[00:00:05] >> OK very honest are a great but you got a problem here there. Was a problem for many great scientists got to start physics chemistry physics and all. Because. I needed to cage the accounts that we're done with the giant. Field congregational pro-science. Weary public some classic papers on. [00:00:38] What is a moment series is actually really clear series of regrets to. Have to do speech he really wanted to come up with all of the stuff. That's when a whole Romo to study working out every particular parametrically in this entire century. Primates and their writing this is better fundamental to the way things are there's a science that. [00:01:07] Doesn't belong. How to look at animals in the mirror system integrators sensory information how will this in memory and how we make a decision about this really gets important over the long road which I think is going to go home. After this post office service on my wall street Harvard where he was a character. [00:01:29] In the climate studying decision making going out rodents and I think this inspired him to work with great things that are good basically to like do experiments in my lab and now he's got a really elegant lively to do experiments investigating the same things of memory century that you clearly think you have to be making very simple competition brings extremely late the 2 together. [00:01:57] Very well funded brought his career by the. Recently Science Foundation and he's been refused medical. Thank you very much thank you all for coming and thank you Bill Allen and Garrett for inviting me it's a pleasure to be here. So I'm going to tell you about what we're interested in my lab My apologies to that side of the room I'm going to be pointing on this screen so we're interested as Bill mentioned in what we like to call the neural architecture of cognition so what do we mean by commissions we mean processes such as the working memory how do you keep in mind some information how do you manipulate that information over a span of a few seconds. [00:02:43] Planning how do you use knowledge of relationships between actions at the route and their outcomes in order to decide what to do for example if you're playing chess and you know that moving a pawn would put pressure on a bishop how does that inform you in terms of what moves you're actually going to do. [00:02:59] What psychologist like to call executive or cognitive control and I like to describe that as your ability to control how information gets routed within your brain for example I could tell you right now today Red means go and green mean stop and you could go out and you could drive like that and it would take a little bit of effort a little bit of mental effort but you could do it so how is that just by deciding that you're going to follow that instruction you can reroute read from going to the brake pedal to now read go to the accelerator pedal How do you control routing of information in your brain and decision making evidence accumulation for decision making how do you accumulate evidence when you're making a decision how do you came late evidence over time for against different options and how do you use that accumulated evidence to eventually commit to a decision and implement these are the kinds of things that we're interested in and I got to that interest has been all described partly by being in room 4 rooms monkey lab where they were studying these questions and when my lab 1st started doing experiments on these kinds of questions about 10 to 15 years ago studies of these kind of cognitive. [00:04:05] Sometimes referred to as higher condition which is a term that I like to use because it makes it sound really sexy although I don't really know exactly what it means but I still use it higher commission makes it sound interesting so studies of this kind of higher commission add so little resolution if you want to know how did single neurons or many single neurons participate in represent this kind of information these studies were mostly done in non-human primates and the general idea was these are very sophisticated processes you need a sophisticated brain in order to implement them and study them and one of the difficulties in using our human primates is people that do this kind of work know very well it's extremely valuable work but it's really difficult one of the major difficulties being that it can take a long time to train an animal in a particular behavioral task that you've designed carefully to isolate a particular cognitive process and allow you to study it it's not atypical to see training times of 6 to 12 months parental And so as you guys know if you've read monkey papers you know what you typically see are papers that have 2 animals in them right typically one monkey was great and the 2nd monkey was kind of OK but at the back it monkey to show that it wasn't just in a single animal and so that makes the cycle if you have an idea you want to test it out and you find the results and that refined the idea that cycle but experimental hype on the cycle can be really slow and really hard and so when I was a Cold Spring Harbor I had 2 neighbors that had huge influence on me and they were Tony's atr and Zack Maine and they had started trying to treat rodents the same way people were treating monkeys and by that we mean trying to do kind of rigorous psychophysical studies of them rigorous behavior and trying to address the very same same kind of questions so I decided to join them and together we pushed a lot on that and the reason why we were trying to rodents initially with rats is that they're comparatively small comparatively inexpensive you can get good behavior out of them and they're mean able to modern molecular tools. [00:06:10] So here's a picture of one of the walls in the current behavioral training room in my lab These are 9 different rigs this is the computer racks a person is about that tall just to give you a sense of scale in total in the lab we now have $36.00 rigs and we run 9 shifts per day per rig. [00:06:29] So that means that we're to training a total of $324.00 individual rats per day and that's what we can do because of the relatively small relatively inexpensive that gives us about a 1000000 behavioral trials per week so it gives a lot of information to do careful statistical analyses that in itself is extremely useful but typical student or post-doc in my lab supervisors 20 to 30 animals if you're supervising 20 to 30 animals there's no way you can train them the way people train monkeys which is to stick with each one individually and kind of see what it does and then almost are to some kind of training you can't do that on that scale so the only way to do it is to develop computer code that will do the training for me that will take the animal through all the shaping steps that you want for a particular behavior and the advantage of doing that is that once you've got the code well you can just copy it as much as you want and you can paralyze as much as you want so the task that we study in the lab they typically they typically take about 4 months of training per individual animal but if we take 40 out of these $324.00 slots and dedicate them to a particular task which is a typical kind of thing that we do then on average we get one trained animals out every 3 days and that's a completely different world to having a couple of animals per year. [00:07:46] Let's try out many more things and so the idea is that we will use this as a platform to make that hypothesis experiment cycle to try to make it faster and make faster progress in a way that should be complementary to the studies that are being done in the prime So together with Akon Tony were part of labs pioneering studying neural circuits for cognition and rodents of course there are many. [00:08:09] Other labs that were doing this before just kind of like the thing that we pushed on was this idea of treating them almost as if they were small monkeys trying to address the very same kind of questions with the same kind of behavior of behavioral and quantitative rigor since we started there's many more labs that are doing that it's a big wave of people doing that some people here and you know there are many many labs participating in that my own lab our particular interest and focus has remained on these more complex Well you know what I like to call the higher cognitive processes that I described in the 1st slide and these in practice are typically turn out to be tasks that require a lot of time for training an individual animal and therefore their tasks where this kind of set up where we do this semi automated high throughput training can be extremely useful so that system that high throughput training has really allowed us to push the envelope on what we know rodents are able to do so for all 4 of the processes that I described in the 1st slide. [00:09:12] All projects and all 4 of these started with a paper that showed either that rodents were capable of something that people didn't think rodents were capable were capable of or that they do it in a way that is very similar to humans and monkeys in a way that people didn't think was possible so we're really pushing the envelope of how far they can go we don't yet know how far we can push it to be limits to it so far it's being pushed in a very interesting way at least for us. [00:09:38] So we do you know $300.00 odd behavioral sessions per day in contrast when we do electrophysiological recordings once we go beyond behavior and try to get into the brain to do recordings or perturbations. Each of the recording rigs we have for recording rigs and those require much more manual intervention so individual people sit with each animal during the recording session so as a result we only get about $3.00 to $4.00 electrophysiology call recording sessions per day so there's a tours of magnitude difference between these and that has actually felt very painful over the last 10 years because the pole project was we're going to move so fast that cycle is going to be so incredibly fast and actually we have a bottleneck but it's not so fast so we want to work on that. [00:10:28] And in particular using some technology that our car post lab at Geneva has developed and there's other groups that are doing this kind of thing they're really helping us for that So these are wireless electrophysiological recording and wireless generic perforation recording in particular is now through a company called Spike gadgets that came originally came from his lab and the reason that these systems are so useful and important for us and combine so well with our high throughput training system is that because with the wireless systems we do not need a specialized rig to do recording or up to genetics and so that means that we can really do it in all of our behavioral training rigs any one of our rigs now becomes and he she is. [00:11:13] So we're really pushing on that so that we'll be able to get $324.00 electrophysiological in up to the next sessions per day and we match these 2 numbers and hopefully move a really as fast as we were hoping to do when we started. So that's telling you a little bit of the background a little bit of that kind of framework in the lab a little bit of the simple technology that we use but to get into the science. [00:11:36] I describe some of the processes that we're interested in and the regions in the brain that we study we start by trying to guess which regions of the brain are going to be important and it turns out that for all 4 of these there's an overlapping set of brain regions that are relevant These are the initial guesses right it's not that we're limited to those we want to explore more and trying to understand what other brain regions are involved but these are the initial guesses and because they overlap across all 4 of these studying the these 4 different processes becomes really use because it gives us each one of those gives us a complementary view on these different brain areas and so we can get understanding that we wouldn't have gotten to if we focused on a single process and so what I'm going to do today is tell you a story from 2 of these projects I'm going to tell you a story from decision making and from working memory and you'll see that the 2 combined in a particularly interesting way and I'm choosing these 2 because they illustrate this idea that by coming here it's different projects of the same brain areas we can gain understanding that we wouldn't have gotten others. [00:12:35] So I'm going to start with evidence accumulation and by the way I would like to emphasize please interrupt me and ask questions as much as you like I'll keep an eye on the clock and I know you guys have to leave exactly at 1215 I'll be responsible for the clock so don't worry about it if it gets late I might say if you ask a question I might say you know what I'll leave that to later but don't worry about the clock I'm responsible that you please ask as many questions as you like. [00:12:59] OK So this is for me I'm going to start by pointing out perhaps the most commonly observed behavior of observation in the sea in decision making which is simply the fact that when you've got a really difficult decision to make it takes you a long time whereas when you've got a really easy decision to make you can do that very quickly if you're deciding between 2 awesome grad schools that have accepted you a long time to make up your mind if you're deciding between a fantastic grad school and a horrible grad school you decide really quickly. [00:13:30] So that has been observed not only anecdotally in your daily lives but also in the lab in a whole host of various types of decisions social decisions perceptual sensory decision is the stations very famously by Bill Newsome and Paul the economic decisions gambling decisions memory decisions visual search the station's value base decisions all kinds of decisions and because it's been observed so commonly it's thought that there's a common core process that sometimes goes into the name of noisy accumulation that perhaps can account for this observation across all of these different decision making the means and the idea behind it is very very simple really so I'm going to illustrate it now perhaps some or many of you have already seen this before but imagine that your subject in a lab and you're going to do a trial where you're going to get shown a stimulus and based on that stimulus you have to decide between orienting to the right or anything to the left. [00:14:24] And now pretend that somewhere in your brain there is some abstract variable an abstract scalar variable that we're going to call the evidence accumulator 8 and what exactly it's represented in the brain and how it works in the brain that's going to be some of the main questions but not pretend that that scalar exists and pretend that at the beginning of each trial the value of that scalar is 0. [00:14:43] And then as the trial of unfolds and you're seeing the stimulus every time that you see something that makes you lean towards going right add something to A and every time that you see something that makes you lean towards going left you subtract something from me and you just keep going like that over time until you get to the end of the stimulus and at that point you have to make a decision and the idea is that what you do is ask what's the sign of a if by the end of the stimulus the sign of a is positive you go to the right if by the end of the stimulus the sign of a is negative you go to the last super simple model. [00:15:15] But you can already see that it accounts for this observation. If the evidence is weak namely what you're adding or subtracting from a each time step of small then it's going to take you a long time to move away from the decision border. If by the end of say hold 2 seconds you're not far away from the decision border and there's noise that's what's going to mean it's a difficult decision it's hard to tell on which side you should look whereas if what you're adding or subtracting from A is each time stick times that is really large you're in to move far away from the position border very quickly and so that's going to be an easy and a fast decision so super simple model accounts for a ton of behavioral data of course people up and it showed up in their behavior literature about 50 years 50 years ago and it's been very very widely applied now in this model the idea is that each individual trial the trajectory of a as a function of time could be quite noisy. [00:16:13] But if you average within classes of trials with a particular strength of the evidence you should see the trends that I described before namely if the evidence is weak on average a should grow relatively slowly so shallow slope here where as if the evidence is strong the evidence should grow much more rapidly. [00:16:31] And starting in 1906 my cattle and Bill Newsome did a series of experiments that were really seminal experiments they started a whole school of many other people following this. What they did was they had monkeys doing a perceptual decision task a sensory decision task where they could control how strong the evidence was on each trial and they recorded from a particular part of the brain that I'm going to refer to as P.P.C. particular area in the monkey brain that they recorded from was lateral intra Prialt cortex L.A.P.D. which is part of P.P.C. here I'm going to refer to it as people see the record from neurons and P.P.C. while the monkeys were making up their mind before they reported what the decision was and what they found there were firing rates that ram it kind of like what you see here and critically when the evidence was weak the ramp the slope of the ramp was shallow it was small but when the evidence was strong they ramped much more much faster just as the kind of picture that you have over here and that similarity led to the proposal that people see firing rates in code the value of this abstract variable and that was a critical finding because this model this noise accumulation of evidence model had been around for decades tons of behavior papers about it but this was the 1st time that anybody had seen a signal inside the brain that corresponded to this putative variable so super exciting things said of whole school of thought and really we're kind of like following in that in their footsteps so their 1st experiments as I said were in P.P.C. and the idea was there's a one to one relationship between firing rates and P.P.C. the value of this abstract variable. [00:18:13] The evidence accumulated but there's been recordings in several other brain regions since then for example the frontal eye fields where the same phenomenon was found so for weak evidence shallower ramps for strong evidence much sharper ramps and these 2 critical regions are thought to be key areas for this process of evidence accumulation but this same kind of qualitative conclusion has also been seen in other parts of prefrontal cortex subcortical areas like the straight M And there's in the brain stem like the Sapir quickly so if you look at a subset of the connections between these different brain regions you basically see that everybody is connected to each other and of course I'm not even putting in other brain regions where they are which gives rise to indirectly to be this you know figuring this out or looking at making these observations as been happening for the last 20 years or so and multiple labs and the similarity in what's observed across all of these different brain regions has led to the proposal that maybe the mechanism for this the mechanism for how evidence is added over time how it's integrator over time maybe it's a diffuse circuit mechanism widely distributed across the brain we don't know. [00:19:25] The way I like to refer to that is essentially to say that we don't really know anything yet about the circuit mechanisms that's what to me this proposal really sounds like. So some of that when we started working on this some of the questions that we thought were really basic So when we published our 1st paper 5 years ago 6 years ago now some of the basic questions that were still in answered were questions like this week which of these brain regions is necessary or not necessary for the 7 is accumulation process not know where is the evidence accumulation computer is it any one of these brain regions or is it somewhere else that hasn't yet been recorded from not known how is the evidence accumulation computed what is the actual mechanism that leads to that accumulation Well if you don't even know that you've recorded from the place where it's happening of course you don't really know precisely how it how it's happening it's really hard to have a moment. [00:20:17] So your thought was Well let's see if we can use our rats in our high throughput training system to do this faster cycle and maybe we'll provide some information that's complementary to what's being done in the monkeys and unravel the mechanistic circuit logic helped to unravel that that was that was the idea and when we set out to do this the 1st question that we faced was well can rats accumulate evidence over time for decision making on time scales that are similar to what had been seen with primates both human and non-human maybe rats just can't do this and in fact some monkey researcher friends that's exactly what they told me you are an idiot rats are going to be able to do this you should even try so that was a critical question to address right can they do it and so the 1st paper that we worked on for this was a paper where we set out to design a behavioral task that would try to get rats to do this process and they would allow us to quantify or to state of analysis on how they're doing the task that we put in front of them to try and answer this question. [00:21:18] So that's what I'm going to describe to you now so anyone trial of the task are rats are in a behavior box where they face 3 nose ports that are in front of them and a wall in front of them and I know sport is a place that they can stick their nose into and we have an infrared beam across it so we know when they put their nose in there and I'll take you through one trial the task the very 1st thing that happens is that we have an L.E.D. in the Centerport that turns on and that's the signal to the train rat come and put your nose in here and the rat is trying to keep its nose in that center port until we turn that center and off we call this the nose fixation. [00:21:55] So they put their nose in there and then we have speaker to the right and a speaker to the left and we play randomly time clicks from both speakers both speakers going at once click click click click click click click click click. And then when those clicks those 2 trains of clicks similar streams of clicks and we turn the center light off so the right is free to withdraw and it should choose to go either to the right or to the left and if it goes to the correct side port it gets a drop of water we keep them a little thirsty that's what they were for that word if it goes to the incorrect side port then we give it a white nose and a time out and then the next trial starts So what defines which of the 2 side ports is the correct port the one that is seeded with water it's Which of these 2 sides right or left had played the greater total number of clicks that's what tells the rat which way to go so why do we give it that particular stimulus Well the reason that we chose this kind of stimulus is that if you're going to solve this task if you were to do this task 100 percent correct get it right on every trial then would you have to do would be to add clicks up over time 3 clicks up over time from the 2 sides need to be able to compare the 2 totals and see which one had the greater top things. [00:23:08] Adding auditory clicks up over time that is the same as accumulating sensory evidence over time so we're trying to set up a task where the optimal thing to do is to accumulate sensory evidence over time to follow the cognitive process that we want to study now nobody does this task 100 percent correct not rats not humans not my kids not anybody so there's a lot of wiggle room in that while the rats are doing it but not quite 100 percent correct and we're going to worry about that quite a bit. [00:23:37] OK So in this task I'm going to I'm going to skip over the evidence that rats actually are accumulating evidence over time it's in this paper and if at the end of the talk you want to ask me about it I'll be happy to tell you to go into it at length but I want to get some other stuff for now I'm going to ask you to please just go with me Keep your questions I know it's an important question so please do ask me. [00:24:00] But for now imagine that rats are actually accumulating clicks over time in order to solve the test I want to point out that there are at least 2 very critically different time periods in this task the 1st is while the clicks are playing and the rats are accumulating the clicks over time and you can imagine this as each click each say right words click being a plus one a positive delta function and each negative each left click being a negative delta function and you just integrate over time integrate those over time. [00:24:31] So while the clicks are playing you're doing this integration over time but then once the clicks end and we turn that central light off that's the go signal that says to the animals now you have to decide and now one has to be computed is not integration over time but ask what's the sign of a right because that's what tells you should you go left or if you go right and the difference between these 2 time periods is going to become important later and I'll remind you about it. [00:24:56] So we made these clicks random in time and if we make them random in time that means that on each trial we literally have a lot of bits of information because the timing of each quake is random can be anywhere but we know to the microsecond exactly when it was placed and so that gives us a lot of statistical power that allows us to do a lot of analyses in particular the model of the accumulation process that we had had a lot of parameters but with that statistical power we can have relatively low error bars on each of the parameters and distinguish some of them were able to quantify some things for the 1st time and it was also a big part of what convinced us that the animals were actually using gradual accumulation of evidence over time to solve the task and also another critical thing that this gave us. [00:25:43] That knowing exactly when the clicks played allowed us to build a model a model that modeled on a moment by moment moment by moment basis how this variable A was evolving over time within each individual trial. And so in response to a particular set of clicks we have an estimate of how a revolver Now we also have we think the process is noisy they're also known as parameters in there so actually what we evolved is not a deterministic Ethen deterministic estimate of A but A Probably distribution over values of A That's what we evolve over time so the equations that we're actually using for those of you for whom this means something where you can talk like equations to hold this and then when we get to the end of the trial we know what the animal actually did so for example if this was a trial that the animal went to the right we know that that means that this probably distribution must have all its mass on the positive side of 0 so that puts a constraint on this and we can use the backwards Kolmogorov equations to reverse propagate that constraint backwards to get the most precise estimate at any point in time what we think that probably distribution of values a is. [00:26:51] So that allows us to use the exact timing of each click it's a model that takes into account the timing of each click and the animals decision to give us to use all the behavioral information that we have to give us that estimate of the distribution of Ray and how it evolves over time when we're doing recordings we're doing recordings from neurons while the animals are performing the task then from each individual neuron we also have an estimate of how the firing rate of all the over time and we have these 2 estimates simple tenuously having those 2 estimates similar tenuously means that it each point in time over trials we can build up a map of how those 2 relate to each other we simply say when the value when the distribution of values of A is over here how much is the firing rate when it's over here how much is the firing rate when it's over here how much is the firing rate and so on and so forth and so that allows us to build up a map of how find rate of the Near and relate. [00:27:42] To the value of another words a tuning curve if you are recording from a neuron in visual cortex What are the basic kind of things that you would want to ask in order to build an intuition or sense of how information is encoded Well for example if you're using oriented bars you would plot firing rate as a function of the orientation of a bar firing rate as a function of the variable of interest in this case in decision making studies before we had this estimate of what a was at each point in time we didn't really have this horizontal axis so nobody could really plot that but once we had that estimate that allowed us to make these to make these plots and really estimate what these tuning curves are in different brain regions. [00:28:27] So that is going to be a tool that has been going to be very useful for us and I'm going to talk about it about the results of using it in a few slides and having all these tools in hand yes. What are the aware of the error bars. [00:28:47] Here they're really tiny because they're averaged over many neurons and they're averaged over many rats so they're actually only a little bit bigger than these ball. Different neurons are quite quite different so I'm going to mention that later particular with any one brain region it is not the case that all the neurons behave the same and I'll. [00:29:10] Come back to that point again any other questions at this point. Not right now that probably means I'm speaking to fasten it just all gobbledygook. Don't hesitate to ask. OK So we have a tool we've convinced at least our set ourselves and a small fraction of people or perhaps a growing fraction of people that rats are accumulating evidence over time as we go so that means that we can study this process we have a tool to assess how information is encoded and neuron that we're recording from and so now we want to go to that question of what are the critical regions that we should study. [00:29:49] So our thinking was that any particular brain region that was causally involved in computing the gradual accumulation of evidence over time should 1st of all carry information about the graded value of the accumulated note is that this accumulator is a greater quantity and so we're going to use our tuning curve measurements as a tool to help us assess that as we record through different regions of the brain 2nd if that region is involved in computing the gradually accumulating evidence well if we silence it that should have an effect on the behavior given that the behavior is based on accumulation and 3rd if we perturb it during any part of this accumulation period that should also have an effect on the behavior so in practice what these 3 considerations boil down to are one computing our tuning curves to form a colon gee we can activate regions pharmacologically and see if it has an effect and 3 up to genetics we can activate regions obstinately to give us time position to look at the difference to look to see whether this whole time period is relevant. [00:30:53] So off we go now we're ready to go into the brain and then we face this difficult problem like nobody has done this in rats before where should we go I don't know the whole brain is fair game. So we did the next next next best thing namely we look to the primary literature to get inspiration from that and there are all these regions that had been recorded from the prime of literature and so we thought well we'll start with brain regions in the rat that appear to have similarities to corresponding brain regions in the present and so we started with 2 regions P.P.C. and the ret this has a lot of similarities in terms of its connectivity and some of its functional properties with prime it P P C and this has a lot of similarities with the prime of F O F and we started the prime at X E S The frontal I feel we call it the frontal orienting field. [00:31:45] In the rodent and we started with these 2 because these were the B. 2 most prominent regions in the prime of literature for evidence accumulation so let's start with P.T.C. And so we put tetrodes in P.P.C. while the rats were performing the task and were recorded from many in Iran and we used our tuning curve tool to try and get an estimate of how firing rates in P.B.C. related to values of the accumulator and what we found was that there was a relatively smooth relationship between these 2 so that means that if you know the value of firing rates in P.P.C. You can read out an estimate at each point in time of what the value of the accumulator is. [00:32:25] Which is essentially the same thing that Mike Chaplin and believe him had proposed all the way back in 1906 we used very different approach and now we're showing it in a species that they have not studied We showed it in rats and we think that we showed it in a more rigorous way but the main message is essentially the same one. [00:32:45] So sometimes I like to say all that work just to show that those guys were right. But then we did something that they had not done which was inactivate P.P.C. while the animals were doing the test just to probe causally what's going on and that gave us a big surprise so here what I'm showing you is the percentage of trials in which the animal went to the right as a function of total number of right minus total number of left clicks that were played on each trial and black is the control so that's if we just regular trials we don't do anything and then we used to maul I gag just to inactivate on one side so Queen activate on the left that's the green curve reactivate on the right that's the red curve if we activate both sides imitating sleet that's the blue the blue curve and essentially you don't see anything so then we double that those we don't see anything so then we could group of those and we don't see anything and then we look to pull those and we didn't see anything I mean thought we're just doing things wrong and so we added a different type of trial that was intermingled with these that was a controlled type trial and we found a big effect in those controlled trials so we know that our museum is working there's just no effect on the P.C. so that rules out the P.P.C. as being a main contributor to this computation maybe the P.C. is part of a broader redundant network. [00:34:06] That could be but it certainly can't be a key central node in this process right. We'll come back to P.P.C. later in the talk and these experiments inspired some very similar experiments in monkeys so Alex hucks group U.T. Austin then using the same task that the monkey researchers had used an activated P.P.C. in the monkeys and found essentially the same result you're active 8 L.A.P.D. there's no. [00:34:37] So now we know that our result is not specific to rats it's a more general. So we moved away from P.C. and then we turned to eff off and one of the 1st things we did there was an activate F O F hoping to see some kind of effect and then we did see a big effect so if you activate the right of the animals go much more to the right if you activate the left foot they tend to go much more to the left so that tells us that the F. A is playing some role in the task it doesn't tell us which role it's playing it just tells us that it's it's in there somewhere. [00:35:10] So then we did electrophysiological reportings in the F O F And here in red I'm showing you the data from the Aqua and think purple the data from the P.C. and what you can see is that the F. one has a much more step like tuning curve. And that was extremely interesting to us in particular because of where the border of the step was and the border of the step was exactly a 0 and I want to remind you a bigger than 0 that means go to the right where there's a less than 0 that means go to the left. [00:35:42] So what this was telling us was that in the end if you know the Valley of Fire it's in the air for you don't have a good estimate of the graded value of the accumulator but you do have a very good estimate of what the sign of the accumulator is. [00:35:55] So that suggested that maybe it was more involved in this process than in this process so it's part of the task it's part of the decision making behavior but it's not part of the gradual accumulation of evidence over time so coming back to even this question here I'm showing you averages over near ants and different neurons can be very different so some neurons in the F. are as smooth as when you're inside the P.C. So that's something to keep in mind another thing to keep in mind is that even in this average this red line I'm talking about it as if it were a perfect square step function but you can see that it's not right there is some great information in there so these electrophysiological results suggested that the effort was more involved in this than this but certainly didn't prove. [00:36:40] So to try and get a better look at that we then turn to up to do next using Hillary Dobson to inactivate and here and in particular we're going to use it to inactivate the F. word unilaterally and for us here the big advantage of using up to generate expose the time resolution you can silence neurons within about 16 milliseconds and when you turn your silencing leads are off activity comes back in about 60 milliseconds so that meant that we could silence difference of the trial and so we did that and I'm showing you the results of that experiment here so on the Y. axis I'm showing you the percentage lateral bias that was caused by inactivating the and the horizontal axis the extent of eternal bar tells you when the airflow was silence so what I'm showing you here is that if we silence the effort before the click started it had essentially no effect. [00:37:32] On the 1st half of the clicks and had no effect if we silence the F.Y.R. once the motion had started it had no effect but if we found during the 2nd half of the clicks just before the scope signal that's when we did see a big effect and so then we did the experiments again but now having the time that we were using So now looking at either of the last $250.00 milliseconds before the go signal or 250 milliseconds before that and what we found was that once again only in the very last few 100 milliseconds before the go signal is when we see an effect of perturbing there. [00:38:10] This is exactly what you'd expect if the A fire was involved in this process but not this process if you perturb during the accumulation process you don't see anything if you prefer when the animals are trying to compute or use the sign of the accumulator that's when you see an effect of perturbing if it so that suggested to us that indeed as the electrophysiology gold results had hinted the effort was more involved in this then this and there's a question. [00:38:37] More. Absolutely I mean you know the clicks are coming in and you're told make a decision that's when you are committing either to go left or right and you make your motor plan to go to the left or right yet so these results tell us. They don't tell us that the F.B.I. is involved in computing the sign of being they tell us that the F. is involved in either computing or representing it or using the scientific So for example if the effort is necessary to make the left versus right orienting Motor plant then this is exactly what you would see. [00:39:14] Yes. So one of the things that we did was calculate curves at each time point and in both the P.C. and the F. one it turned out that different time points were very very similar so that's why I'm showing you. That point we averaged over time and that's what I'm showing you here it's the curve does not stretch it did not have to be that way but that's that's how it turned out. [00:39:43] OK So that told us that the effort was very interesting but again it suggests that there is not part of this accumulation process of this integration over time which is the thing that we were trying to hunt down so then we turned to the Superior to look at this and I won't take you through the data on this a pretty quick list except to say that it looks very similar to what we found in the in the F O F So we think that the Supreme is also more involved in this process than this process and then we turned to the straight him in particular to the interior dorsal straight I'm if you're straight I'm afficionado you will distinguish between straight and person lateral straight in the place that we got to we got to because we were following projections from the into the straight and and it turns out that the projections kind of like land right on the border between there are some you know and they're so lateral and so that's where we went and so that's why we don't call it either Dorsami or dorsolateral all we really know is that it's and tear and dorsal so hence hence the name that we used and we did the same things that we'd done before record there so this was well got to have in the lab and we found a relatively smooth curve kind of like in the P.P.C. So if you know firing rates in the well you can read off at any point in time the value of the graded valley of the accumulator but then unlike the P.P.C. when we did the pharmacology there we found a big effect. [00:41:02] So certainly suggesting that the A.D.'s is is involved in some way in the task and then when we did the up to genetics we found something that at least to us was extremely interesting so here in red I'm sharing again the data from the and once again I'll point out that in the effort only immediately before the go signal do you see anything and in green I'm showing you the data from the straight and from the A.T.'s and what we found there is that in the straight and no matter when during the accumulation time period we perturb we found an effect on the task. [00:41:35] So these are the 3 things that we had set out to hunt for right silencing effects the accumulation of evidence behavior the finding rates represent the graded value of the accumulating evidence at each point in time as it's accumulating and perturbing at any point in time has an effect on the accumulation process so we're very worried when we ruled out the P.C. in the effort we thought well who knows how many years we're going to spend hunting and ruling regions out so it was actually quite a relief to find at least one region that matched our prettier and that we can now say we think that this is at least one node in the circuit of the nodes that are involved in computing this accumulation of evidence over time so extremely well we're very excited about that. [00:42:24] We're not excited about this method. So we're very excited about that. And we're also excited that we think that the F.B.I. and the I.C. are involved in this process because then we think that we've got a set of brain regions that represent the entirety of the process all this stuff and this stuff and we're really interested in of course how does this graded signal get turned into a binary signal the categorical go left versus right signal and we think that the interactions between these might tell us a lot about that so what we're doing now is doing similar recordings in pairs of regions and also simultaneously recording and up to genetics and pairs of regions in this cartoon I had 3 regions similar pain of sleep but really we're doing we're going to enter 2 regions at a time to try and understand how these regions interact with each other both through looking at correlations between the recordings and the effect of perturbing one region and seeing the effect on flying rates on the others. [00:43:25] Now I want to point out we think that the atheist is a node in the circuit that computes accumulation of evidence breaks I did that you know which finally identified one. But we don't necessarily think that it's the only node in that circle. And so to prove that one of the things that we did was inject retrograde beads into the A.D.'s and just look in the rest of cortex ask where do those beads go and we had gotten to this particular part of the strait of the interior Dolphins traded by following projections from the F.O.'s to the straighter and so we weren't at all surprised to see that some of the bees some of these retrograde beads and that up back in the 5 but what we were surprised by was that the strongest signal most of the bees ended up in a region that's interior to the F.L.S. 2 millimeters away which in rat scales is a totally different brain region so this and tear into region is now of huge interest to us we're very excited about that nobody has studied this region with our rights on monkeys in the context of evidence accumulation over time and we're only just starting the experiments there at least I can tell he that minimal in this brain region has an effect on the behavior so we know that it's not going to be like the P.C. but its role is that we don't know that's what we're pushing on and the biggest point we think is that having identified one node that was the key thing because now we can follow the anatomy and unravel the whole circuit right here we're following it upstream but we're also going to follow it downstream to try and understand how the whole circuit works at least that's the plan. [00:45:04] Now I'm going to turn back to the P.P. see that a few slides earlier I said For my clothes to clean when you activate it has no effect forget about it right there put this big red line over it. But it's kind of mystifying because another thing that I told you was that the P.P.C. did have a really good representation in its firing rates of the valley of the accumulator So what is that representation doing there. [00:45:28] And if you look at connections between the P.P.C. and the rest of the circuit Well it's pretty strongly interconnected with the rest of the circuit so what role does it have and that's where this other project that we've been working on gave us some really interesting hints so I'm going to switch to a different a different story and just to tell you where I'm headed the results of this told us that at least in the tests that we're studying P.P.C. media it's long term where by that I mean effects over multiple trials of sensory history the history of the stimuli that the animal has received have an effect on the animal's behavior and they do that through P.P.C. So let me tell you take you through the experiments were found that. [00:46:12] And in this experiment again in each of the trials the rat faces the wall with the 3 nose pokes and it's trying to put its nose into the center poke but in this experiment once it puts its nose into the center we do the fall we 1st play a stimulus a white noise a pink noise stimulus that we call a. [00:46:32] And then the there's a delay that can be quite long we've taken it up to 12 seconds and then we play a 2nd stimulus that we call speech. And then there's a little delay and then it'll turn to tell the animal OK now you're free to go and now what the animal has to do is respond based on which of these 2 was louder the 1st or the 2nd one and and to do that you have to remember the loudness of essay throughout this delay period so that you can compare it to the loudness of a speech. [00:47:01] So it's a working memory task in that sense you're keeping this information in mind and we call it a parametric working memory task because the loudness of the essay is a parameter it's a continuous promise or so that gives us something to plot against which can be extremely useful in the analyses that we do. [00:47:19] Now if as you go through this task if you ask how difficult is each of the trials Well that depends on how close to this diagonal S.P. equals as a line you are right S B is equal to say that it's impossible to say which of the 2 was louder but if you are very far away in this S.B. versus a as a part if you're very far away from this diagonal line then it's very easy to see where the child was and in different trials we use different pairs of S. and P. and S B to cover a big chunk of this stimulus space and these are the kind of the 8 standard types that we used and little number here is telling you the percent correct trials that the animals did in this experiment and you can see that they're doing pretty well right above 80 percent correct on all of them and on some of them quite a lot. [00:48:09] Now as I mentioned how close you are to the stagnant line tells you how difficult the trial is so we also had another subset of trial types that spans this line horizontal e from here to here to give us an estimate of the finesse with which the animals can respond and distinguish these 2 streams. [00:48:28] And so if we plot the fraction of the trial the time that the animals respond by going to the right as a function of S. aim. We get these curves which are like the psychometric plots that are shown in the decision making task before the ideal performer would look like this any time that essay is bigger than S.P. no matter how much even if by very little go to left whereas sorry if essay is less than S.P. you go to the left all the time whereas if S. a is bigger than F. B. you go to the right time this is the ideal performer and rats of course are not ideal like any any observer any biological observer Now one of the things that I didn't tell you is that in this experiment we are particularly interested in the role of P.P.C. and we are interested in the role of P.P.C. because there's a lot of rodent literature talking about P.P.C. and its role in short term memory so we thought well here's a working memory short term memory task it should play a role what we expect is that if we activate P.B.C. this should impair the animals' ability to perform this working memory task. [00:49:33] And so what I hadn't told you is that these data are from animals that had. An optical fiber implanted into them by lead to fibers actually bilaterally in P.V.C. and we would also injected a virus to express hatred option so that we could inactivate P.V.C. and these blue lines are from the subset of randomly interspersed trials in which we had turned we had kept the inactive anything blazer turned off but we could also turn in activities and later on and when we do that what we found was that all of the rats looked more like the ideal performance. [00:50:08] All got better so we're expected that it would disrupt their performance and it's 10 they got better so this was highly mystifying not what were you expected at all. And when Athena crummy the postdoc doing this who was doing this work who now has her own lab in London and is looking for a post office and she's awesome So if you're looking for a postdoc 1st come to me but she's great for you. [00:50:32] So when she did the electrophysiology recordings doing this task Here's an example of a typical nearing that she recorded while animals were from the task and the different shades of brown in the in in these lines tell you the different valleys of this for a stimulus say and what you can see is that during this delay while the animals are remembering as a going to compare it to be there's essentially no difference between these so there's no information in P.P.C. firing rate about the value of SE during this delay interval but as you can't recording into the intertribal interval so the animals make their response around here and they've either received a reward or not by about here and so after that once this information is totally irrelevant the trial is over then there was a lot of information in the firing rate of this near and as to what the value of essay all the way back here you know one trial ago was so that was a huge surprise what the heck is this information doing here the animal doesn't need it anymore. [00:51:32] But that got us into thinking well maybe we should Adeline's our behavior Nell analyze it our data not only in terms of the current trial but also in terms of previous struggles and so we set up a very simple model in which we take the values of essay and S.B. multiply them by parameters add them up also add a bias parameter and put that's realistic to produce a probability of the animal going right that's a super simple model of the behavior and then we add terms that look into history into previous trials previous sensory stimuli previous choices previous rewards to trials back and so on and so forth and there's many different combinations of how you could do this and how many trials back you could go we did a lot of them here I'm showing you the particular formulation that had the best Cross prediction of crossed out of the data. [00:52:19] And if you look at the values of these weights these parameters those are when you fit the data those are a measure of how much each of these parameters affects the animal's behaviors so this is very big That means this have a has a big effect on behavior so if these were 0 that would mean that the sensory stimulus one trial back would have no effect on it and what we found was that sensory history weights were significantly different to 0 so previous sensory history has an effect on behavior and if we take this model and we just predict data from it of course we had fitted to the data so it's no surprise that it predicts they did reasonably well but then we can also fit the model to the subset of trials without inactivating laser was on and when you do that what we found was that the sensory history weights all got significantly reduced whereas all other weights for example for previous correct side are previous award or the current sensory stimulus they were unaffected. [00:53:16] So this is if the P.P.C. was selectively reducing the effect of sensor history on the animals behavior and when we keep these parameter values when we use these yellow parameter values but keep the blue parameter values from all the other weight so we only affect sensory history then we found we could predict that in performance improvement very very well so what's going on what's going on it's it's as if the animals use previous sensory history to guide their behavior but in the lab each trial is totally independent from previous trials so that is a sub optimal strategy to use and we speculate that in the wild of course using past sensory history to guide your behavior is a good idea because there are a long time correlations in the wild but in the lab we artificially eliminated all of those and so the animal is maybe still trying to use them and when we stop it from using them that's when its behavior improves relative to what's being done in the lab Yes. [00:54:19] Right. You. Know so these are you know these are very well trained rats these are say animals that have been doing this experiment stable for 6 months so we don't see any evidence of them having You'd imagine that they could we can train this out of them right because this is suboptimal thing to do to use a sensory history and we also looked at you know once the animals are very well trained look at the 1st month and look 5 months later is there any difference and we haven't found any difference so this particular thing we have not been able to train it out of. [00:55:01] I'll also point out that in the subset of trials that we're using. For the 2nd metric curves getting rid of sensor history predicts an improvement in behavior but there are other trials we're getting rates rid of Hansard history predicts an impairment of behavior and that's exactly what we see what we see so it's not just that inactivation leads to improvement in behavior and activation of P.P.C. leads very specific leads very specifically to behavior that looks as if sensor history effects have been eliminated. [00:55:33] So what I've shown you is that P.P.C. represent sensor history media it's the effects of sensor history and behavior and people are standing up because you've got to go and I've got 3 minutes why why is that result interesting because there's such a large literature on sensor history affects all of bays in perception is about using sensory history to guide how you behave now but nobody had any clue as to which regions of the brain are involved and as in the decision making task now we have a foothold right we don't think the P.C. is everything but we can go from the people where it predicts to what projects to the P.C. and try and unravel that circuit and when we look back at the decision making task at the click accumulation task and now looked at sensory history effects we found those there are 2 So this is the case of the working memory project telling us something about the role of P.P.C. in this other project that we would never have found if we had stayed within this one so people need to leave so I'm going to wrap up merely going to tell you that we're also looking at other parts of the brain. [00:56:33] And I've spoken mostly about looking at effects of whole regions and about behavior at the level of the whole animal but we're of course we're also very interested in how networks of individual neurons drive all these phenomena and so we're interested in working at this scale and spanning the scales from microns individual neurons to millimeters to centimeters and I don't really have a lot of time to tell you about. [00:57:01] I'll just tell you that's where we're trying to go and that we're actively looking for more post talk so if you're interested or more graduate students in this kind of interesting to you please think of Princeton let me end just by thanking the people that all the people that did the work crummy did the working memory experiment being Brunton did the original behavior in the decision making task differ of the content Hanks did the F. I think P.P.C. experiments and make out you had to be straight in experiments in the decision making task and finally I thank all of you for listening thank you. [00:58:05] So of course I'm finding something on a rat causes a change in behavior rate and yes you know the rat wakes up from surgery and suddenly it's got this like 30 grand thing on its head yet it's so little freaky for. And so we spend a lot of time so there are 2 things to say about that one this is true of all recording. [00:58:27] And in fact when we use our regular tetrodes not only does the right have something on its head but during the recordings it's plugged in there's a cable that goes to the amplifiers and having that cable in the field of that cable. The rest it takes about a week for the rats just to get used to that so both of the tetrodes and for the wireless we actually 1st start putting little weights on their heads so they get used to the feeling of something weighing there and then after a while they seem to be fine and they groom and they seem to be happy once they've gotten used to it and all seems to be fine the wireless speak. [00:59:01] If it doesn't have the cable is even better than the wired so that's another reason why we why we like the wireless. So we're also using the functional ultrasound and we now have rats moving around and behaving and they seem to have gotten used to the weight of that too so that's one place where it's useful to use rats rather than mice of course when you're using something heavy. [00:59:37] So he's a he's asked 3 questions already so so when you move from the use of all you know activation. To get more cute faster times time resolution Yeah good direction like you like our conversation really. Reaches or aspirate it could be. Other parts of the circuit we haven't done that yet so we don't know the answer to. [01:00:13] So so we worry a lot about that So for example where was the last quake we know that that that that's not what they're using because that's one of the things that we've checked for so there's a variety of ways that we've checked for this so so here are the 3 things that have convinced us that they really are using accumulation of evidence 1st the model that we used to quantify the behavior I said It has 9 parameters and it turns out that as you said the 9 parameters to different values you can actually implement many different strategies so for example it can also do burst detection and it can also do a mix of bursts of texture and accumulation and can do kind of everything in between so that model alone and come past wide range of strategies and we fit into the data it always does the accumulation strategy so that suggests out of all the things a model could be doing you know it's really accumulation then there's the universe of things that the model doesn't do and so everything that we've been able to think of you know we check that and it's never come out as matching the data better than accumulation and finally the one that really led us to drink our own Kool-Aid was you have an idea of what's going on you make predictions from it predictions that aren't in the original data do they match your data and it turns out the predictions were match really really nicely that if you won the details it's in Figure 3 of the $2013.00 Brunton a town paper. [01:01:33] But what we did was take trials where they had been multiple clicks from each side and use those to fit the model and then from that model fit not predict the behavior on trials where there was only a single click. On one side OK so completely different class of trials and it fit beautifully predicting rat by rat exactly what was going on so that last one seeing that predictions from the model are borne out in the data and that's when we thought OK we think we can believe this. [01:02:15] So much. So. I'm. Just. Like. So we haven't yet but guess that's where we're going a lot of things in my lab seem to be driven by my saying we should go here and I graduate sooner post-doc saying no you're an idiot we should go over here and so my bias opposed to words frontal errors thinking they'd be more comedy but a post like Thomas Lew in the lab said no we really should go to order to cortex in the Imperial eclipse and so he's he's working on that so we'll have an answer for you not quite yet but in progress.