Larry Abbott Transcript
SPEAKER 1Alright, everyone, it's really an enormous pleasure. Just introducing Larry Abbot a second in a series of talks as one of our Bauer visitors So Eve introduced Larry yesterday. She told all the of the anecdotes, that I would take So I'm not gonna e do just want to say anything, I do just want to say that it's been real privilege, you know, for me to know Larry all these years. And Larry was sort of my unofficial postdoctoral, so I learned a huge amount from him. He's one of those people where when you talk to him, it makes you feel 22 IQ points smarter. Unfortunately, that effect doesn't last, but it's still always incredible but So I think Larry has amazing energy thought this incredible ability to communicate and, in terms of theorists, community deep insights into biology and can actually communicate the way he thinks about it. And the computational way theoretical way to biologists. It's a wonderful to have him and we're looking forward to his talk
SPEAKER 0
you hear me anyway? All right. It was okay. I'll try t o speak loud Let me just say a little bit about BRANDEIS and my experiences with BRANDEIS. So there probably are thousands and thousands of people who have benefited from BRANDEIS and hopefully all of you in this'll room. But you might ask, Who is the single person who benefited the most from BRANDEIS? And that is me. I could talk to you the whole hour about why that's true and convince you, but I'll just give you one example as far as education. Not so I got a PhD in physics here, but then I also got an undergraduate, that graduate and a postdoc education in neuroscience, mostly from me, but also from other members. So I got two complete educations. I also got two complete jobs, and I could go on and on. But anyway, if anybody asked you who's at the top of the heap? As far, a benefitting from Brandeis, it's me. Okay, so this guy here is is an electric fish. It's called the Elephant Nose Fish, even though it's a lift. Not nose, but it's not a bad name. Andi, I just want to start off with a calculation that all of you could do in the room. Uh, that it's gonna be related, so I put a bunch of shapes you can see on the slide here. They're all pretty much the same, but if you have a good eye, you'll see that? Not all, exactly the same. So if I gave you the problem off finding out which ones are different and how are they different than the others? What would you do? I think a sensible thing to do would be to take the average of all those to try to get out the common factor in them and so average of where all those things get a shape like this on. Then subtract that average out from the patterns I gave. So here in patterns, here's the average that you computed. You subtract them out and right away, out pops the ones that are different and you can see the signals by which they did so Simple calculation. If you follow that calculation, you are qualified to be an electric fish. So I'm gonna tell you how electric fish essentially that now in the electric fish case, uh, there's a slight difference in the way it goes. Thes patterns come in sequentially in time, but it's still the same problems. Look at those three and see if you can tell which one is different again. A good strategy that requires a little bit more sophistication this time is that every time I give you one of those pulses you have to generate the average yourself from pre computing is on. Then do the subtraction and you'll find again the difference So that's that's the calculation off the day, and it involves a bunch of calculations that I think are general interest in neuroscience. So if you have to compute an average, you have to sample, you have to accumulate evidence. That's that's a big topic in decision making. You have to correct any errors and your accumulated, Um, some, as in the case I showed you, that average has got to be pretty accurate because the deviations in my patterns were pretty small. And that's gonna be true for the fish to And, then, because this is a dynamic process, you have to learn how to generate that pattern. You have to remember it over time. um and so As I say, a lot of of sort of interesting computations are involved. So So let's get to the fish, by the way, if you want to interrupt me for questions, I'm fine with that. I pay attention to you, but don't do that. Okay, so here's the fish. Uh, that I'm talking about elephant nose fish on, but it has a feature that it shares with many fish that aren't called electric fish. and That is that the surface of its skin is covered with very sensitive electro receptors. This is true of catfish of sharp, so this is not what makes electric fish, but it's actually the system we're going to study. so These sensors are incredibly sensitive. And With them, the fish can pick up the electric fields produced by the motions off above in the water, Let's say, detect that bug going. So that's the system. And, as I say, catfish sharks use the same kind of system. Any time of water is money or a shark wants. Find something in the mud or in the sand. That's the system. they use. Okay, now, what makes this on electric fish is it also produces its own electric field. It does that with a modified muscle in the tail of the animal. On it produces a very short, millisecond long You see it at the bottom there of about 10 volts in the water that is sort of like an echo location or a radar system. With that system, the fish generates a field around itself, and it detects any deviations of that field due to, for example, the prey or objects in the water it to text them by kind of looking at the distortion in the field. So that is a cool sensation. It's called active electro sensation. It's also used for communication between animals. But it is not the topic of my talk, so you know that. Forget that one, except in that it's an insane thing to do with regard to the other systems. So I mentioned that these are super sensitive life receptor, that they could detect tiny fields and all of a sudden deficient, blasting the water with a field that's 100 or 1000 times bigger than these poor electoral receptors are supposed to sense on. What that does is it sends them into a ringing, uh, pathological condition that last for about a couple of 100 milliseconds. Looks like that. So when you see e o. D on my slides, that sounds new houses. Yes, I usually don't use this because, um, when you see you EOD. on my slide, I can't stand for electric organ discharge. That's the firing off this pulse that drives these receptors absolutely nuts. And if you notice this last for almost 200 milliseconds and these pulses of reaching, generated anywhere from a few hurts to tens of hertz. So essentially, the fish has to pick up this tiny signal riding on top of this huge way. I hope you see the analogy with the problem that gave you at the beginning of the talk. It's actually a slow, low frequency wave imposed on this. Just like what I showed you the beginning. The fish has to do that calculation. Okay, So historically, this has been a curse, spells area of study and throughout his career just did a gorgeous job working in this preparation, along with a theorist. Patrick Roberts has also been a gorgeous job I'm referred to never working on. Then I'm not really gonna talk about this. The work I'm gonna talk about is done by this trio. Um, Solomon is a graduate student who works with Nate. So tell me, and Nates Nate, and Abby I assume some of you recognize you she came from Brandeis on, but this is actually an old picture over from BRANDEIS. I right. Anyway, she was from Brandeis Okay uh so, here we go. The computation off of this electric fields. Maybe, maybe if I If I go back for one second, just like I had to do in the beginning to talk with the electric fish is going to solve. This problem is gonna subtract out the field it generates, um, that it's not interested in and leaving over the small signal due to the bottom. That's the general idea. And that occurs in this electro sensory load, which is kind of a modified piece of cerebellum that is called the E L L for short Now there's a command nuclear nucleus. Um, this sends out a motor command. Remember, this electric organ is really a modified muscle, so it's just like a motor command to a muscle on and when that gets to the tail pile. Electric fish electric field goes off on and, then command nucleus. The whole thing that makes this possible also sends a signal to the EL L to say, Hey, I just triggered the electric Gordon, you should expect a big shock in the water to come. The ELL also receives the sensory input from those electric sensors on the skin, and then it sends the resulting signal up to the rest of the brains. Is that go eat a bug or whatever the animal wants? Todo, um, players in this game are shown here these green guys on the output cells. So there, the neurons that are sending this signal off the rest of the brain on the purple guys are called mg self medium ganglion cells. And they're kind of the heroes of the day. They gonna show up. I'm going to do the calculation for the system. Okay, So I'm gonna take you through this system. And, I'm going to live it. I'm not gonna show you the full thing at the beginning. Just go with me, because I get a little more complicated as we go on. But I'm gonna give you a simplified version on. Then I will fix the various things that I've left out by the end of the talk. I promise. The Other thing is, if you've heard Nate or Curtis Bell or whatever talk about this system, I would say Just forget everything they told me. I'll tell you why. Okay? And including me. If you've ever heard me talk about. Also forget what I said. Alright, so let's trace through the yellow. Let's in there And what does it do? So you have a command nucleus that sends off pulse that says fire the electric organ that produces electric organ discharge that produces this pulse in the water that is picked up by the electro receptors and causes. Is this really pattern that I showed you before? Um, if there's a bug in the water, it also picks up whatever the bug is producing so that some of those signals goes in through the Afrin fibers of the electro receptors. Um, there it is. It gets picked up by an Interneuron on related to these output neurons on. Of course, the Interneuron just does the same thing. It it reflects some of these signals. So if this was the whole circuit, what would happen is, uh, not the best thing, which is the output cell will just send the total signal to the rest of the brain without filtering out this uninformative signal, which doesn't doesn't provide any information to this system, which is just the electric organ discharge. So the point of this, this, uh, circuit is to get rid off basically this pulse here and send off this pulse here. How does it do? I'm gonna give you the answer before I explain what's going on. But the answer is the second class of cells, which are the majority of the cells in this structure. Um and they receive those. Sorry, let me do this. First they inhibit the output cells. So that little triangle is supposed to be an inhibitory Synapse hit the output cells and they get this'll, um, copy of the motor command through a set of granule cells. So the granule says that's how they know that the electric organ is being discharged on. What they do is they produce this average pulse. That's where the average pulse came from. I have to explain how to compute it. But today, right introduced the, uh, the average pulse Whenever they know that the electric organ has been discharged. The key is that that black wiggle is the average off off the red wiggle. So it contains the repeated self generated field. But it doesn't, of course, contained bug Because the bug can't be average. That is different every time. As, a result of that inhibitory synapse. The subtraction occurs on the bug. Get sent off to the rest of the brain That's basically our circuit works um and Now we're gonna get into the nitty gritty of how works. Um, so obviously that what we're gonna focus on, at least for most of the talk, is this part of circuit. This is where the averaging is being calculated, where the average signals produced on. Then you know, there's just a subtraction that occurs on the output stage. So let's focus on this part of the circuit now. The other thing is that we're gonna get rid of the bug because this circuits supposed to calculate the average. So the bugs just gonna average to zero anyway, So I'm not gonna talk about the bug to the very end of the talk Um, and then I just simplify the diagram. But then, if you look at this, you're gonna say there's something missing. This thing can't possibly work because you haven't sent the signal. This is the signal is supposed to be average. It's not going here. And so that's because I told you I was gonna lie to you and one of the lies that so far is that I left out A connection on this connection is the key to the whole thing, which is, uh, again, an inhibitory synapse. There it is that sends this signal into the NG cell and allows you to do the averaging, etcetera, calcula calculation. The weird thing about this input is it really is a kind of a learning input. Um, it doesn't drive the cell. You wouldn't want To drive the cell through the sensory pathway because otherwise it would reproduce the bug signal and then subtracted out of the output. That's not what you want. So this is not an ordinary input. And I'll explain what that means as we go along. What? What is our learning? What? Me? Okay. Now, in an older paper with a bunch of graduate students at Colombia, uh, Nate had recorded from these granule cells and the granule cells turned the basically the delta functions. So, you know, they get a command that says, bam, I just triggered the electric organ, and they turned it into a kind of a complex temporal pattern. So just just a sampling off the cells across time. This is a couple of 100 milliseconds. You can see. Here's where the command signal came in. There's a lot of early stuff. There's some late stuff, you know. Anyway, they make a basis like this and then Um, there's plasticity at these synapses on by a process that I will I will explain to you, I'll tell you about They extract the average. So that's where the averages computed. It's built from this set of basis functions. The average comes out and then the average just gets transmitted to the output and you're in business. That average sent it to him. The synapse Bingo. One thing I should mention. I mentioned that this was a cerebellum like structure. So if you know the cerebellum, you recognize granule cells, parallel fibers, these are the cells off of the system. They're not called cells, but that's what they are on. Then The output cells, which I don't have on here, are like the deep cerebellum nucleus cells. It's just that they're up with with the such up with. the cells, it is located in a different place in than this Okay, so how does all this happen? How does this learning take place, um and, How does the system part eso one thing that if you know per Kinji cells, you will know that frequency cells have have what are called central spikes and complex spikes. These cells are exactly the same. They just have a different name for them. They call them narrow spikes are the ordinary spikes on They are the action potentials that carry the the signal down. This acts on and send it to the output cell. Get there. But in addition, there's this second kind of spying complex spikes in cells which is called broad spikes here have no idea why they're not called by The same name And the key with the broad spice was Here's a picture of the broad spikes you can see here that they're broad. And, you can see how that they're also tall s so they might have been called tall spikes on. But they're only tall because the normal narrow spikes here are very attenuated in the SoMa. This is somatic recording, and actually, that's very important. People keep that fact in mind. It's important that they're attenuated in soma, but they're actually full fledged act potential. Thanks. Okay. So that these two kinds of spikes the narrow spikes occur much more rapidly. Maybe 50 Hertz, whereas the broad spikes are one or a few hurts so many more narrow spikes. The broad spikes on the key to the story is it's the broad spikes of driving plasticity. Just as in putting yourself on day a number of years ago in the bell lab, they worked out how this plasticity work. So they worked out the rules off this plasticity mechanism. And it works like this. If you have a pre synaptic action potential, that means one of the brand new cells fires or spike, and it arrives at the synapse here on and then Uh, nothing happens so you don't have a broad spike It doesn't matter what the narrow spikes do. You don't get a broad spike. If that happens, the synapse states stronger. Okay, so Spike comes in here, gets this synapse the synapse will get strong. Now, if you think about the rate of that process, I guess I do this first. Never mind. Sorry about Okay. Now what's the other? If you have a spike in the granule cells comes in and hits the the M G, the dead right, and then you do get a broad spike within a tens of milliseconds. Then instead of synapse kids, it's It's a spike time independent rule, kind of a spectrum independent rule thats it Okay, so now I get to the rate So if you think about this first process, the I'm again pushing her on if you think about this first process. So it's only Spikes coming in the M G cells that doing anything that's gonna occur at a rate which is, whatever the rate of the grandiose l spike this. On the other hand, if you think about the reverse, the other process the weakening process. Since that requires both a grand yourself, Spike and abroad, Spike, it's gonna go at a rate proportional to the product of those just proportional to that. And now, if you want to get an equilibrium, so you want the total amount of strengthening of snaps to be matched by weakening. So on average synapses is holding constant. Then you clearly have a relationship like this, right? The rate of the granule cells spikes has to be in proportion to the product of the rate of granule cell spikes and broad spikes the constant of proportionality depends on the amount of synoptic strengthening and weakening, and I'm not gonna bother with that. But the upshot of this is that equilibrium. Once this learning rule has done its thing, you expect the broad spikes to be held at a constant rate on that constant rate is a few hertz So this snaps plasticity. Well, just remember, it's gonna do anything it has to do to keep the broad spike. Great, constant. Okay, so where are so let me show you how the learn works, then. As far as the broad spikes go. This is, as I say, there's been known for a long time. This is not our new work, but it's kind of okay, So imagine. Let's say that the synapses are set to the wrong value and what's being coming up? What I'm showing here is sort of the sum of What? What's coming out from all those synapses? I suppose it's just a constant. So it's, you know, it's not gonna produce any kind of useful signal. Um, if that's true, then what's gonna happen is because we have this in a vision here, you're gonna get basically the reverse of this signal coming out in the broad spikes That's exactly what happens. So this inhibition is causing a big dip in that ring. Exactly. Inverted from here in the broad spike. Right um This is doing nothing. It's just a constant current throughout this pulse. All right, so what's gonna happen here? Well, if you think late in the in this pulse, it's gone back to normal here. So at this point, uh, you imagine that the broad spike rate Is that the rate it wants? So this is the equilibrium, Value And that means that any of these, uh, activities and granule cells that occur late like this, for example, these guys here very late. So time is going this way. Um, they're not gonna produce any change in the synapse so that the granule cells that are contributing to this part of the signal are already in equilibrium. Nothing happens But if we move backwards in time, you can see there's a bump where the rate is too high of a broad spike. So that black thing is broad spike. Great. It's now above the equilibrium point. And that means that the plasticity will get engaged and for the the granule cells active in here that contribute to this part. What you're gonna get is a weakening of the synapses. Okay, Now, on the other hand, if you go with big pulse there you get a period when broad spikes are lower, then they should. Well, that means you're gonna get a strengthening because the system is gonna try to bring the broad spite rate back up. So you're gonna get a strengthening off. For example, now, these early guys So the early, um, granule cell inputs that contribute to this part of things are gonna be strength. So what's gonna happen? Well, that's gonna distort this thing. This curve, right there's gonna be no change. In the end, the weakening synapses gonna drop the curve, strengthening is gonna raise it up. And of course, that's exactly the inverse off the output signal, and so it's gonna shrink it. And so over time, what you're gonna get is a transferring us so you'll extract from this basis of granule cell things an exact copy this signal and put the the broad spike back to equillibrium So that's the learning process in these cells. It's kind of nifty it's completely self contained um and It basically does the averaging because it's fairly slow. So I mean, the way I said it is, it makes a copy of this, but really, it's slowly develops a copy of this. And if this has both signals on it, it doesn't copy those because they average out To zero and so This is exactly building the copy we want, and it built it from the reference copy of the Motor Command. And every time the motor command comes out, it extracts this signal Alright, so that seems like great success. This process is, uh, a cancelation. It's always been called cancelation. I think I put the word out beyond, uh, what we always thought about. I think what the first Bell Group who did what we always thought about it, um is that it's a cancelation process by which the inhibition here is canceled out by the excitation here, resulting in a constant products. Okay, on that word, it's it's an interesting thing in in research. Sometimes you just think about something the wrong way. It set you back for 10 years or whatever this is an example of that. It set us back a long time this cancelation idea on Why did it set us back so much? This is kind of what I'm telling you. Forget what we used to say, because if you think about what I told you in the beginning of the talk, if I go back now to the narrow spikes, narrow spikes on the output of this system. And I told you that the output of this system was a copy off was the average right. In other words, the output of the system and narrow spikes was exactly this command on dso What? I was acting as if, Well, this current goes blasting through here and how it goes. But what about cancelation canceled? There's nothing there. So as I said this this set us back a long time on by explaining finally, how this goes. But there's a There's a paradox, which is exactly what I said in the beginning. It's is if that input the acts on broad spikes and doesn't act on narrow spikes, right that when you're talking about narrow spikes, if it's not even there, there's no cancelation. You just get that grant of self. No. So how on earth does this how so In order to do this, we we built a multi compartment model, and that model was actually a sort of recycled from an old model that Patrick Roberts built. I should say Patrick Roberts did, Ah, huge body of theoretical work on this system in old days on. But it's an incredible body of work. He did absolutely everything that could be done. The piece that he was missing, the two pieces are the pieces that nate has provided. One is this granule cell basis. Hey, guessed it and he didn't get it quite right. On the other is that people had not recorded from the M G cells. So So the whole MG self story was kind of half there and half not there But he did an amazing body of work, and we benefited because we could just go grab one of these models on and um Try it out. Try to see this paradox we had canceling signals that, for some reason didn't seem to be canceling when we talked about narrow spikes. So here's the model. It's a pretty good model that Solomon sort of adapted from From that Robert, um, here, you can see, uh, that data the looks of the narrow and broad spikes and model produces that quite well from a somatic recording and But this is the structure of the model. Actually, Solomon rebuilt it. You following the structure of one of the cells that make them filled, but used the conduct Ince's and all that from from model Okay. And and the thing that it works like this, the broad spikes are generated up in the April dendrites You might think they're calcium pulses, but they're not actually in this system. There's some kind of modified sodium spikes that are super broad sodium spikes. Okay, so they're generated up there. Um, the narrow spikes are generated on the axon the little way out, which is why they appear quite, um, right diminished in this soma. The inhibition that I'm talking about that has this sort of paradoxical effect occurs sort of the basil part off this dendrite on that remembers the sensory signal that comes in and suppresses the broad spikes until they learn through the granule cell input that's on here. They learn To, cancel that out and make constant prospects. So that's the general layout of the model and Basically, there's one of these cases usually, you know, for me dealing with multi compartment models, this kind a scary situation, because if they don't work, they have four million parameters that you have to tune. But in fact, this just worked immediately and the problem with how to work rather than that works. So almost immediately, when it got set up thes eare the results that Solomon found so first of all, um, this is the model receiving a pulse of inhibition resemble, you know, modeling this sensory input through this inhibitory channel on Here's a broad spike. Great. It's it's four hertz and equilibrium, and you can see that that caused this tick So it does exactly what you have. A mission lowers the bread spike. Great on. You know, much to our pleasure. If you look at the narrow spikes generated out on the axon, they don't even notice this. There is nothing happens at all. You can't even tell anything happen. So it is true that the narrow spikes don't don't notice that inhibition. Now, what happens if you do the cancelation? So what we now have to do is introduce, um, a input up here the granule cell input that cancels this pulse. That's what learning does. So you can see here. Solomon has done that. So now there's a grand yourself. Input thes two inputs. Cancel. There's no effect here on the broad spikes, so that would be the equilibrium of learning everything back to the equilibrium 4 hertz And sure enough, there's this positive pulse coming out in the narrow spikes. So this this input got copied to the narrow spikes, which is the whole point, right? You wanna average that input, copy it and then generated every time you get a command and send it out narrow spikes in order to do about it cells. So the thing worked perfectly, and then it took us, you know, weeks and weeks to figure out how it was working on. But here's how it's working, at least in the model. And I think, uh, in the real cells, we don't know because we haven't done this sort of rice electrophysiology at least yet. But I think largely this is how it works as well in the model with okay on that is that if you look carefully here, especially here you'll notice that in the data, almost always, I think it always the broad spike is generated by a narrow spike. You know if there's a little narrow spike, and that's true in the model too. So there's a little tickle from a narrow spike generates abroad. So in the way that happens is that the broad spikes generated the axon back propagates, gets highly attenuated and gives a little tickle from the broad spike generating region sets it off now. That obviously doesn't happen very often because I mentioned there's something like a factor of 50 ratio between the rates of these narrow, spikey broad spikes. But what you get on idea is that there must be some relationship like this, right, that the broad spikes are being generated by the narrow spikes, but with very low probability. It has to be certain fluctuations. Maybe the broad spike a little taller than normal. Maybe it comes with other sequence, but you know, we formulate that as as as this factor P, which is something like 1/50 that narrow spikes make broad spikes but very ineffectively. And that's why the attenuation was important because it's really the attenuation if you look at the model that narrow spikes just sort of barely make it up this pathway, just barely get there on occasionally caused a broad spike. Okay, so here's what the process looks like that so baseline means we haven't introduced in the model. Any sensory input? We haven't done any learning. We're just looking at how the model behaves all by itself. Its adjusted to have a roughly 50 hertz. Whoops. That should say narrow Spike. Right. I'm sure that, um it's like on do you know, if you look, this is just a schematic at the bottom. But the idea is, if you look at the action potentials that we're triggering the broad spikes, they sit on some baseline and they have some height on, but have a low probability off generating a broad spike Okay. Now, what happens if you add the inhibition? So now we're doing a inhibitory pulse. I already showed you doesn't really do anything to the narrow spike. Great again. Don't pay attention to the wrong label. Um, and, uh, it does. It does affect the broad spike. Great. However, and it does that by two effects. One is as you would expect of inhibition in the region of the Aprilical dendrite. It pulls the voltage down to sell. And the other thing. If you look carefully, it also reduces the height of the action. It's a weaker, region, if attenuated on the result of that as You saw that broad spike break goes way down, but it's going down because the P went now we always thought that cancelation means everything just goes back to normal. But in fact, that's not true in the model. What goes back to normal, for example, is the peak of the action potentials in the region where you make the broad spikes come back. Right was before on average, but the baseline doesn't baselines actually higher. That's because now we have granule cell excitation on. That means the actual potential. Still, shrink here on the result of that is that the probability does not recover. The probability does not go back because a short action does just worse than generating broad spiike And that means to get the same number of broad spikes you have to make more spike. That's how it works. Um, I won't beat it to death if you didn't get it. Just trust me. It works. What? Okay, so we end up now with a prediction off how this model works. I'm gonna getting experimental data, which I've completely, uh, not given justice to until this point, I hope to do a little better So there's a prediction in this model, right? If what I told you is true, if the cancelation of this signal is being produced by M g cells that if you look at the m g cell output, you'd better match this signal. You better be able to figure out that the signal that's inhibiting on this pathway is the right signal. in other words, that's that calls for an experiment. That's a conjecture of model. So let me tell you a little bit about how these experiments are done because a lot of cleverness and allows you to measure all these things. So I'm going back to my first slide. Remember? That was the whole set up with the bug and everything again, we're gonna ignore the bug for the time being. Um, but in these experiments, the fish is, um, you know, not in inestitized but sort of has a pain relieving, but it's also paralyzed. But if the fish is paralyzed, can't wiggle around and all that. Andi, remember I told you that this electric organ just charge it was due to a muscle. That muscle is now paralyzed. So when you in the preparation in the lab or the fish cannot make electrical discharge not happens, on the other hand, it still makes commands. It commands like crazy on that allows experimentalists to do a clever thing, which is to trigger a new electric field in the water. Every time you pick up a command on electoral around a nerve every time to do that. Now the experimental produces fielding state. So you tricked the fish in the thinking. It's making a field that it's actually not making. Now, the cool part of that is now your you've got complete control of the system, so you could do, for example, this trick. You could generate the electric field without the fish, knowing about it without fish, that sort of unlocked to the command. When you do that, you're not allowing the MG system to work because it doesn't know an electric field has been produced, So that means nothing comes out of the M G uh, cells. And that means if you measure the output cells, you're gonna see this signal that I talked about beginning to talk. The UN canceled signal will come out, and that means you can measure the sensory response that this system is designed to cancel. It is very cool trick on dso you basically you're measuring that. Bye bye, Fouling of the system in this way. Okay, so now you can check this prediction because you can look at the MG output. That's the novel thing that they did. As part of this project was to actually record outputs from M G cells. You can figure out what this signal is using the trick. I just told you about it and you could see if they match and they match quite well. So is this. MG response is actually obtained by summing from several cells because, in fact, a number of M G cells, maybe 10 or 20 will converge on a single output cell. So it's not just a single one. The height here is not predicted because we don't know the strength of the inhibitory synapse, but at least you can see that the shape it seems to match quite well. So? So that prediction Good enough. Good enough for theorist Okay, so What's the next thing? The next thing is not I have to tell you more True truth from I haven't drawn the whole surface. Okay, Um the unlike in the cerebellum, I mentioned that these output cells are like the deep cerebral a nuclear ourselves. But they're up with Akinci cells and they actually get also get granule cell input They don't get us many thes guys will get maybe 20,000 synapses and these will get 5000 synapses they don't get asmuch. But they do get, um, synapse from the granule cells I left out to make it simpler. Andi, there is plasticity. We don't really know the mechanism or the details of that plasticity, but it does exist, um and We also know that that plasticity kind of does calculation again. So it extracts from the granule cells What's called a negative image? It subtracts a signal. Uh, sort of like this one. Remember, this gets inverted, so it does its own calculation Now, if you've heard the old talks, what people did in the old talks, including me, was talk about this system in isolation, but they acted like this was the whole story. But we used the plasticity of this system. Is if it was over here, so achieved a little bit. Um, What we're now saying with this, now that we've measured the MG cells, is that yes, indeed. There is plasticity on the output cells, but that as I've drawn a bulk of the plus the bulk of the negative image, the bulk of the cancelation is actually done by energy. So So I think of this as empty cells of the workhorse. They subtract that most of the signal and then individually output cells companion tweet things to their own purpose with their own plasticity. So how do we know this is true? Not more tricks. So if you generate a command, the fish actually generates the command and you don't fire the electric field. Then you get rid of the whole red pathway. And then what you could do is measure the effect off these two signals. Okay, so bye bye again, doing a little trick. You could measure the some of these two as what's called the negative jump. Now, what we also like to do is measure least one of them themselves. Because we want to know from the some who's giving the bigger contribution and the way you do that way or the way Nate does that is to prevent the plasticity in this output cell by voltage time. So you you refuse to let the you don't Voltage clamp that spike But you you put in negative current pulses. Basically, it's not your perfect voltage, but you hold the broad spikes at a constant rate. You hold thes air, general broad spikes. Sorry. You hold the fire making a constant rate on That means that this plasticity doesn't have so you can cancel out that apostasy and now when you do the other trick, you know off. Get canceling this. Now you're Onley measuring the mg contribution. So the key is you can pull out tease apart these two effects on this is the This is the result. So the green here is what's coming from the M G cells. This is the total negative image that does the cancelation. Andi, this little bit is coming from the output cells. So indeed it's true in this system that the bulk of the cancelation is being done by the energy cells. And there's some tweaking going on from the output cell If you know the cerebellum literature, you know the famous plasticity is in the cells. But there's also plasticity in the Deep cell growth and nucleai things just okay, what's next? What's next is truth, right? So I have added this connection to the output cell. I'll add one more connection and that is that the M G cells, for reasons we really don't understand, also get a kind of a more conventional input that does afftect the narrow spikes I I think that a Sfar, as I could tell them this was put in there by nations just confused daily this'll makes the whole story much more confusing. But what I can tell you, I promise you, is that ultimately that input. just gets canceled doesn't do anything. I don't know why it's there and I've just ignored it. To spare you from the confusion that we are through working this out. But there is that input without bigger lie. Is that only described half the system. Don't worry, I don't have to go through the whole talk again because the other half is just a flip side. So I'm sure you all know that many nervous systems come in sort of on off varieties, positive negative sides. This one's exactly like that. So in addition to this system that I've told you about and just To get you that nomenclature in this system, I've talked about an output cell that is an excited by the sensory stimuli. This is called Output E cell E doesn't mean the cells excited, very excited by stimulus. Andi, I've talked about M G cells that are inhibited by the stimulus, but on the other, there's a flip side to this, a system that completely works backwards exactly everything I told you the same. The only difference is in this case, the output cell is inhibited by the stimulus. So it's inverted of the MG cell is disinhibited there? There's an extra Interneuron um and so It generates the upside down signal to cancel the upside down signal blah, blah, blah. So anyway, it's just a flip side. But that means that in addition to being two kinds of output cells, that was known that there were two kinds of output cells and that they were either excited or inhibited by the by the system. There are also two kinds of M G cells on. Do we call them at Bs minus? That's because the broad spikes are in inhibited by the stimulus. And you have to say yes because the narrow spikes and broad spikes have this weird difference as a how this input affects them. So now we need another class of M G cells on those BS plus ones. Those air X broad spikes are excited effect stimulus. So all this is just, you know, it doesn't really matter to the story, but it matters to the model a lot because, um, we have these two kinds of M G cells and they better be wired the right way. So what you realize is you must have the bs minus cells inhibiting the output E cells because if the B s plus cells did they give a long sign of signals which completely screw things out, and it must be that way BS plus cells in output cells. And I must say, you know, everybody says those you know modelers should make predictions, but this is a prediction that really scared because there was no way out of this prediction We were gonna be dead, Solomon, and go back to doing another rotation. This'll one didn't come out right on. The trouble is in the fish you could test it s o I sweated for for a few weeks while Nate did test the reason that you could test this Is that the output e and output I cells that has been drawn that way? They're in different layers that they're separated in space in the structure. And so what Nate could do was to record from an mg cell figure out if it was a Bs minus or a BS plus and then fill it and see which layer it's axons went to um and so That was a prediction model. Andi he did that is in this for pictures that he drew from it. But the punch line is over here, and that is indeed true that the axons from the B S plus cells went into the layer. This is supposed to be a layer. This G is a layer where the inhibitory output cells and the axons Bs minus went appropriately to the right way. We would love to be able to do E m, for example in this thing and really verify that this is correct. But at least that the light level it's true that the axons seemed to go right place. Okay, so where am I? All right, so one more thing, um, this is I'm not This is an earlier work that we did, but it's kind of nice to end with, So I explain to you how this circuit works. Um, does it actually matter to the fish? You know, this is we always assume that we know what this circuits doing on that the fish is using the way we think it is. Um, and, uh, there's a chance here to actually show that. So let me show. So in order to do that, we have to introduce the prey on. That's easy to do, because remember, we're mimicking, are not we? They they are mimicking, mimicking an experiment. The electric field produced by the fish you can also mimic on electric field produced duplicate it So that's what's being done here. Now, the problem is this a behavioral experiment in a paralyzed fish. So how do you know when a paralyzed fish whether, uh, fish sees the prey, for example. And the cool thing is that when the fish sees the prey or see something interesting in the water senses, um, it accelerates its command right. Which makes sense. In other words, it's usually pulsing that commander friendly, low rate suddenly picks up on electric field that looks interesting on it starts pulsing really quick so you can actually tell whether the fish the fish tells you whether it senses, uh this prey or not. And so here are some results on the ability off the, um, fish to sense the prey. So again, that This signal is a change in the command right? That's the indication that the fish senses something interesting going on. Um, initially, no mimick means that there is a prey mimick, but there's no electric field. So this means the these commands are not doing anything. There's no electric field in the water. This is an easy case for the fish. There's nothing to subtract. It just consents. the prey all by itself is not the natural situation. This would never happen in nature, but you could do it in the lab. Okay, Now what they do is to turn on the mimick So now all of a sudden the fish is producing electric field. It hasn't been doing it for a while, so it's actually lost the plasticity This this takes about five minutes or 10 minutes to change. So if you keep a fish without its electric field for 10 minutes, it's completely forgot is a continual learning process that takes a sort of 10 minutes to roughly make a a match. to the average maybe an hour to get perfect. So anyway, not surprisingly, when you turn off the electric field suddenly the fish is being overwhelmed by this ringing noise, and it's electro sectors. It can't detect prey. But then, as it builds, the learning happens. So you can actually watch learning happen as it builds this negative image. The average that I talked about subtracts it out back, it goes, and now it can detect the prey in the presence of this field. It's a really impressive actually noticed. It comes back right to the level it was before, despite the fact that in this case, you know, nothing was in the way of the field. In this case, you know much, much longer. But the field is being played at the same time. It doesn't bother fish at all. And then you could do one of the tricks that I showed you, which is even though the fish has learned how to subtract. If you produce the electric field when it doesn't make a command, it can't subtract it because it doesn't realize that made the electric field didn't make the electric field and then it's lost again So this system really does what it was always said to do. It allows the system to see people. Yes, okay, I'm just gonna wind up then, by giving you some lessons from this thing, Um, that the thing that why I find this work interesting of this interesting natural system Because as far as I know, it's the first time that learning has been studied. Not at that output layer. If you want that the language of machine learning, in other words, people are certainly looked at the effect of learning. Let's say, on out on a cell, never mind whether there's not someone just using it as a stand and cell and then What they usually would look at is the effective plasticity of that synapse carrying in some signal, uh, due to some manipulation on watch it happen. So only in one layer. But, you know, we all know from machine learning that the power of learning occurs when you go through multiple layers. And um So and you know it's pretty obvious in the brain that we don't only learn by output. We learn all through the brain on. So as far as I know, this is the first case we're learning one step back. So we're talking about learning a bat synapse affecting the output of that thing. Uh, and with that comes all of the provisos those that are part of machine learning. So, you know, as I said, you know, deep networks are many layers, but at least we've started to think about learning in a t least a two layer system. It's a step in the right direction. Now, the tricky thing about learning when you're not right at the output is that the effect of this synapses direct. You know, if you want the this inhibitory synapse, you want this guy to fire more. You better weaken that synapse. It's a simple is that, but the effect off to back can often be very subtle. And that's the origin of back propagation in artificial networks. On there's been a whole lot of papers about how to make back propagation work in a biological system. This is not back propagation, but it's interesting that we Have a system that has kind of a distributed error. signal machine learning is actually build systems like that. In other words, the error signal at the output that does this plasticity and their signal here their identity Well, in both cases, you want to cancel the average And um So this is a system where you apply the same error everywhere, and as a result, you don't have to do back propagation. Um, the other thing is that typically in machine learning, um, learning and processing are seperate, uh, you know, in standard deep networks, you train them, you know, for a whole long period, and then you just freeze the system and then you settle, for example, or market or whatever. But it doesn't learn after that. And even in things like helm holds, machines or whatever, there are these learning and processing phases, and that's because there's often a conflict between what you want to do. The cell, what do you want the cell to do when it learns what you wanna do when it processes? And here we didn't have to do that with fish. That thing this thing learns continually. I mentioned that anytime you change the field for five minutes, it will re calculate the average on. By the way it does. It is to compartmentalize learning and processing in different channels. In other words, there's an a pickle channel that uses broad spikes for learning. There's a more basil channel that used a narrow spikes spikes for communication, and the cleverness of the biophysics that I was trying to tell you in the model is that that allows the process to go on together at the same time and not interfere with each other. That is something that is not yet achieved in deep learning systems. Finally, I think something to remember here that comes about from this connectivity. These mg cells, you know, went to properties. These went to the upward eyes. If you want to understand the wiring of this circuit, you have to think about learning. If you think about the signal processing that goes on the way that these guys respond in their narrow spikes. For example, I left those connections, but you cannot make sense of this connectivity, which is why nobody expected it. Nobody proposed in the path that this was a pattern of connectivity. You could only understand the wiring of the circuit if you think about learn. And that's absolutely true. If you looked at the architectures off deep networks, they make no sense. If you think about him in the forward, since they're all built so that the back from works. And so I think that's a general thing, no matter what system of just leave with this you studied, we tend to think. Okay, this is connected to this because then, you know, when I put on the stimulus system, the office will go down. That's great. But maybe sometimes the wire you could only be understand stood. If you think about how the system learns now, process this information. All right. I'll stop there and take anything. Think about plasticity, I think plus and minus. It's the same rules, right? You know, even though a signal Yes, it's exactly the same. Learning. Yes, that's what it is. I presume it has to be really fast. That is the window has to be fast. Otherwise, yeah. Yeah. Window. You probably members. Well, as those old patients 20 more seconds or I don't know, 30. It's more that I don't remember. Yeah, um, so it starts me about looking electrical discharge. You figured the moral. A stereotype. Animalistic produces them. It seems like you have, like, an extra extra set of knobs to turn on this to be a deal with any arbitrary. Very good. Yeah, I should have said that. You know, why is this thing so incredibly plastic? The reason is that at least seasonally and probably depending where moves the electric properties of water changes and um so that the fish has to adjust to that fish also grows. So during development, you know, electric body moves relative to things and um so For some reason, this thing just incredibly plastic. It's a very general learning system. I talked about this yesterday, so this is kind of a theme. I'll do a wild speculation. You know, both of the learning systems that I talked about yesterday today I had to do more than you would think they had. You can You can make upside down electric fields here and they'll cancel. You could make things that are too wide, Too narrow. Doesn't care. And so So that's my exclusion, you know. Why can't we played it down? We never evolved. There was no pressure to play the piano. No, your youth. and um so But, you know, we are general learning systems and at some point, evolution. Just just build a general learning system, and then we can do all these things that way. That's kind of a magic of humans, so