[00:00:05]
>> We're very pleased to have so welcome Dr Marc Michael Shannon Mikey's And I think either of the current heroes Medical Research Institute and professor of neuroscience and Commander Universe think you'd think he's be any from around and he's from u.c. Berkeley where if you work with her for a month you're also paying them and the from wrong meaning that I'm up and like most of us he's an actor not there if.

[00:00:29]
You don't like the pose Dr room craning quit playing you some a Stanford and finally join the department of your 1000000 by physics at the University of Washington and where he remains unclear on what I think 12 before joining them via my has made fundamental confirmations who are in the standing of this ng they think by combining behavior I look for physiology and compassion I'm wondering if you're Hispanics and the Korean is also an avid You know the terrorist So let me think the shorten were done for and do my think that way thanks a lot Rob Thank you Bill thank you all for inviting me I'm sorry about the technical problems that we can't see each other so I'm going to talk to you today about our new work.

[00:01:18]
That involves a collaboration with these people saying who goes by hurts and a serious Ashok Kumar and another experimentalist Richard Axel but this is a you know I want to just give you some background which is that in evolution the emergence of of cognitive capacities was accompanied by an expansion of the cortical mantel I'm assuming you can also use my house and.

[00:01:45]
And most of it's devoted to what we call association areas and many of the neurons in these areas especially the pride and prefrontal cortex exhibit persistent neural activity and that's often associated with working memory or knowledge states that are called noses gee you know we're practices like planning Ok so and decision making I think kind of holds the key to cognition and.

[00:02:11]
Offers us some traction on the elements you know what Plato called carving nature at its joints and so I think that the my basic approaches are as light as the principles that are lucid in the monkey by study decision making pretty complicated decision making but still simple compared to the things we're doing right now out.

[00:02:32]
There when they are less principles or are large they're likely to underlie sophisticated cognitive processes that go on in humans that ultimately give us some indication or some insight into what's responsible for disorders of cognition in humans now to treat those kinds of disorders and make a difference of brains I think we need to understand those principles that are much more refined level that we're capable of learning in the in the in the monkey and so I've been devoting a portion of my lab doubts with the mouse in order to sort of break things down and I mean the talk that it's been a Instead I'm going to talk about it a monkey inspired project then led to some work in the mouse for you some background about decision making and I'll.

[00:03:15]
Introduce a problem that is interested me for a long time but I think it's difficult study in the monkey which is the control of the circuit configuration I'll explain what that is later. And and then we'll study that for in the mouse using to support the late match the sample.

[00:03:33]
For years and years I've been studying this problem like trying to understand the brain one dot at a time and and we asked monkeys to decided that the reaction of these red dots they sensual which. Is the wind blowing then and we control the difficulty by controlling the probability that the dots moving in motion are displaced you know for a moment to moment as opposed to randomly place with the noise and the monkey answered by making an eye open to most of our experiments Sometimes that.

[00:04:00]
Summarizing over 20 years of work now we understand computations of to some extent the neurobiology behind this the idea is stated kind of from a computational perspective is that then that information noisy information from the world this case that red dots is Prince duce by red letters Nevil it especially cortical areas that have direction selective neurons and prime it into what I will call the momentary up and small and quiet moment fluctuations that might favor right versus left as represented by this noisy distribution on its side where this is said this direction means the evidence for writing this for left and the idea is that you can relate that to the moment by moment Spike discharged from neurons in the effort for His will for it specially airy and he or b 5 and I know there are experts in that area in the audience right now now the idea is that is that neurons with longer time constants in the association cortex they represent over a longer period a type the accumulation as a function of time of that momentary evidence this might be the path that such an accumulation of signal plus noise takes until it stops and has absorbed upper bound which means terminate the terminate the decision at this point in time with the right word choice had to reach this bound the lower one it would be a luck where choice now immediately you can see that this idea this competition will idea explains the tradeoff between speed and accuracy because if these balance were closer to the starting point of the accumulation then say this probably had been a left choice which would have been incorrect but at the savings of time on average anyway Ok so it turns out that this framework does really well in explaining the were.

[00:05:44]
Call evidence accumulation or accumulation of pounds on has called ripped a fusion with balance. This framework explains the behavior of the monkeys and human it's beautifully so if I plot on this graph the reaction time of the monkey that's the time that elapses from onset of the motion to the beginning of the monkey's I hope it responds if we plot that as a function of motion strength here sign so that the positive mean right where the negatives we let roots of this is easy right were easy leftward in the difficult cases in the middle which are most of the trials are in the middle and you can see that when the when the motion is sweet the monkey takes almost the 2nd on average when it's drawing takes about 400 milliseconds on average Ok so so that's kind of cool the nice thing about this framework is then it gives us simple of mathematics to to explain essentially the 1st pass of science of this virus random walk or if you should and team being the reaction time is claimed by this equation and very briefly it only has 2 important parameters it has a bound which I've already introduced which kind of mediates the speed accuracy trade out it has a just a single constant came here which multiplies the motions to see it converts it into units of signal to noise and then there's a non-decision timing of the part of the reactor head that's not spent doing this deliberating on the output Ok now that's the equation that spits of these so this data and that's the bit but the cool thing here is that this mechanism a computational that it isn't at this point but at the moment I'll show you the allergy that a computational mechanism reconciles not just how long it takes to make the decision but the probability that that decision will be the correct one or a right word choice in this case and so that function.

[00:07:33]
That function is short here is a logistic functions the probably of right is a little just a function again this product out of these 2 bear these 2 brothers now the. As to free parameters we're used to fit this these reaction times and now we have a complete prediction just based on this bit we make this prediction for how the choices should go so let's say these 2 motions strengthen these to the should lead to perfect behavior all left choices all right choices and the graded change in behavior should go exactly this rate according to this equation and this is where the data live so it's a pretty good job making sense of predicting not fitting but predicting these choice proportions but just by measuring the reaction times and gives us some reasonable reassurance that this is the kind of mechanism that play in the brain we've been to we didn't invent this idea it came from classified work in World War 2 both from Abraham Wald was doing munitions quality control in the States and from Alan Turing actually as part of his of the crip sion of the Enigma code I can tell you some about that in the q. and a he liked but of course it was used by many psychologists mathematical psychologists of.

[00:08:52]
You know and I'm just was that a few of them here that are in champions of these kinds of models to react with Stephen Link in particular. Ok so when I was a post-doc with builders. We started we thought we'd try to find a candidate they had. With that would read as the information from the visual cortex and so we thought Ok they must proceed permission from direction selected That's the dia stands for from direction selected neurons maybe directly maybe indirectly and they have to project to argue about hers as not directly or indirectly because ultimately the monkey answers with his eyes and.

[00:09:32]
They should deliver exhibit long time constants like they should have persistent active the other words the capable of computing something like the in for which we suspected was going on the accumulation of evidence mathematically integral and and so over the years we bill I other labs as well have studied the cortex for like the old We've been in the medial intra parietal area now and and and but I'm going to show you some are my favorite areas the lateral is proprietary.

[00:10:07]
In l.a.p.d. which is a is is on the lateral bank of the lateral of course of the enterprise Also office which I'm outlining here roughly where the oranges there were identified by Richard Anderson as the part of proteins there he said and the projects of the frontal I feel the circular with so you can consider these neurons part of sort of our knowledge system which is what we think the pride of love but they are involved in in what we do with information with our eyes like making a good steward wrecking intention.

[00:10:39]
And these neurons have persistent activity the spatial a selective Let me illustrate that by showing you a very simple working memory task over a flash a spot roughly here and we're going to ask the monkey to make it it's what's remembered location he doesn't get to see that great you got that's for you to remember that spot was but let's listen to neurons hopefully you'll hear then as the monkey performs this very simple working memory task.

[00:11:08]
I hope you heard that and I can't see anyone and I don't hear anyone so I have no idea I'm just talking to a walk and someone to say that they've heard this report and it's not a great Ok good so this is the spice the red spy came on in our mugging told to make a slight robot buggy maze like and that was the yellow streak you saw of.

[00:11:28]
That's persistent activity and if you ask me and many others the this person actively holds the hall mark. Holds the key to understanding our mission because sort of all of our mental acts that we would call cognitive have a freedom from immediacy that is they don't have to you don't have to do to act in the time frame of the immediate changes in the world order we have to to keep up with a body in real time it was a little real time to bounce but between that we want some flexibility I think evolution that's what led to the in coate versions of what we recognize as mission in us now in the in l a p not all areas of the association cortex just you know chock full of these kinds of neurons but now I keep these neurons are specially selected they have other properties and other and other association areas so if I flashed a spot over here.

[00:12:25]
Ok now you didn't now there is no elevation activity so that's the persistent activity that spatially selected and what we do in the lab is we contrive tasks that are they are contrived not natural we can try past a certain decision making by setting up experiments where one of the targets one in the response field will represent at least the choice that the monkey thinks that the right is that does the moving to the right and makes that I'm unfit as opposed to making this sign and when we study those neurons I think we got kind of lucky and they didn't just show us the choice nation to give us insight into the decision process leading to the choice because when we started them in the reaction time setting we know.

[00:13:07]
The window of time that the monkey's making his decision on every trial and here I'm showing you just a few motions traced and the solid curves of the trials where the monkey chose the target in response field you can see there rising and and these the firing rate to fire it by the way is a function of time 0 was when the motion was target sergeants that these daft hers are the trials are trials where the monkey shows the target outside that response feel Ok these are average of about 40 or 50 yards so but this is true on the single neuron level as well and which you immediately appreciate is that these neurons do tell us what i know but the mark is going to make a solid person rip up the deaf ears but they're also telling us something about the quality of the prints they used to make the decision because they rise faster for the strong emotion that's the one I keep showing in the examples 50 percent of parents but they also they rise at a lower rate or even the energy might say for this after when the when the motion is weakening the monkey might as well be guessing because we're short it brand of and we put no average motion signal in but but of course there are fluctuations in the display and the monkey makes answer based that and you see that revealing itself gradually as a function of time here of the decision this red set of curves is just one of the intermediate motions right thing you can see it occupies an intermediate level of rise in decline Ok so we get the idea that these average is quite conform to who you know might be in a representation of the accumulation of evidence for that in Target in response field against the target out of the response field that's the against part because the responses can decrease in response from their starting point here where the process begins to unfold Ok so I'm just writing that as in a row.

[00:14:53]
Ok now this is the reaction time Tasso these curves you know are clipped that the median reaction time they cut out any activity in the tires a cat but But what you can see is if we align the same. Sense of a fire of of trials to the time of the eye movement as a cat when the monkey makes its answer you see that these these response of come together in a very stereotyped wait for the sick cat and if we look at all the trials we just break up all the trials that ended in this choice for the target response and the response field and we are and we break them up and we group them by reactions high so this is not the fastest quantize up and the blue the Indigo is the slowest quite tell now you really get the idea that they reach a common level about 80 to 100 milliseconds before this a cat begins so we think this is a sign of a determination threshold that something else outside of l. Ike He is effectively blinded reading off the spike read about like he or many any of those other areas of it you know that do the same kind of thing that I like he is doing but they're behind it correlated with their like he because at the time of that readout like he is roughly at $62.00 or $64.00 spikes for a 2nd plus or minus Ok let the sign of that barrier threshold Ok now I know you might notice that these responses don't come to a stereotype level and that's an indication that we're in the brain we don't have the fusion between our upper and lower bound but we have a race between the mechanism that say has the the right word choice target it's responsible so it's accumulating evidence for right and it gets left but there are other neurons mostly on the other side of the brain that if you're late evidence for left it against right because the left.

[00:16:41]
Eye choice target is and it's responsible and of course right minus left the left minus right at least at the level of the video monitor are perfectly and correlated they may not be perfectly at the correlated pray but we'd see some process like that so when we record from a right we're prefer a neuron when the choice is left work we see something like this coming from a right where there are Ok or vice versa.

[00:17:05]
All right now I want to say this chip this idea generalizes and it probably is not as simple as integrating it different fiery from from n.c. and I'll come back to that point again later but we can train monkeys to reason from probabilistic symbols and so your monkey has been trained that there are 8 different kinds of symbols that come and go and whenever he's ready with an answer get a reaction time experiment he can say I want to choose the right target because these symbols they confer or they confer a weight of evidence that either favors right or favors left and the amount of weight of evidence is something we control Ok so we signed them different degrees of reliability it's sort of like you know so the right thing to do is combine evidence but combine it weighted by the reliability it's kind of like you know all the sort of news that let's say you happen to walk past sun some fascist you know a place where there are people showing Fox News so you see you know is it on reliable source and you walk past it but you know or maybe you ask the owner of the shop to shut it off or you have to get your haircut there anymore I do that that's that's my acts of civil disobedience and or or maybe you know you hear n.p.r. something else that the think might be a little bit more reliable and you will and so you might pay more attention to that Ok I hope I didn't offend anyone with that I only hope I didn't offend you not out of compassion or something but just because I hope that you wouldn't be offended by that.

[00:18:33]
Anyway so so what we do now is is we study the same kinds of for us we know where they're responsible there's I'm going to show you an example of a neuron that's accumulating evidence in this past and this is neurotic prefers this target is going to happen would have persisted activity before in the late I mean it's this start.

[00:18:57]
Ok so that's the trial Lucky's like position was in yellow you saw that he ultimately chose this spot and of course you couldn't remember what all shapes were but I'm going to tell you what this graph is at the bottom now those are the spikes you just her and I Some someone will interrupt me if you can't hear the spikes anymore and what I plan in here is the weight of the evidence the log likelihood ratio in favor of this target the one that's in response feel and what you can see is that the 1st shape favored that these yellow is neutral evidence Ok And this is the cumulative evidence in units of log likelihood ratio for the target in the response below so the 1st shape did favor the target and there's a little elevation in response you might see this the spikes a little closer together than they were back here Ok Now the next shape reversed that it was pretty strong evidence and it actually got us to the other side of neutral here and so it's so the net evidence is not favoring the target it's a longer here the one to the right and you can see that there's sort of a little decline here now you know you can use me reading tea leaves here with these that you know this barcode But you know the over many trials we convince ourselves that it's actually a real dip in the fire and then you see there's mounting evidence starting its shape 345 and 6 before the monkey stopped and answered and and you can see that mounting evidence is associated with a crescendo with the fire it and you heard that Ok let's hear another another trial.

[00:20:28]
So this one had kind of the opposite patter in this case the 1st few ships were favoring the target in the response field Ok And you hear the rise of activity and then followed by the client is there's meandering but bouncing up and it's all to the favor of the target outside the responsibility now you notice what's happening here is the firing rate and single trials from random kind of random there are in the pride of cortex the Deep in the brain is showing you that it's doing additions and subtractions it's keeping track of running a running sum of basically positive and negative numbers Ok so it's true it's intriguing those increments of documents and show you a few the accumulation inspiring rate don't let anyone tell you that when I show you an average that looks like a like a ramp that that average is a bunch of individual steps on every trial this is these are these neurons I have not only persisted activity they have graded persistent activity they can change many times over the course of a single decision Ok.

[00:21:34]
Now I'm not showing you all the behavior from this you can read the paper if you're interested as the monkeys behave as if they are accumulating evidence in units of like likelihood ratio which they they then combine and stop when they reach a critical level and that explains their behavior Ok now so this is we think is a pretty general mechanism for deliberate the liberating are really making the plan but basing it on evidence in this case the worlds of perceptual decision.

[00:22:05]
It's a little bit hard to think about this in terms of in terms of just integrate into integrating evidence coming out of the visual cortex that seem to involve some kind of associate memory that links the weight of evidence to each one of those shapes you know we train the monkey like crazy so they can learn them again it's not a very affable task the only natural part of the task for the monkey I think is that effectively this is like foraging he's looking for information that there is one where he's going to get it before Ok so now I want to introduce the circuit configuration problem of the do it the easiest way I know how which is the interest one particular kind of configuration change which is rather reading problem Ok so.

[00:22:46]
So so imagine that we train this monkey to maybe left right decisions upon a path of dots in the upper periphery near near the you know more or less on the left of the midlife but you know you know in the upper Henry field and he would do that by saying left with this sign movement right this side of it let's imagine this target is in the response field of a neuron well on another trial of the monkey strain we might put the red and that's over here and ask him to choose between north east and south west in this case now he's going to use different information to say Who are these and and we have another problem like with both targets at the right Henry field and I haven't choose between up and down and again this is still you know kind of like a neuron that we might expect to accumulate at.

[00:23:33]
It is for motion but now it has to get its evidence from a different part of the visual field I'm I'm going to I'm going to look back at the want to make this explicit this is an old to be acceded Luke O's patter you know another showing the activity of this this.

[00:23:51]
Sort of brings in the visual cortex the old people right right if you tell from the eighty's and let's just look at it on Mitchell cortex where these things are so the target representation is over out here on the periphery Ok it's in the upper field rights the upper upper periphery and the Northeast Southwest doctor you know not too far from the vertical Meridian which is over here and and you know Mike to occupy this much space given its magnification factor the up down over here and the left right over here now neurons and ally p. and l. The elsewhere but a single neuron or a likely that we report kid has the capacity to get information directly from all 3 of these regions which are centimeters apart Ok so how does it do that well we don't really know Ok it doesn't have that kind of connections to all those parts of the actually spread cortex Well probably not directly a definitely not directly from the one but probably not directly from Mt either Ok so that's a routing problem and it's a it's essentially a worry that I would say it could be stated more generally have been the same decision clan or proposition imagine that represent the state find support from diverse sources of information our brain has to solve that problem it's a kind of tipping constraint everything isn't connected to everything we are not a neural network you know an artificial neural network Ok that that we're sure Ok And and so also let me let me introduce another kind of a circuit figuration.

[00:25:34]
Change which is that it's context of the depends on context how can an associate of memory for example like these those event marry a lot of weight of evidence with those shifts change the level of level of persistent activity So now I'm kind of stating the integration problem of one that's more like l. like he receives an instruction based on the content of the associated memory Ok so that's a very different kind of way of casting the computation of cumulation a better that's it's more like the language of computer science of thinking about these the l.a.p.d. is affectively out like a like a finite state machine and the in the end the the shape one of those shapes the Pac-Man shape or something trigger in a source of memory that turns into more of an instruction to change the firing rate by some amount Delta which is effectively the content of the working memory I'm sorry of the associated memory so.

[00:26:33]
Someone could say how does the saying information lead to one behavior or another or contingent on a 2nd contextual cue so this is another way that a certain configuration might be thought of as as being under the control of neural responses so circuit configuration to me is the Europe was the Aldous translation of a hard it's like college ee concept called control and I'm thinking of a very famous paper by shallow as a Norman called controlling and processing so I would argue that systems neuroscience has done a reasonably good job of of advancing our understanding of signal processing in fact when I use the word integrate and momentary out of it's things like that and direction selectively we are trying to understand that a confrontational level under the rubric of of signal processing but when we start talking about how a signal changes the way another circuit operates.

[00:27:34]
Now we're talking about controlling and routing is an example and context dependent chains of example and I think it might well Act Think about it it's hard to understand how that works I mean because these changes in a certain configuration overlay must be signaled in some way by information of spiking neurons because we don't know or the way that they're actually communicated permission but they at the same time they're firing it shouldn't change very much or will that circuit we contaminate the actual signal processing so I think this process flies under the radar of our usual techniques of understanding or of computation which is with firing rate.

[00:28:12]
Ok so and we thought this 2nd one this context dependent change in the way a circuit operates based on a context will cue another cue is on his one it could be studied in the ballots and so this is our 1st foray into that and that's what I was spend the rest of the time talking about and I don't have to rush like it's the 1st part of it kind of losing track of my own time is all I have is my watch.

[00:28:37]
Ok anyway here's the task we this is a simple The olfactory delayed match a sample test and it was brought to the lab by Herbert he goes I heard his Chinese and his general whoop and and but I had a problem birds talk when he's on the job market and hurt train my eyes to do a simple The simplest The lady asked sample is only 2 possible orders I'll call them what they really are so in the way it works is that the mouse gets a little snip for half a 2nd of sample over a or b. he waits through the lead which is typically a 2nd and a half sometimes longer and some of our experiments and then and then he gets another snip which I call the test Ok though it's in the late the question is is there is there a match between this and if there is if it's a followed by a or b. followed by p. the mouse should look to the left port for a report and if it's being followed by a or or a followed by the other it's not math you should look to the Right now that's the limited sample but I'm going to concur if you think about it slightly I want to say one of the thing about that is that it's formally Quillan to the exclusive or problem it's the same truth table if you think about it Ok so and that's kind of interesting for computational people because it's not savors one of these poster.

[00:29:59]
Archetype problems for the needs of a non-linear kinds of operations and I want to be more about that except I'm going to show you an interesting way that the mouse brain solves and sluices or and it's not by embedding in high dimension. So so here's a here's where I want you to think about the problem I want to touch want you to consider that the sample odor might play the role as a conceptual q. and then her context sample equals a then test order a needs the left and test the intervening so right but under can text you will oder be then then test the intervenes gold and tested her aids go right Ok so we could think of that as a change in a certain configuration and I would be telling you that I'm not going there wasn't evidence but that what I'm about to say so so here we are thinking about doing the task.

[00:30:58]
You're looking for below here is looking he's looking to the right. Match trials and so forth and I would just tell you that the heart only hard part of training these mice to do this is it's a match I'm actually using a go nogo but we wanted a separate behavioral task and the hard part is just getting them to with to delay their licking Ok because we wanted we wanted to not see contaminated responses during the sample and Aleck.

[00:31:32]
Ok so very quickly because this turns out to be the least interesting part which I hate to say because I am an Earth was the allergists of love the sounds of neurons but we did a survey of the neurons in just a selected set of areas that represent the odorants some association area and the motor response Ok and so l.m. stands for the extra lateral motor area it's part of Mary and to study by.

[00:32:02]
Carl's boat especially contains the bird you know essentially think of it as a frontal I feel for licking of a whisk to Ok and then. The pure form of course is the primary all factory according to the order from the cortex and that is a classic Association area Ok not surprisingly in the pure form cortex we see lots and lots of sensory neurons surprisingly a lot that are just sensitive selected for these 2 odors and here is a neuron that responds at the time of sample which I'm delineating with these faint half lines this will be the dilemma she a lot of rats like this and then the and then between the testers also the linear 85 leads to a faint dash light sample of the late test and so this is an order this is a neuron that responds to odor odor be used to read the perceived by your people perceiving a I mean and and then also response to be when it appears the test when it occurs the test so here is you know a d. and b. these 2 this thing in blue Ok And so I you know maybe there's a slight difference in the way this neuron responds to the Pentagon whether it was preceded by a or air b. but.

[00:33:14]
But that's that. So this is what you might expect mostly from a sensory neuron many of these neurons have persistent activity that last through the late that's important because in order to solve the delayed math the sample task you need some kind of representation of the sample at the time of the delay at the time of the test sorry to make a comparison Ok well if you're from cortex we see a lot of neurons that are selected and and what I'm showing you here is the seal activity in text based on her c. major In other words it's a sort of get in the overlap and so it's a wonder is an overlap between the response to a over b. or hear b. over a and and and so that's not surprising we see in around like that in the orbiter of frontal cortex too and although some of them have more big selective in the latter part and they are Len we see very few that are selected for the odor.

[00:34:13]
And it's very at that Ok so when we do see them Ok our course the choice itself which is effectively which way the monkey get a look that is the dominant role type we see in a land so you know almost all the neurons that are that are related by that task in all our lives are sensitive to look left or right they don't do much they don't see much before that.

[00:34:39]
And and let's see what I want to tell you about that yeah that's about it so this happens to be a non-action neuron we see as many battles on that there are the words right and left licking each atmosphere Ok so here's the l.n. you like to be indexed for living left looking right or out and and we see some a little bit of that in the o.s.c. as well but it's not very strong and very little of that in your form court case that's on a surprise I want to just emphasize one thing is that at the end of the of the delay so this column here where where you really must have some kind of representation of the old records office task we don't really see much of that or all in all that but you know if you a few the significance but with very very weak selectivity Ok Now what does all the regime neurosis lines these days sadly you.

[00:35:41]
Used to train classifiers to decode information that we're reporting from you know up to 80 to 100 neurons are often from the from o.t. prophylactic pros in these areas and so we train some poles and for vector machine to classifiers because you know everyone wants us to do it if I had my way we would never do it but many cases we do it and and you can ask me why I feel so strongly about that later but I'm about to make a point but we see that in that these the classifiers can decode from pure form and ofc almost as well or at least blow it all if the about as well as the mouse performs the task by the way every now and then I might slip monkey for mouth I'm just so used to it and I apologize if I could see your face I would see someone laugh but but I can't do that.

[00:36:31]
So anyway so basically we have a representation in the sensory cortex and that's probably you know that little trial type of selective use little model is. At the time a test based on what this sample over was is one of you know yet the decoder exploits in order to solve this No of course we can solve them and not the match on hatch in ai learning even with a single Lawrence let alone with a population of one single trials.

[00:36:58]
Some of them are you know they're sort of like that's trivial because of course those are ons that are telling us which way them the mouse is going to look Ok now so this I'm supports what I call the standard hypothesis like which is that up screen sensory areas and sensory Association areas result the decision matters a son Matt and his truck I learned or maybe even a I learned could be decoded could be acting like what our S.T.'s doing to the inputs from the.

[00:37:27]
Upstream areas I'm channeling what I think is characterized mainstream computational systems neuro science. Ok I'm not that's not my own prejudice so Ok so so we decided that rather than just accepting that at face value we would test it and herd came up with time a clever way to do this which is he uses these mice that express.

[00:37:57]
Chair adopts and only in viet we're kind of created these mice out of ones that express the yellow fluorescent protein but in which we export expressed our the absent in the league at the other it's the gatherer neurons we do that in both hemispheres when we and we interact a bit bilaterally with the light but only in the sample of the late period though our logic was we wrap this down we have to reckon with some pounds back of the neural response so we taped this off so that by the time the test comes which is the 1st time the mouse can make the that decision the cortex is back to normal the monkey can hop the valves can lick parts of the mouse can live normally and so forth and I'll show you a controlled ferment that demonstrates that we have Ok so the idea here is there is that if upstream areas could solve the task they can only solve the task when the task and of the and at the time of test they have access to the lab so we're doing this in activation I'll tell you what I would see just as kind of a control of the moment but our main emphasis is on Ok And here's what we found we found that that manipulation that intervention impaired performance for manically So here's the control trials and here's the laser trials and each line is a single experiment Ok so there's 14 individual experiments this Dark One is the average and what it showing is just the proportion correct and you can see we went from you know on average 97 percent correct down to you know something below that it was.

[00:39:35]
It's a pretty striking deficit. And and you know as a as I said you know if see this is a surprise a lot of people we saw a much less of that that was that of a deficit it subtle but not significant Ok. And. You know but of course you keep in mind that there are many areas that connect the factory have to feel you know or the ball that's the way to the.

[00:40:02]
Many path yet the order information which obviously is necessary it's all task into Ala. Ok so. I'm not going to talk about all the control experiments that this one is like I think the most informative or among the most and what we do here is we train them out on another kind of trial where we get into new odor c. and d. and and these trials always work that c. is always in the sample position but then the test of the scene is gold out and he needs to look right Ok I should look and look right so in this case this Q Do the sample or or what I mean to encourage you to think of the context is is irrelevance it's only the test order is needed to solve the task Ok And so we have to leave those trials in single experiments and so what we do is to here are showing that we reproduce the result I showed you in the previous experiment that is a pretty substantial impairment in performance when we enact it in a lab and be 88 task but when we activate in the sea cross with the task we see very little apparent a little bit in these 2 mice but for the most part.

[00:41:25]
So let me let me tell you what I think we learn from these activation experiments is that Alan is playing the poor all in the late bachelor sample tast pacifically under conditions where the late as instructed. Ok so not for example the c.c.d. to us in activation does not impair the animals' capacity to translate decision about order to a lip response Ok so you saw it was a he was able to perform the task when we did the activation and he was able to perform the c.c.d. control test this well obviously one per minute look if we can produce an apparent liking if we keep a light out through the looking period we can shut down looking.

[00:42:04]
Ok so then and then this is so I think another important point is that it be in activation of a lead does not affect the response of the ofc in pure form I didn't show you this but if we do that same as the n.t. coder we can do that even under trials where we had an activity in a Ok so what we learned from this is that signals upstream in upstream areas they can impact.

[00:42:29]
That's what happened must happen in c.c.d. control for example but they don't resolve the math vs the not that despite our ability as a statistician running a machine learning direct the end decoding swipe of our back and we can decode so I've posed this as a cautionary tale you know because it is the range or ofay whatever the right expression is for everyone now when they have datasets and we're getting them in the punky now were among the 1st users of the new neuro pixel Kakkar pixel probes you know if everyone wants to decode you know information I might he was that it's basically a waste of time because decoding is no guarantee that.

[00:43:13]
Any downstream structure achieved this in this case not only does a high level not decode what the information ofc in pure form but we could say that any area that had access to a dollar and didn't and was not blocked was also unable to send its signals to l.m. to allow l.m.c. perform the task so so I think it's a good cautionary tale the next time you're someone telling you that they decoded information and and they think that they're for the animal use that you know a strategy like that.

[00:43:47]
Ok so I'm not arguing that you create can't decode information and say that therefore the information is present in a population I mean it's a useful statistical tool but that but but that's a different story so this hypothesis that I mentioned what I was called was sort of the standard view is that of Premier is resolved at 1st sign that they instruct l.m. or Alan being that a coder I think we can exclude that and so we now consider an alternative is that A.O.L.'s is capable of implementing one of 2 sensor response mappings and Len and he and this computation must occur in a lab and that decision and this this sense response happens you can think of it as a decision between 2 rules that test odors the instructor left and right respectively so that's under under under actually context.

[00:44:41]
When the sample order is say. And or the other rule is the test orders and the instructor right and left respectively in other words that's when the sample over is the is is be Ok so that's. So the sample or the terms which will is going to be instructive.

[00:45:03]
Now there is a problem for me with this alternative hypothesis and that is when we did our survey of the neurons we did not really see much of a representation of a sample odor in a lamp especially during the last day at us and definitely that at the end of the leg if they need what we saw it was inscrutable week so we were worried about this and we decided to take a closer look because her had the wherewithal to realize that much of our samples were from deeper up deeper layers 23 and and and then the deeper layers of the cortex and so we didn't really sample much in superficial layers 23 So what I'm showing you now is to photon imaging this is not the raw fluorescence signal this is our.

[00:45:54]
Direction index computed from the calcium signal Ok And again this is the sample period of the late period this gap between these the these double their flights and this is the test period and what I'm showing you this is the heat map shows the selective the absolute value so we don't care whether it's selected for a or b. but just selectively and you can see that there is a sample in you know 40 something neurons in this field of about 220 yards that.

[00:46:27]
$200.00 plus there's a of of of of the sample of the sample over in this case often spanning the entire the late period now now it turns out what I'm showing you is the is that this is one field and so has happens this field is was at a depth of $1200.00 microns.

[00:46:50]
That we see these are us when we go deeper than that we rarely encounter them maybe a few you know in a deeper layer in Layer 3 some are deeper in Layer 2 reactions. I will refer to this because the 1st 100 microns of the 1st really soluble are level of I mean you're a level of layer of Layer 2 which is so I'm going to call these later to Nora this concentration of her hands that had never been seen before we kind of felt that something like this had to be inhale and and there it is and it's sitting up there in the superficial layers now by the way we're sampling these other layers I'm trying to show you this but we think look cells throughout all the layers back there the dominant starts in this 1st 100 microns Ok now they come in flavors now I am showing the Rossignol So these are now single horizons Ok and some of them have pretty stereotyped response when you know those spots of the piece electric neuron this is anything like that neuron priests there are stereotypical It's response you know they're not perfect there are some response to the other odor or suffer spots of the other or right and some of them are more variable they respond it's sort of at any time during the delay during the samples or delay so they kind of can span I could order them in terms of per order that in terms of the roughly when there are people peak response was Ok and I'll mention these again but they come in all kinds of flavors of scattering in time and duration so this is are you no longer ration or are a scattering her on and something that has a little in between Ok.

[00:48:35]
Now now we don't know that these layer 2 neurons are mediated the behavior we're working on that right now because when we get activated been activated all the pyramidal cells we assume they're among the primal cells because they also were only expressing the calcium indicator decaff 6 have been in brambles also but this is the evidence we have that they are related to the task and you know in monkey neurosis we call this a choice probability analysis looking at single trial activity and it's relationship choice and what I'm showing you here for example is is intuitively like if let's imagine that we have a neuron that prefer sample order in a Ok and then we just read its responses to sample odor b. which are you know there are some responses when that response is on the higher end Ok that's when the mouse makes more herbs Ok when they when we take the cells that's the neuron the preferred is a and and look at it and rank it then it's and again you know just just take the right order of the responses we find that it's when it was responding it's in the lowest percentiles that we see the error and so they pile up in his left hand corner and you know there's not a big effect because the overall error rate in this experiment was only 11 percent so you know the vast performing at 89 percent correct but.

[00:50:05]
But but it's but it's the just simply reliable So that's the best evidence we have that these neurons that there variability in their calcium whatever's getting rise these calcium signals presumably spikes but we haven't actually measure that directly. Is a is is related product trials of the behavior of the animal the success or failure Ok so what's going on I want to share 2 relevant observations with you.

[00:50:33]
What is that the Imp. Merit is alleviated if an activation doesn't span the entire sample in the leg so here in addition to in activating during the sampling delay there is a condition where the sample where we only an activated the sample Plus early delay Ok built or just a late away can see that compared to control there's a slight decrease but it's the biggest decrease we get is when we when we are in activity for the entire time now with a shock when Whom are we trained tractor we train our current neural networks to perform a task and this is just a kind of a fancy way of stating something that might be obvious to you it's this these these curves represent Weatherby train the recurrence at each level technically you could have done what I'm about to say just to love but it's kind of fun to think of it this way so imagine that only a lab you know there was was was was there sorry that all the only areas that could have recurrence and therefore persistence were prior to.

[00:51:43]
And then in that case what we predict is well they have to kind of solve tasks and then if we if then if we. Simulate an activation experiment we should never see an effect of an activation because we're not activating the area that's making the is that's making the the decision effectively that solving solving the problem the fact is the logic behind Herbert's in activation in sample through the label leading airline intact at the time of the test when the delayed matches sample the system can be made now if we activate if we put recurrence only an ala Ok then that makes it out another obvious prediction which is that is which is that with recurrent network you if human activity at any time during the Senate there in the sample in the way you should get almost I'm.

[00:52:34]
Sleep in activation Ok Because the complete apparent because there is no representation of the sample at the time of tests which is what you need with recurrent network needs that in order to be able to solve the problem so so it says so here's what we get if we allow for recurrence in all 3 areas in training network and what we see is that we get effect of increasing.

[00:53:03]
Increasing impairment as we as we activate more and more of the delaying especially late with the late aha but that's a problem because that neural network neural network solutions like that with persistent activity effectively providing the leakage between sample and the ability to compare the test of the sample implies that in late to lay really knocked out all of that recurrence then we should have got the same impairment here as here but we don't see that so we don't really think the neural network solution is the recurring neural network is the way to think about this and instead we're so we're toying with and testing now other models but I'll come back to that 2nd by saying the 2nd relevant observation which I kind of really introduced which is that the mice don't generalize to the sample that matched the sample if we tried it with new owners like see the cross with c.t. they have to be trained all over to learn the lady asked a sample Ok so they just learned another 2 by 2 contingency table like the exclusive or table back to plate but if we train them on this task we just sample it is either syrup and the test order is still era b. now we really can't call this the least massive sample but really it is a context dependent changing the rule and wife can learn that just as easily as they learned the cross with and this is just an indicator to us that maybe this is really the way.

[00:54:34]
About solving the problem they really are treating the sample letter of thinking textural Q So with those just really an intuition it means that we something must be going on in Aleppo during that a leap year if we think that all of those later tuners are doing something to the their arms that are left in the left right there are they don't allow it to.

[00:54:58]
Do interpret the test order as an instruction to look left or right and one way to do achieve this that we've played around with is with that episode with Taishan and depression it takes more than that of course but you know that's a flavor of models another way to achieve this is with getting models I'm actually my favorite part was actually the services like a priest and after getting my favorite as opposed to have think and then read it branches where one could say that like the imagine that the sample selective.

[00:55:28]
Sample selected nor oscillator tomorrow you know maybe synapses say onto dendritic branches where test order is that information that conveys information about the test Order era be. With say concentrate and bite by turning on or off and writes in this case show has a preset haptic control but by turning on or off it puts to that the neuron we might get to the next to the to the match our maps of his lip left the graveyards we could change which where.

[00:56:04]
We were going to drive the response Ok so we're these are things that we're studying now we have no direct evidence for that but the idea that the circuit would establish a change in the way it operates on the basis that contextual cue is what I was introduced to this work is being.

[00:56:25]
In the service of and we think we now have a system we're actually studying it so let me close by saying the premotor in our land remote area lands involve the key step of an all. Factory decision now I would say that's not surprising from the work of perceptual decisionmaking in primates because the key insight that I get from all this work that's just my own is that the way the way that that knowledge exists in the brain is a persistent activity that's the way it transpire across time and in those that are supposed persistent states don't live structures that keep up with the changing world sensory areas they look more like propositional things about what we're going to do to the world I call those afforded this is the seal in terms of our return from Gibson and I call them provisional forms because it's not like you have to act in order to know you have to organize information as if it's in the service of a potential action so the idea of finding that they have knowledge States or the use of information to do a behavior would live closer to the voters as the sensory system I think is it is just a just a page out of the playbook from systems or a science of system making.

[00:57:37]
Now upstream sensor in association areas must supply information about her identity by the lake that they don't supply a lot of with the decision I believe I think I've said that enough already it's a cautionary tale about the coding the working hypothesis now is that a all that precedes input about sample her identity it's also the effect of kind of activity of little left and right neurons like right here and such that one of the other older inputs and test leads to exploitation Ok And and the layer 2 neurons are promising candidate to mediate is change in the search configuration they're the only in pounds I've shown you that they have a trial by for those with success and failure but beyond that we don't know what we don't really don't know that they are causing but we're working on that now.

[00:58:27]
And I will just and also say that this complicate it somewhat complicated it's lose a war problem is structured as a decision about a position it's implemented as an instruction as instructed change in a certain configuration. And and although we don't know how and I would say this is a model of what's called executive function or Arctic control and it's a sort of control a control operation as opposed to simple processing.

[00:58:57]
I should call it simple we need both and so with that I'll stop I'll thank my collaborators Herbert who was a success on the job market Ashok Kumar did the complications and also was selected above all stages and Richard thankful who was an expert in all faction and factory systems or a science of the absolute delight to work with my support lots of guidance from very helpful people especially Carl Sloboda and and you might have noticed I started not end with this rare painting by the Greek here he is rendering his own decision process shall I paints a scene or puzzle peep or look later going on and here was his decision out.

[00:59:39]
Of the clear boils and with that I'll take your questions that we have time thank you my. If everybody wants time you open next drive me Mike Mike can you hear me or Skeeter you are from Emory Hey yeah so we were going to have quite a bit lately and I have read a lot of about this lot of folks from what I pay a person and getting maybe more confused than before because he's describing this interpreting feedback during the last hour for tech to add and auditory cues between layer 6 and the out of town mess that doesn't seem to be on your radar so I don't quite see whether order is just different or whether the task a slightly different override which P.C.'s The task is different remember the mouse does not know it's a great question and we're very aware of this were the the mouse in all of those the late responses the mouse nodes there is going to lick left or right so these are what we call instructed the late parrot no Remember in the late match to sample the mouse does not know which way he's going to look if he knows anything at the time the sample order and through the delay all he knows is what rule to apply.

[01:00:54]
Which sensory response mapping to apply if it's a test order a or b. look left or right or test order 8 or b look right or left so he doesn't this is so that that activity is is in you could say. Represented by this these later tuner are there representing the order.

[01:01:18]
In From the point of view of a motor circuit they're representing what rule to apply and we think what they're doing is they're free readying the circuit to apply that rule so that the difference between these these parents that if the mouse was can plant look left then you see this persistent activity invariably Clarke and Carl believes that this is supported by that look at that you mention.

[01:01:42]
I think their interactions between the hemispheres 2 of you know I don't remember that's where maybe there are forget that. There are. Just as a little comment because I could write in a bit of Matthew Mark and Berlin and he has a couple of papers I've been 2020 apart decision making process is a context dependent and he would certainly claim that it is to direct a commom in the areas that can provide the context of decision absolutely and Matthew and I have been talking about this for years and so on when and and absolutely I look I look at the mechanisms that interest me and so I said my favorite ideas about them are closed and critic and they involve things like you know calcium plateau potentials and and the spikes and all kinds of things like that and I believe that this problem this search these kinds of operations and to speculate really responsibly now but Matthew would agree with that that these these.

[01:02:39]
This is this problem the circuit control problem is the key to understanding tons of things that we blindly think we have any understanding about in systems or of science for example feedback there's not one single good idea about feedback and systems near us and it's because they all make believe that neurons receiving input from feeding back structures can somehow process that information without ever revealing it in their spike this church and the make believe that neurons that are providing this feedback can somehow pick up address the proper neurons that they're supposed to be meeting back to if there's some kind of a handshake that has to occur between areas and it probably occurs by neurons identifying themselves by virtue of some kind of change in state if there is a cold and rights and the reason it has to be unequal the rights and maybe just the basic rights is because to make a joke out of it if I was on was less to lose I'm committee these controlling operations should not.

[01:03:42]
Should not contaminate the actual signal processing this shouldnt lead to change that Spike rate they should lead to changes in the way to serve it can then lead to changes like rate in the process of computing which is the processing part I wasn't on a design committee I think a little at least it and I think that's what the field of attention studies but there happens like it's like in modulation the one or something Thanks a lot for him probably have time for 2 more quick questions from Mark Miller and then parents that me and that's programming on time have Mark you're after question.

[01:04:20]
Yes sure Mike I just wondered if you could say a little bit more about the the album touring work the enigma. Ok Very quickly you know the the the process itself it is cumulation of in units of blood like the ratio or products of like the ratio to a spread troll is called sequential probability ratio test gives the most efficient test for deciding between 2 heart disease if you can do it and with the least number of samples efficient sensible given the desired level of accuracy touring in the in one step of the decoding of the Enigma machine was to find random random letters that were intercepted Ok that would have been produced by machines in exactly the same state and it turns out that by if you line up letters one about the other from these random things there have to be cryptic we don't know what they say then then of the probability of seeing the same letter even in the.

[01:05:24]
Mutated or you know code the encrypted versions of ritual messages will fight it is that to hurt this deterministic operation that it should work and then to preserve probability of matches so he accumulated evidence from pairs of letters that match versus none at under the 2 hypotheses he machines in the same state or it effectively read them States and that's an e.q. related likelihood ratio until some threshold because they just had to get through as many as possible and thereby finally to end the machines in the same state from that they could figure out what the code was that 1st 1st set in the the rotors and from that we could decrypt all the messages from the day if you watched that horrible movie about snoring and you see these women on these flat tables lighting things around they were basically sliding around pairs of of letters to see in fighting I would.

[01:06:17]
Provide support for or or refutation of the hypothesis machines identical vs random States he wrote on the back of a piece of paper but his deputies that decision he became very famous big he and I did good bye bye Jack that he wrote about that in the in in a book called Good think thinking that I read I read it at this by accident.

[01:06:45]
All right thanks for a great thank you yeah so yeah I might. Hear from Georgia Tech. So my mind was more of a comment than a question I was just as you were talking about the the fact that the factory task that you design here couldn't really generalized for different odors I started thinking about the fact that you know if you had chosen a different path where you might be able to titrate this so you could play around with the quote distances between the sample and the match and I was just curious I guess or faction you're thinking it's more about identity of odor rather than something along the continuum that you could measure in play with in terms of the distances between.

[01:07:32]
These 2 dogs if you were in a coma Yeah you know it's a it's a great comment it's not the act this is something that one might want to do you know try to understand distance in order to space or chemical space and so for the repetitive questions it's the kind of thing that I absolutely want to avoid remember I came to this looking for in a context appendage change in a circle I want to study that because I see it happening every day in a month and I give you kind of a simple example of it but I but I want to understand because I think I have to do so biology and really to find scale kinds of elation that are not available to us in the monkey so I thought I could do with the mouse so we had our mice actually achieve the task that a lot of us some of us in the project were hoping that we could solve the task with the yes the m.b. code of stages and things that were going out of your form but I would always remind them that the moment this problem was solved upstream of b.l.m. which was my paper had parts from the start or c. or some area parietal or if it was a time or something but when he was done that I would just abandon the project because I wasn't it was didn't study distance and similarity in.

[01:08:44]
Actual perception perceptual sight I can do that in the monkey and we'd better than a mouse so I'm not saying people should do this themselves but I just think that was what I want I want is that it is probably I can't study animal which is the problem of love of affectively Cheney is a control problem yeah I got it thanks that helps clarifies what the motivation of the ng had been why do we need we can train them we haven't tried hard enough I think the furthest up we went was 3 different order pairs you know so if you know if we got out to 10 or older pairs for a mouse and see me live that long to get it all then maybe most could see our lives but that's nothing else I want to study Thanks All right thank you so much.

[01:09:32]
So Mike you have a couple meetings lined up in the 1st one who graduate students that starts Well it's important because for start 4 minutes ago and I think very progress and there are links that are presuming something you know people can actually see you but again thank you so much for.

[01:09:52]
Having our new patient and for your niece on the park.