.004 --> 00:11:54.164 It can be a microsecond. It doesn't matter. Yeah. 00:11:54.764 --> 00:11:58.344 In all those cases, I could step in with an attractor network formulation. 00:11:58.464 --> 00:12:02.184 Ah, you see, it's an attractor because the system in some way is returning to 00:12:02.184 --> 00:12:04.784 this state and then I define that state in some way. 00:12:05.744 --> 00:12:10.844 So that means the attractor framework is super powerful. 00:12:11.024 --> 00:12:14.264 You can describe anything with it you want, as long as there's some persistence 00:12:14.264 --> 00:12:19.024 in the system, in some definition of its possible states. 00:12:19.484 --> 00:12:22.864 I see. So what's special about, say, prefrontal cortex? Exactly. 00:12:22.984 --> 00:12:25.564 Why does it give you leverage to understand prefrontal cortex? 00:12:25.804 --> 00:12:31.244 I guess the important thing is that you have multiple attractor states at the same time. 00:12:32.204 --> 00:12:35.044 If you think about if you want to design 00:12:35.044 --> 00:12:41.144 a working memory system it basically is a system with multiple states so that 00:12:41.144 --> 00:12:47.824 you can switch on and off between different states and that may not be so universal 00:12:47.824 --> 00:12:52.904 in different systems for example you don't need and you don't want multiple. 00:12:54.224 --> 00:13:01.224 States in not necessarily in early sensory systems for example And so even if 00:13:01.224 --> 00:13:02.284 you have some persistency, 00:13:02.344 --> 00:13:11.524 the question is whether you have this ability to go back and forth between many 00:13:11.524 --> 00:13:14.024 different states with very brief input. 00:13:16.680 --> 00:13:22.080 Does that make sense? So that seemed to be the distinction between what you 00:13:22.080 --> 00:13:26.760 want for working memory system versus a sensory processing device. 00:13:27.300 --> 00:13:32.820 Okay, so you're saying the framework of attractors gives you leverage in this case, 00:13:32.900 --> 00:13:39.400 because actually you're looking at a system that can be in a large number of 00:13:39.400 --> 00:13:43.140 possible states, and this is an efficient way to describe these. That's right. 00:13:43.240 --> 00:13:49.640 And the persistency should be at the long time scale compared to the time constants 00:13:49.640 --> 00:13:57.120 you have in the system, like milliseconds versus tens of milliseconds by physical time constant. 00:13:57.120 --> 00:14:06.420 Let me also just briefly say that, as I was discussing in my lecture, that, 00:14:06.560 --> 00:14:13.080 in fact, fast switches tend to be not the right conceptualization in the model I described. 00:14:13.780 --> 00:14:19.680 So you also have very slow transients, too, and that's not just steady states. 00:14:19.940 --> 00:14:24.420 And those slow transients, like a gradual ramping activity, turns out to be 00:14:24.420 --> 00:14:30.600 a very good computational mean to, say, integrate information in decision-making. 00:14:30.940 --> 00:14:36.840 So this is kind of an unusual type of attractive networks where you, 00:14:36.940 --> 00:14:42.880 at the same time, have multiple stable states, But you also have this ability 00:14:42.880 --> 00:14:47.620 to have very slow transients for time integration. 00:14:48.960 --> 00:14:56.440 Okay, but do you see that as an intrinsic property of these circuits or as an 00:14:56.440 --> 00:15:01.000 add-on that might be supported by a different substrate? Right. 00:15:03.632 --> 00:15:11.092 So at least for the mechanism we found in the model, it's the same circuit. 00:15:11.232 --> 00:15:17.572 So the same circuit, because the recurrent dynamics is mediated by, 00:15:17.592 --> 00:15:21.612 you know, kind of slower cellular mechanism, 00:15:21.892 --> 00:15:27.572 you can have both slow transients and multiple statist. Okay. Yeah. 00:15:27.872 --> 00:15:33.012 Whether, you know, in the brain, those two things, actually, 00:15:33.072 --> 00:15:36.772 there is evidence, I guess, at a single neuron level that in the brain, 00:15:36.832 --> 00:15:41.732 those two things can happen at least in the same circuits, right? 00:15:42.112 --> 00:15:48.672 Yes. And actually very commonly, you observe decision-related neural signals 00:15:48.672 --> 00:15:53.832 in a decision task, and working memory-related signals in a working memory task, 00:15:54.692 --> 00:15:59.412 in very different experiments done in different labs, but from the same code. 00:16:00.592 --> 00:16:03.552 Adjusting, you have a shared mechanism. Okay. 00:16:05.112 --> 00:16:10.392 Although you can never exclude that these might be interspersed circuits that 00:16:10.392 --> 00:16:13.892 perform different functions, no? Right, in theory, sure. 00:16:14.652 --> 00:16:19.072 But before we move into this issue of timing and decision-making, 00:16:19.292 --> 00:16:23.892 another issue around attractor networks is that in order to be, 00:16:23.912 --> 00:16:29.952 let's say, a certified and card-carrying member of the club of attractor networks, 00:16:30.292 --> 00:16:32.792 you must satisfy certain minimal criteria. 00:16:33.112 --> 00:16:39.592 Like when it's under perturbation, you must resort to the same attractor state, and so on. 00:16:39.692 --> 00:16:44.612 So are you sure that these cortical, the frontal cortical network should look 00:16:44.612 --> 00:16:46.932 at do satisfy all these conditions? 00:16:52.974 --> 00:16:59.274 Well, I guess that's rather difficult to test, especially in behaving animals 00:16:59.274 --> 00:17:01.974 or behaving primates, especially. 00:17:02.354 --> 00:17:10.814 So I imagine that if people now are making efforts to develop maybe simpler 00:17:10.814 --> 00:17:13.934 model systems like rodents, 00:17:14.534 --> 00:17:18.034 if you can design a good, say, working memory or decision-making task, 00:17:18.034 --> 00:17:20.994 it's probably more likely you 00:17:20.994 --> 00:17:24.394 can really go down into microcircuit mechanisms and get 00:17:24.394 --> 00:17:30.114 into more detailed information however you 00:17:30.114 --> 00:17:36.714 can you know do certain experimental manipulations to test some model predictions 00:17:36.714 --> 00:17:43.974 so for example as I showed that you know mentioned earlier that we found that 00:17:43.974 --> 00:17:45.374 that reverberation should be slow. 00:17:45.574 --> 00:17:49.334 In particular, it probably depends on a particular receptor, 00:17:49.474 --> 00:17:53.074 NMDA receptors at the recurrent synopsis, right? 00:17:53.254 --> 00:17:57.694 And we, in collaboration with Amy Anstin at Yale, actually we tested this idea 00:17:57.694 --> 00:18:02.754 in behaving monkeys, you know, working memory task, using a technique called 00:18:02.754 --> 00:18:06.214 ionophoresis to inject the drug locally. 00:18:08.274 --> 00:18:12.614 Onto neurons you record from. So you see persistent activity in those neurons 00:18:12.614 --> 00:18:15.494 in the prefrontal cortex of a behavior monkey. 00:18:16.654 --> 00:18:20.554 And then when you apply the drug, the blocks and MDA receptors, 00:18:20.834 --> 00:18:22.994 you actually see that process activity goes away. 00:18:24.014 --> 00:18:30.494 That's, in my mind, a very direct confirmation that MDA receptors that are slow, 00:18:31.334 --> 00:18:36.594 mediating slow repopulation, are critical for process activity. 00:18:37.434 --> 00:18:42.054 Okay, yes, I see that. But then I could argue, well, but maybe you're blocking... 00:18:42.614 --> 00:18:45.194 The NMDA receptors are the part of the thalamic projections. 00:18:45.474 --> 00:18:48.694 Can you exclude that? No, you cannot. 00:18:50.654 --> 00:18:55.934 That's right. But at least we can say it. 00:18:58.402 --> 00:19:01.762 Are likely to be kind of intrinsic synapses 00:19:01.762 --> 00:19:05.162 rather than you know synapses coming 00:19:05.162 --> 00:19:08.142 from external stimulation because this is 00:19:08.142 --> 00:19:11.002 during working memory during the delay when there's no direct 00:19:11.002 --> 00:19:14.182 sensory stimulus right um so 00:19:14.182 --> 00:19:19.182 to adjust your question you can do a different set of experiments but not with 00:19:19.182 --> 00:19:24.902 monkey so with rodent you can do in vitro slices right and actually we have 00:19:24.902 --> 00:19:31.102 done in collaboration with another group by Wenjing Gao in Philadelphia where 00:19:31.102 --> 00:19:34.162 you do prefrontal cortical sizes directly. 00:19:34.362 --> 00:19:39.922 You can look at, you know, NMDA or AMPA receptor immediately the transmission 00:19:39.922 --> 00:19:43.262 between two cells neighboring neurons. 00:19:43.582 --> 00:19:49.342 So there's really local connections between two neurons and show how much NMDA 00:19:49.342 --> 00:19:54.042 you have at this very local synopsis. 00:19:54.582 --> 00:19:59.942 In peripheral neurons, and you can compare that with the same kind of peer recordings 00:19:59.942 --> 00:20:03.942 in the primary sensory system. You see a big difference. 00:20:04.442 --> 00:20:08.542 Okay. That's also consistent. Right. So with these experiments... 00:20:08.542 --> 00:20:10.762 So that would be really cortical, right? Local. Yes. 00:20:11.082 --> 00:20:14.542 But with the antiparesis experiment, basically what you're saying is, 00:20:14.602 --> 00:20:19.682 well, this makes it plausible possible that the memory state is implemented 00:20:19.682 --> 00:20:22.342 by a recurrent or reverberating circuit. 00:20:22.582 --> 00:20:28.262 But in some sense, what we still don't know is whether it really can be called an attractor or not. 00:20:28.502 --> 00:20:33.122 So would you argue that, for instance, these micro-stimulation experiments that 00:20:33.122 --> 00:20:37.602 have been performed in front of light fields that would sort of bias decision-making 00:20:37.602 --> 00:20:43.042 of an animal, would you take that as corroborative evidence of having attractor states? Yes. 00:20:44.152 --> 00:20:48.512 Or would that be the kind of way to get at that question? Well, 00:20:48.612 --> 00:20:54.232 I think in order to go further along the question, I mean, what I suggest, 00:20:54.712 --> 00:21:01.152 we should propose something very well defined that's an alternative to the attractive model, right? 00:21:01.532 --> 00:21:08.452 And otherwise, it's hard to say, how do you prove or disprove the attractive network paradigm? 00:21:08.992 --> 00:21:13.272 So I guess one alternative is slow transients. 00:21:14.152 --> 00:21:20.152 Just because of some cellular or whatever, you know, biophysical process with 00:21:20.152 --> 00:21:22.392 very slow time constant, right? 00:21:22.552 --> 00:21:28.012 And so as a result, in response to transient stimulus, it just keeps going for 00:21:28.012 --> 00:21:31.332 a long time on the timescale of many seconds, right? 00:21:32.592 --> 00:21:37.772 So that's something we can try to, right, to contrast with a charted network. 00:21:38.992 --> 00:21:45.232 And I'm sure slow processes are playing a role The question is whether it's 00:21:45.232 --> 00:21:47.532 the workhorse, the main thing, right? 00:21:48.312 --> 00:21:55.772 I do see a possible problem with a mechanism that's completely based on very slow process. 00:21:56.352 --> 00:22:01.112 Because if you have a system with intrinsically very slow time constant. 00:22:02.112 --> 00:22:10.092 As you know, you would have a hard time to change the system with briefing. 00:22:10.972 --> 00:22:15.172 Because the time constant is so short and so long, You have to use very long 00:22:15.172 --> 00:22:18.072 Z inputs to do anything with the system. 00:22:18.232 --> 00:22:20.772 So switching on and off becomes a problem. 00:22:21.232 --> 00:22:27.852 Right. Yeah. So maybe people can think about a clever way to really propose 00:22:27.852 --> 00:22:33.832 very clearly defined alternatives, right? And then try to design. Right. 00:22:34.732 --> 00:22:39.472 So what was very exciting about your proposal is that you also made the argument, 00:22:39.532 --> 00:22:44.212 look, I can represent the memory state. I can represent the decision-making 00:22:44.212 --> 00:22:46.092 by switching between my different detractors. 00:22:46.152 --> 00:22:50.492 But on top of that, I can also capture the kind of ramping functions that have 00:22:50.492 --> 00:22:54.072 to do with the reaction times as observed in prefrontal cortex, right? 00:22:54.152 --> 00:22:57.912 Where you see that, okay, these neurons that seem to correlate with a decision 00:22:57.912 --> 00:23:02.252 ramp more slowly when you have long reaction times and they ramp very fast when 00:23:02.252 --> 00:23:04.272 you have a fast reaction time. 00:23:04.632 --> 00:23:09.992 So how could that drop out of your model so easily? 00:23:10.172 --> 00:23:12.652 What was the trick there? Why does that work? 00:23:16.938 --> 00:23:21.338 Well, I guess the main thing is this idea of slow reverberation, I guess. 00:23:21.678 --> 00:23:24.698 And that was a surprise to us. 00:23:25.018 --> 00:23:31.398 You know, our priority, as I was saying, you know, working memory itself, in principle, right? 00:23:31.458 --> 00:23:35.538 If you're an engineer, you think about how to design a working memory device, 00:23:35.918 --> 00:23:38.378 you could say that can be done by faster switches. 00:23:39.338 --> 00:23:44.438 But when you try to do that with neurons and the realistic synaptic connections, 00:23:44.438 --> 00:23:51.938 you know, it turns out that the system is very unstable because of the strong feedback, 00:23:52.198 --> 00:23:54.598 you know, machinery in the system. 00:23:54.838 --> 00:24:02.858 And one way to solve that is to say positive feedback needs to be slow relative to negative feedback. 00:24:03.038 --> 00:24:07.158 And that so was kind of forced on us, you know, on us. 00:24:07.238 --> 00:24:10.918 So we say, you know, we have to have a working memory mechanism with 00:24:10.918 --> 00:24:13.938 slow reverberation rather than very fast positive feedback 00:24:13.938 --> 00:24:17.438 and from there it turns out um 00:24:17.438 --> 00:24:22.578 you know uh it becomes easy so so this slow river version turns out to be exactly 00:24:22.578 --> 00:24:28.338 what you need to get a slow ramping activity in decision that's i thought it's 00:24:28.338 --> 00:24:34.438 quite nice yeah but now one one question i would have there is um. 00:24:35.478 --> 00:24:38.158 In some sense in in this manipulation you could have 00:24:38.158 --> 00:24:41.058 a confound that um what we are 00:24:41.058 --> 00:24:44.578 what we are detecting what we are detecting is based 00:24:44.578 --> 00:24:47.518 on moving dots essentially and 00:24:47.518 --> 00:24:50.198 it's about the coherence of the moving dots that you 00:24:50.198 --> 00:24:53.178 make your decisions so i could 00:24:53.178 --> 00:24:56.298 argue look if i have these prefrontal neurons that 00:24:56.298 --> 00:24:59.498 are sensitive to the moving dots then that 00:24:59.498 --> 00:25:02.818 if i if they have an orientation tuning 00:25:02.818 --> 00:25:05.898 then of course i'm driving them more effectively if 00:25:05.898 --> 00:25:09.178 i have coherent movement in my scene than if i have if 00:25:09.178 --> 00:25:11.998 i drive all these are incoherently so that basically means per unit 00:25:11.998 --> 00:25:14.898 time they get less energy and therefore as a result 00:25:14.898 --> 00:25:19.918 i'll be ramping up more slowly or faster you see and this then also exactly 00:25:19.918 --> 00:25:24.918 correlates with the task condition so imagine i would change the task now that 00:25:24.918 --> 00:25:30.218 the animal has to make a decision based on let's say um when when there's incoherent 00:25:30.218 --> 00:25:33.518 movement it gets reward and it should ignore coherent movement so i sort of 00:25:33.518 --> 00:25:34.678 i change the contingency, 00:25:34.818 --> 00:25:39.058 would you predict the model would still work or would you have to add a new feature? 00:25:39.958 --> 00:25:46.658 So what we added, we do need to add something in that case that is reward-dependent plasticity. 00:25:46.778 --> 00:25:55.458 So in that case, you have to learn what are the potential outcomes from your choice options, right? 00:25:55.618 --> 00:26:02.638 And that, we believe, is done through learning that depends on say, reward. Okay? 00:26:03.998 --> 00:26:11.998 I think what you are pointing earlier at is the confine that maybe slow RAM 00:26:11.998 --> 00:26:15.038 just corresponds to weaker inputs. 00:26:17.065 --> 00:26:22.965 But in part it's true. So basically, you probably need to integrate more over 00:26:22.965 --> 00:26:26.545 time if your input, your evidence is weaker. 00:26:27.345 --> 00:26:33.585 But it's not just that, because, for example, we showed that if you have more 00:26:33.585 --> 00:26:39.225 several options, more options to consider, then the reaction times are also slower. 00:26:39.225 --> 00:26:44.865 And that in part is because you have this competition between your own pools 00:26:44.865 --> 00:26:48.065 selected for say four or five options. 00:26:48.705 --> 00:26:53.585 And that involves inhibition because it's a competition mediated by inhibition 00:26:53.585 --> 00:26:55.965 that also slows down the ramping activity. 00:26:56.285 --> 00:27:00.405 So you can also show nicely, you know, something that people see in psychology, 00:27:00.605 --> 00:27:04.945 you know, the more options you have to consider, the slower the reaction time. 00:27:05.225 --> 00:27:08.505 Okay. Yeah. Yes. so the other. 00:27:09.605 --> 00:27:13.065 Aspect that I was curious about if you look at the model it's 00:27:13.065 --> 00:27:16.465 also an issue we discussed earlier the interpretation of 00:27:16.465 --> 00:27:19.345 this prefrontal cortical function also in the literature at 00:27:19.345 --> 00:27:22.025 large it's in the end very much a labeled line kind 00:27:22.025 --> 00:27:26.565 of system I have choice options the choice options depend on certain sensory 00:27:26.565 --> 00:27:31.365 cues and if you want by magic they just come together in these units or in my 00:27:31.365 --> 00:27:37.605 attractor network and now I can start to make decisions with them but you could 00:27:37.605 --> 00:27:41.025 then of course pose the question, and sometimes it's a form of the symbol grounding problem, 00:27:41.345 --> 00:27:46.365 okay where do these labeled lines come from and should I really assume that 00:27:46.365 --> 00:27:49.825 these are labeled lines, that let's say your prefrontal cortex has all these 00:27:49.825 --> 00:27:52.245 labeled lines projecting into it from other areas, 00:27:53.085 --> 00:27:57.305 representing all possible cues you can ever encounter or you have ever encountered 00:27:57.305 --> 00:28:01.845 in the world combined with all possible actions that you can ever trigger in 00:28:01.845 --> 00:28:04.585 response to these cues, but that would be the labeled line view. 00:28:04.785 --> 00:28:10.505 So do you think that that's a reasonable assumption, or do we have to get away from it in some way? 00:28:12.825 --> 00:28:16.905 Right. So I guess you're totally right. 00:28:17.145 --> 00:28:23.545 We have been focusing on certain elementary and fundamental, 00:28:24.965 --> 00:28:30.645 machineries about working memory and decision-making without paying too much 00:28:30.645 --> 00:28:33.585 attention to real-life stimuli, right? 00:28:34.085 --> 00:28:36.065 And how do we really process real-life stimuli? 00:28:36.985 --> 00:28:44.565 So I guess in my mind, to go forward in that direction, we have to understand 00:28:44.565 --> 00:28:49.225 better how objects are recognized and represented in the brain. 00:28:49.445 --> 00:28:55.585 And I could imagine, for example, in the inferior temporal cortex. 00:28:56.925 --> 00:29:02.965 There we don't really understand yet how we recognize objects fully. 00:29:02.965 --> 00:29:06.325 So it could be gilet l'ail, but it could be something more. 00:29:07.406 --> 00:29:11.266 You know people talk about grandmother cells that's more 00:29:11.266 --> 00:29:14.406 like data line I guess in your terminology 00:29:14.406 --> 00:29:17.726 but or maybe something 00:29:17.726 --> 00:29:25.206 more dynamic right we don't know yet what that is but to answer your question 00:29:25.206 --> 00:29:30.046 I can envision that you know starting with the building blocks basically elementary 00:29:30.046 --> 00:29:35.126 elementary machinery we kind of have some insights into 00:29:35.326 --> 00:29:37.206 four working memory systems, 00:29:37.586 --> 00:29:44.526 we can envision to connect that with something like a sensory, 00:29:45.086 --> 00:29:49.766 visual system, including IT, for example, and see how the interaction between 00:29:49.766 --> 00:29:52.226 a working memory system and the rest, 00:29:52.406 --> 00:29:58.606 the posterior part of the visual system, together in a larger scale brain system 00:29:58.606 --> 00:30:05.666 to carry out more realistic kind of representation and the working memory. 00:30:06.066 --> 00:30:11.886 And I do think that's a major challenge in the field. How we have in Pakistan 00:30:11.886 --> 00:30:13.806 mostly local circuits, right? 00:30:13.926 --> 00:30:21.906 And we have to come up with a theoretical framework, new concepts perhaps perhaps 00:30:21.906 --> 00:30:28.286 to understand large-scale brain system with interacting parts. Exactly. 00:30:29.046 --> 00:30:35.106 So to go in that direction, if you look at the physiology on prefrontal cortex, 00:30:35.446 --> 00:30:39.726 these classic experiments by Assad, Miller, and so on, in the end you look at 00:30:39.726 --> 00:30:44.226 populations of neurons that actually have a very broad range of representations, 00:30:44.646 --> 00:30:48.426 of cues, actions, and their combinations following different rules. 00:30:48.886 --> 00:30:53.186 So, could you then imagine that maybe you have, let's say, more specialization 00:30:53.186 --> 00:30:57.206 in these prefrontal circuits, that some neurons, let's say, contribute to cue information, 00:30:57.626 --> 00:31:00.886 others to action information, and then in their interaction, 00:31:01.106 --> 00:31:04.806 they will build up this, what you would call the attractor, representing this 00:31:04.806 --> 00:31:06.806 key state in which you want to make your decisions. 00:31:06.886 --> 00:31:08.986 Would that be a reasonable alternative? Yes. 00:31:11.394 --> 00:31:21.774 Well, actually, so we have done more recent work with Stefano Fusi and Mattia, his graduate student. 00:31:22.274 --> 00:31:27.594 That work suggests, this is a computational work, suggests that, 00:31:27.674 --> 00:31:34.694 in fact, the neuron signals in the peripheral cortex probably should have a 00:31:34.694 --> 00:31:36.014 lot of mixed selectivity. 00:31:36.014 --> 00:31:43.054 So a given neuron would be, you know, activated to different degrees by a combination 00:31:43.054 --> 00:31:46.234 of many different things, including, you know, sensory stimuli, 00:31:46.434 --> 00:31:49.174 internal representation of behavior rules, 00:31:49.434 --> 00:31:52.454 and maybe some control signals altogether. 00:31:52.454 --> 00:31:55.354 Together so i my view is 00:31:55.354 --> 00:32:00.914 that neurons are probably not dedicated to one thing only especially in the 00:32:00.914 --> 00:32:05.174 driven accordance you know maybe in contrast to early sensory systems and there 00:32:05.174 --> 00:32:11.574 it's very likely that you know there's a lot of multi a lot of mixed selectivity 00:32:11.574 --> 00:32:14.534 and so that you know neuron Neuron groups, 00:32:14.874 --> 00:32:21.334 if you like, coding things according to a combinatorial code, 00:32:21.494 --> 00:32:31.954 that will be maybe a very good way to be able to combine information, right? 00:32:32.174 --> 00:32:36.134 Another thing that's special about peripheral cortex is that it gets inputs 00:32:36.134 --> 00:32:38.674 for many, many different areas, right? 00:32:38.674 --> 00:32:45.634 And that also speaks to the fact that neural inputs also are kind of coming 00:32:45.634 --> 00:32:49.034 from a lot of convergent-divergent pathways. Right. 00:32:49.814 --> 00:32:54.314 So then on top of that, where do you see these models go? 00:32:54.534 --> 00:32:58.974 What's the next challenge for you in this model? Okay. 00:33:05.723 --> 00:33:09.943 So I guess briefly I can see several directions. 00:33:11.323 --> 00:33:19.783 Number one is to see if, you know, the kind of models we build can be kind of, 00:33:20.363 --> 00:33:22.703 and those are building blocks of cognition. 00:33:22.903 --> 00:33:27.163 I like to call this, right, local circuit mechanisms and building blocks for cognition. 00:33:27.263 --> 00:33:32.223 So the question is, can you extend this kind of approach, at least, 00:33:32.363 --> 00:33:36.163 to more complex behaviors, more complex functions, 00:33:36.643 --> 00:33:45.383 such as task switching or rule-based behavior, right? 00:33:46.103 --> 00:33:50.383 So that's one. Number two, I just mentioned earlier, you know, 00:33:50.463 --> 00:33:57.663 can we develop theory and computational mechanisms and the principles for large-scale 00:33:57.663 --> 00:34:01.003 brain systems rather than just local circuits? 00:34:01.003 --> 00:34:03.423 And third is real-life situation. 00:34:03.743 --> 00:34:10.443 That definitely will challenge our kind of models in a very big way. 00:34:10.623 --> 00:34:18.743 So we want to see if the insights from those kind of models are really useful 00:34:18.743 --> 00:34:25.463 for our behavior in the face of a natural kind of environment. 00:34:25.463 --> 00:34:29.243 Environment yeah so then what what 00:34:29.243 --> 00:34:32.143 i also realized is that recently you have been become more interested 00:34:32.143 --> 00:34:36.083 in let's say brain rhythms and brain oscillations and 00:34:36.083 --> 00:34:41.963 you wrote a rather impressive review article on on these oscillations reason 00:34:41.963 --> 00:34:45.983 why why how is this related to the attraction network so this is really a new 00:34:45.983 --> 00:34:52.483 chapter so this is a very interesting topic i think more and more people are 00:34:52.483 --> 00:34:54.563 interested in this really, 00:34:55.463 --> 00:34:56.883 very active field, 00:34:58.123 --> 00:34:59.323 you know. 00:35:01.141 --> 00:35:06.681 People from very different disciplines, you know, like systems and neuroscientists, 00:35:06.881 --> 00:35:13.761 as well as actually many people in the clinical field, people feel like this 00:35:13.761 --> 00:35:19.001 synchrony, neuron synchrony, might be a way to look at, for example, 00:35:19.021 --> 00:35:21.681 interactions between brain systems, right? 00:35:21.801 --> 00:35:25.701 So, you know, related to large-scale dynamics type of issues. 00:35:26.761 --> 00:35:34.641 We have worked a lot on, you know, or the mechanisms of synchrony and possible functions. 00:35:36.721 --> 00:35:42.021 You know, this emerges naturally in recurrent networks, basically. Okay. 00:35:43.081 --> 00:35:49.221 But it's still rather controversial, especially in terms of functional implications. 00:35:51.261 --> 00:36:01.021 In part, I think it's because you do see often evidence of synchrony and oscillations 00:36:01.021 --> 00:36:04.761 with measurements like EG or local field potential. 00:36:06.061 --> 00:36:08.561 On the other hand, single neurons 00:36:08.561 --> 00:36:13.441 are very stochastic. Even single neuron populations are very variable. 00:36:13.901 --> 00:36:18.201 Okay. And so those two things don't seem to, you know, fit together. 00:36:18.541 --> 00:36:22.721 Okay. That's, I think, part of the reason people feel like, you know, 00:36:22.721 --> 00:36:28.401 it's hard to understand how do you explain those two things in the same framework. 00:36:28.401 --> 00:36:32.961 And also, if the neuron operation is very stochastic, 00:36:33.221 --> 00:36:43.401 then how much, right, in terms of degree, synchrony is really there on top of the stochasticity? 00:36:43.601 --> 00:36:48.201 And if it's not big, right, if you have some measurement about synchrony, 00:36:48.201 --> 00:36:55.881 and if it's still a few percent of the signal you see, why it should be the main focus? 00:36:57.252 --> 00:37:01.512 Okay, I'm raising these questions without knowing the exact answers to those questions. 00:37:01.852 --> 00:37:07.732 So, you know, in this review, for example, I try to discuss how you can reconcile these views. 00:37:08.112 --> 00:37:13.352 Okay, and the general take, I guess we don't definitely know is the final answer. 00:37:13.472 --> 00:37:19.652 But the general take is that maybe it's better to think about the role of synchrony 00:37:19.652 --> 00:37:27.232 in a framework, in terms of neuron correlations, which we know are very important, right? 00:37:27.432 --> 00:37:32.692 Neuron correlations in time, for example, are important for plasticity, 00:37:32.752 --> 00:37:34.952 spike time-independent plasticity, right? 00:37:35.052 --> 00:37:39.272 It may be even important to generate a stochasticity. 00:37:39.752 --> 00:37:43.952 You know, it sounds like a bit of paradox, but we know that because a given 00:37:43.952 --> 00:37:48.832 neuron receives a lot of inputs, you can average out noise. 00:37:48.832 --> 00:37:55.272 And if, on the other hand, neurons are correlated weakly, they cannot average out noise. 00:37:55.572 --> 00:38:02.332 And so maybe correlation itself is important to generate a stochasticity. 00:38:02.452 --> 00:38:06.212 And that in turn, of course, can be functionally very important. 00:38:06.532 --> 00:38:13.012 So to understand the different aspects of dynamics in the recurrent network 00:38:13.012 --> 00:38:18.312 is a big challenge. So I guess that's a big reason that we are interested in signaling. 00:38:18.892 --> 00:38:22.652 Very good. So then to finish up, two questions. 00:38:22.952 --> 00:38:29.292 So coming from physics, going to neuroscience, attracting networks, 00:38:29.472 --> 00:38:34.912 and actually now moving towards brain oscillations, large-scale understanding of the brain. 00:38:36.432 --> 00:38:40.792 If you would have to stipulate a law that we should all follow in studying the 00:38:40.792 --> 00:38:43.572 brain, so what would be the Shao-Yin-Guang law? 00:38:46.392 --> 00:38:47.352 That's a tough one. 00:38:53.354 --> 00:38:55.334 Well, let me phrase it this way. 00:38:58.354 --> 00:39:05.514 I do think that trying to understand neural circuits of cognition is really 00:39:05.514 --> 00:39:08.354 a very exciting, challenging thing. 00:39:10.334 --> 00:39:16.634 And that would help to unify cognitive sciences and quote-unquote hardcore neurobiology, 00:39:16.774 --> 00:39:18.534 which right now is still kind of separated. 00:39:19.094 --> 00:39:25.854 And from what we learned, at least, is the key is slow reverberation balanced by inhibition. 00:39:26.714 --> 00:39:34.834 And so again, quantitative differences can give you surprisingly new qualitatively 00:39:34.834 --> 00:39:36.634 different functions and computations. 00:39:37.554 --> 00:39:43.094 And that'll be, I mean, it's probably known from the theory of dynamical systems field, 00:39:43.894 --> 00:39:53.594 but it should be a very big part of our efforts to understand cognitive circuits. 00:39:54.134 --> 00:39:58.554 Okay, so judging one law is slow reverberation does the trick. 00:39:58.754 --> 00:40:01.314 That's right. Very good. What time do you mean? 00:40:02.154 --> 00:40:07.994 Okay. And then my last question is, if so, I'm going to go visit you five years from now at Yale. 00:40:09.314 --> 00:40:13.374 I'm going to ask you then five years from now, like, look, five years back, 00:40:13.414 --> 00:40:16.574 you gave me this one prediction that you would have believed in, so how did it pan out? 00:40:16.634 --> 00:40:21.014 Was it false or true? what is one prediction you'll be willing to stick your neck out for today? 00:40:28.174 --> 00:40:34.654 Well, I guess still slow recuperation. There are many other smaller predictions 00:40:34.654 --> 00:40:37.014 that came out of the model. 00:40:37.914 --> 00:40:41.414 You know, just one more example. 00:40:41.674 --> 00:40:46.554 This is a very specific example, but I think it's quite impressive. 00:40:46.554 --> 00:40:47.794 If it turns out to be correct. 00:40:48.614 --> 00:40:53.694 That is, you find some scale invariance of reaction times. 00:40:54.054 --> 00:40:57.254 That's at the behavioral level. It's a psychological law, really. 00:40:58.754 --> 00:41:04.614 Very quantitative, very beautiful law in psychology that you can explain with 00:41:04.614 --> 00:41:10.454 a neuron-circuit model mechanism in terms of stochastic neuron dynamics, right? 00:41:10.754 --> 00:41:15.654 And, you know, that can be proved using neurophysiology. 00:41:16.554 --> 00:41:22.414 I thought that would be really quite important, really, to relate what you see 00:41:22.414 --> 00:41:27.254 in neurons in the recurrent circuit and what you see at the behavioral level. 00:41:27.774 --> 00:41:31.214 Yeah, but that prediction has already come out, so I want a new one I can come 00:41:31.214 --> 00:41:32.794 and hassle you with five years from now. 00:41:36.034 --> 00:41:42.274 That has not been done yet. Of course. The thing is, I want you to now take 00:41:42.274 --> 00:41:43.474 this risk that you might be wrong. 00:41:46.554 --> 00:41:51.534 So. 00:41:58.846 --> 00:42:02.446 I guess, you know, slow reverberation. But I guess how to prove that's wrong 00:42:02.446 --> 00:42:05.986 is... It's a safe prediction. 00:42:08.786 --> 00:42:15.306 Right. You know, we also want to have a big one, right? Or just not just any... 00:42:15.306 --> 00:42:16.526 Whatever you feel comfortable with. 00:42:16.866 --> 00:42:19.366 How about the following? Why don't you put some boundaries on slow? 00:42:20.606 --> 00:42:25.926 So in what range of frequencies are we talking about for slow reverberation? 00:42:28.926 --> 00:42:34.826 Yeah, so that, I mean, actually opened up a whole can of worms or opened a big 00:42:34.826 --> 00:42:40.626 door because, you know, when you make decisions, you can integrate information over many timescales. 00:42:41.406 --> 00:42:46.066 So, right, so in the brain, supposedly there are machineries that allow you 00:42:46.066 --> 00:42:48.946 to integrate information, right? 00:42:49.006 --> 00:42:52.606 So you can talk about integration over many seconds or even minutes. 00:42:53.026 --> 00:42:57.746 So we don't know on that timescale what really is going on in the brain. 00:43:00.086 --> 00:43:03.826 Let's see, what prediction that can be proven to be raw? 00:43:06.966 --> 00:43:09.966 Let's make that very specific. I guess 00:43:09.966 --> 00:43:19.186 I bet NMDA receptors and recurrent synapses inside cognitive-type neural cortical 00:43:19.186 --> 00:43:28.446 circuits are the key for slow integration on a timescale of a second or so. Perfect. 00:43:28.846 --> 00:43:30.666 Chai-Hsing Wang, thank you very much for this conversation. 00:43:31.366 --> 00:43:32.926 Thank you. It's a pleasure.