09:08:27 I, thankfully, this is not a rule. Learning meeting, because I have already violated several of Joseph's rules, including I didn't prepare a slide for my 90 s introduction, so I will just ramble on for 90 s, and then I'm also the sovereign of the day, which is funny title.
09:08:44 but I will step in. I'm a poor substitute for Anne Collins, who was originally slated to do this, but I'll give a presentation this morning on how I think about statistical learning and the context of memory research, which I think will probably be a little bit different from what you've talked about so
09:09:01 far, but is related to several people's interests in the room.
09:09:05 So, after my presentation, after a break, we'll have a panel to discuss of various people who've worked on and thought about learning and memory in different systems and in different ways.
09:09:18 And so I set that up for an hour. The final half hour.
09:09:22 I put aside whether we use it for this or not.
09:09:24 For a week, a week in review. Kind of discussion. I know it's only a partial week.
09:09:29 I wasn't here for the rest of the week, although I will be here, I will be here next week as well, but if there are, hopefully the discussion towards the end can veer away from memory into more general, perhaps something a little more integrative across some of the presentations that happen and maybe aaron will
09:09:48 tell us what statistical learning is. Finally, we'll see.
09:09:52 Yeah, I mean by the end of the year I'll tell you.
09:09:56 So, Ok, my 90 s spiel is that I'm a professor in the Department of Psychology at Yale University.
09:10:02 I'm also the director of our new Neuroscience Institute, called the Wusai Institute I'm a cognitive neuroscience scientist by training.
09:10:10 So I study human mind and brain. We do behavioral studies, psychophysics, brain imaging work.
09:10:17 Primarily with functional MRI. So talk a little bit about that today.
09:10:21 We also do some electrophysiology in patients with implanted electrodes, both acutely for seizure, monitoring in the hospital, and also chronically, people with implanted neurostimulators who are out in the world, and then we also
09:10:36 try to encapsulate those data into formal frameworks.
09:10:40 So we do a little bit of computational modeling. I'll mention some neural network modelling work that we've done today.
09:10:47 And so we've studied a range of different topics.
09:10:50 But one of the big areas of focus in my lab is statistical learning, and in the sense that hopefully has become clear by now, because there are many meanings of the word of studying the ability of humans to pick up on patterns through experience which is different from other definitions of statistical learning but we study
09:11:07 that in the context of thinking about memory systems in the brain, and how different brain systems support different kinds of learning and network.
09:11:15 So that's my perspective on this. And I'm gonna tell you all about that today.
09:11:21 Including some recent developmental work that we've been starting.
09:11:25 I think that raises some very interesting questions about statistical learning, and Lauren's going to talk about development more in the future as well.
09:11:33 Last 2 sentences as more social things. I love Southern barbecue food which may or may not mean something to non-americans, but as a Canadian I discovered the idea of smoking ribs and brisket and all kinds of things, so I have a giant smoker in our
09:11:50 driveway that I fire with. You know, Maplewood, that I cut down from our property I'd make all kinds of things, so if we were up near Connecticut I would have a big barbecue for everybody as an excuse to smoke a bunch of food.
09:12:04 That would be the first Monday party. Indeed! Right? I did notice.
09:12:15 There is not far from Munger. There's a Hawaiian barbecue place so don't know what that means.
09:12:21 But it could be good with pardon. Apple. Ok? So another rule I'm going to violate a little bit is I am going to talk a bit about my own work.
09:12:35 But in the context of other people's work, because it's reflective of my perspective on statistical learning.
09:12:41 So, okay, episodic memory is a term that may or may not mean something depending on your background.
09:12:47 What I mean by episodic memory. Here is your ability to remember a specific moment from the past in space and in time.
09:12:54 So, remembering the barbecue last night, remembering the birth of a child, remembering some specific event that happened.
09:13:00 Ok, so this is often considered a kind of one-shot learning. Right? You only have that barbecue once in order to remember it.
09:13:07 And so that's really the core distinction. I'm drawing here between episodic and I'm drawing your ability to form a trace in your brain from a single experience versus to extract regularities across experiences.
09:13:20 And of course that's a spectrum, and there's a lot of overlap. But I'm going to talk a bit about the different computational requirements for those 2 kinds of learning.
09:13:30 I should also say I'm more than happy for this to be highly interactive.
09:13:33 So feel free to interrupt or disagree, or set me straight.
09:13:39 If you're not familiar with the history of research, episodic memory just to give you a bit of a preview and show you something from my own interactions, a lot of the work on episodic memory started in patients who had damage to a particular brainstruct called the
09:14:00 hippocampus, the most famous cases. A guy named that hopefully.
09:14:04 Many of you heard of Henry Moias, and in the fifties the surgeon removed his hippocampus bilaterally, and he became densely amnesia.
09:14:14 Both lost past memories, but also couldn't form so he's a famous case.
09:14:18 There's been books written about him. There's a a Sue.
09:14:23 Corkin is perhaps the most famous researcher who worked with him, and she published a book in 2013, and you could read about it it's in every introductory neuroscience and psychology textbook, but it was really powerful because it showed that this relatively small brain region.
09:14:37 Had this selective, and, you know, critical role in a very specific aspect of cognition, which is your ability to store memory.
09:14:46 I've worked with these kinds of patients. I thought I would show you a video. This is one of them.
09:14:50 Her name's Lonnie Sue Johnson.
09:14:52 She uses her full name rather than her initials.
09:14:54 She's a really fascinating case. She didn't have surgery to hippocampus, but she got a disease called herpes and cephalopodis.
09:15:04 So it's a virus for some unknown reason, tends to attack the hippocampus, so she suffered complete bilateral hippocampal loss.
09:15:10 As a result of this virus. But this didn't happen till our late fifties, and before that she had a very full, interesting wife.
09:15:18 She was a art professional artist. She played the viola and an amateur orchestra.
09:15:24 She owned a dairy farm in upstate New York.
09:15:26 She was a private pilot, so she had a very rich life, and then she suffered this toilitating a disease, and lost the ability to store and retrieve memories.
09:15:36 So I'm gonna show you a video of her. She's highly intelligent, very interactive, very kind.
09:15:41 Funny, but she can't. She has no memory. So this is a video that her sister and her mother recorded of her talking about an autobiographical event from her own life that happened before she suffered this brain damage so hopefully the audio plays people on Zoom may not be able to hear this but I
09:15:58 can say what what a what it says in a moment. So this is her sister and her mother off-a talking with Lonnie Sue.
09:16:05 Do you know what the story is with Daddy? Why, he's not here right now.
09:16:10 What are you doing? Is sick. You remember he died. He died with eidity. Die!
09:16:20 18 years ago, 80. Remember, we tell people.
09:16:30 The last time I saw him was okay, so she can't remember the death of her father, which is obviously something that anybody would remember.
09:16:38 A death of a parent, the reason I'm showing this video is because it illustrates 2 aspects of amnesia that are relevant.
09:16:45 One is that in an ability to remember something from the past, from before her injury in this case or her disease, so he died before she got sick.
09:16:56 The other is that it illustrates what's called anterior grade amnesia, and an ability to form new memories.
09:17:01 So the reason they recorded this is by their count. This is about the one hundredth time after she recovered, that she was told that her dad was dead, so she also, in addition, not remembering the event itself, couldn't store the memory of being told that a hundred times so this yeah, so it seems like this video was cut. At a moment.
09:17:20 And she was going to see that the last time I saw him, and so it's something like that was a memory yeah. So I've seen the rest of it.
09:17:28 She just kind of stammers on so she can't live there.
09:17:32 Isn't that? No, yeah, no, no, she doesn't remember.
09:17:36 In fact, what's really interesting is this is the hundredth time I've seen them talk with her about their dad before.
09:17:43 When earlier, this and later than this, and her emotional response has become more muted over time, you can almost see it here.
09:17:48 She doesn't react like she's horrified, like what happened to Daddy.
09:17:52 So there is some learning going on it's just he can't remember the specific specific event.
09:17:59 Yeah. She never responds things. No, yeah. It's a great question.
09:18:07 So in the clot. In the classic cases, like there's something called graded retrograde amnesia, where he was unable to access stuff that he had encoded recently with the last 5 or 10 years.
09:18:18 But he did have older memories. She has complete retrograde vision.
09:18:23 She has no memories from her life. What's that?
09:18:25 We know why there are. Yeah. So it turned out, only had about 75% of his hippocampus removed.
09:18:31 So I was going to ask, yeah. But, on the other hand, get a lot of other stuff from the Npr.
09:18:37 Removed. Yeah. So it was going to ask, like, How does this show everything of her legion compared to that of natural lesions?
09:18:45 Are messier than the surgical lesions for sure, and so it's a little hard to know whether her complete retrograde amused as a result in medial temporal damage.
09:18:53 That's more extensive. And he did have more. But it was, you know, clean cuts, whereas the this is her brain. So this is the hippocampus.
09:19:03 But you can see on one hemisphere there's lateral temporal damage, and when we get to the front you'll see there's a tiny bit of infere frontal damage.
09:19:09 So it's very hard from a single case to draw any specific conclusion, but other than across many cases.
09:19:14 With this kind of hippocampal damage, you get the same kind of pattern vehicles.
09:19:21 Yes, and that's not unusual in any sort of yes, sure.
09:19:26 Yeah, I mean, it's a brainwide disease. It.
09:19:29 Just yeah, yeah. But she's, you know, we tested her on a bunch of other things.
09:19:34 And so the neuroscience testing is relatively celestial in language, deficit.
09:19:39 She doesn't have other you know. Generic, cognitive problems are, you know, reasoning, abstract reasoning.
09:19:43 Those things are fine. It's a single case. I don't want to make a big deal about one case, but yes, but she still remembers she had it back.
09:19:51 She does. She has semantic memory, so she knows that people have parents, she actually knows the names of her sister and her mother, her mother's deceased.
09:19:59 Now she doesn't know me. I met her I don't know 50, 60 times.
09:20:05 And maybe I'm just forgetting I don't know right.
09:20:11 So she was an artist. So she actually also can recognize paintings from famous arts.
09:20:15 She can sight, read music on her, viewore. It's not like she has no learning abilities.
09:20:23 They're just much slower. I think that's the easiest way to say it so when she moved into it, assisted living facility, she would go swimming every day.
09:20:29 There's a gym in the facility, and to get from her room to the gym requires 3 turns, and it took her about a year to figure out how to get to the gym on her own.
09:20:38 But she did right for you or me. We could do it in one or 2 times we might get lost the first or second time.
09:20:44 So I think of the hippocampus as a really fast learning system that's why you can do things like encode single experiences.
09:20:52 Yeah, just wanted to know that. I suppose you run some experiments.
09:20:56 Yes, I'll show you a bit of data later. I don't want to jump in.
09:21:03 They're only so, what about the inclusive knowledge?
09:21:03 Yeah. Yeah. So she can't do statistical work.
09:21:08 I'll avoid the word implicit, though. Yes.
09:21:16 Absolutely. Yeah, yeah. So there are many other patients. Now, it turns out, patients who have complete bilateral hippocampal loss have total retrograde amusion.
09:21:24 This is where, like Lynn Malmor's Mosquevich, others have shown that some of the original claims about graded retrograde amnesia that only applies for partial hippocampal damage that's the current thinking.
09:21:35 At least.
09:21:37 Okay? Well, that's that's a very interactive discussion of a video.
09:21:41 Yes. Good. Do you? Other types of brain damage or lesions result to food after sugar that Asia?
09:21:50 Or it's only for cameras. Oh, I'm sure that there would be other.
09:21:52 There are other critical steps in reporting memories. So if like, if you couldn't speak or something, you wouldn't be able to report on.
09:22:03 That's a, you know, tried example. But no, that's more hippocampal.
09:22:12 Yeah, some temporal lobe damage can lead to other kinds of memory deficits.
09:22:16 But that kind of memory of being able to mentally time travel back to a moment in the past.
09:22:21 That's 101 hippocampus.
09:22:23 Yeah, sorry that's yeah. The question. But remember, it has normal studies, I think, like Zola and Squire a okay.
09:22:34 Animals some simple task and learning task Asian hippocampus the Gays are rodents, I think, and they basically just loss of memory or something.
09:22:46 This really is that, do you have any? I'm not experts that I hear anybody.
09:22:52 It was at 1 point, but I forget it. How about how these new findings and others relate to the older yeah, it's a great question.
09:23:01 In those tasks you learn things at different distances from the surgery right?
09:23:09 And so it's to test graded, retrograde amnesia.
09:23:12 And they find something like greater retrograde Amesia studies the critical thing is the type of memory.
09:23:15 They are studying is kind of short-term object, memory, and so it's like a delayed match to sample kind of task, and they're trained on the same objects, you know, dozens of times so it's a really hard literature to reconcile with human
09:23:33 episodic right? So that's I don't have a good explanation for that other than to say, I think it's a poor proxy for its kinds of episodic memory studying humans.
09:23:43 Yeah. Almost all the animal work on amnesia is with overtrains.
09:23:48 What's life? Kind of in humans with change it from being but then there's the randomness I can bound.
09:23:54 Yep, to actually do this in that is more. Okay. Yeah.
09:24:02 I come. Mom, is the best version of this. Yeah, the hour.
09:24:05 C groups that's the best work of that touch. Absolutely.
09:24:08 Yeah, yeah. But even in humans, wrecked item, recognition is not hippocampal, sir.
09:24:15 So one is you can do item recognition. So if you show a bunch of objects in the test, remember for individual objects, you can do that even if you guys think the recollection contributed no, just the familiarity if you don't just disassociate that.
09:24:29 With oursc, okay, right? So, I'm going to tell you a little bit about memory systems in the context of statistical learning, although we've already started that I'm gonna talk a little bit about our developmental work.
09:24:41 I think, as a quick preview, statistic learning as Lisa was studied in humans originally was in infants, it was thought of as a really fundamental learning mechanism for requiring language and names of objects and learning about your environment, and so on and there's still a lot of work and development.
09:25:00 we've been studying in adults and understanding brain systems.
09:25:02 And there's relatively little to no work on the brain systems and infants that support statistical learning.
09:25:07 So that's something that we've been working on and talking a little about, sort of new thinking about memory systems which we've already started discussing, and a set of findings to sort of change assumptions about the mapping between brain systems and different aspects of cognition and then just
09:25:27 the discussion at the end, so in terms of systems, this is hopefully familiar to most people, but if not, I'll go through it when I'm talking about memory here.
09:25:36 I'm referring to long-term memory, and so that's different from short-term memory.
09:25:42 If you're holding something in mind actively, that would be a kind of short-term memory or working memory, typically that needs to be actively maintained, whereas with long-term memory, what I refer to as sort of archiving something doing something else, and then going back to access it later on okay, so it's kind, of more
09:25:58 silent memory that can span across interruptions over time, and so on.
09:26:03 Yeah, when does long term memory? I mean, experimentally, it's on the order of minutes to years, whereas working memory is typically on the order of milliseconds to seconds.
09:26:17 That's kind of the time. Scale difference. Well, I mean, none of these are absolutes, but the brain systems required for short term memory are different.
09:26:26 So Monty Sue is totally fine, short-term memory.
09:26:28 Tell her to, you know. Hold some phone number in memory. She can do that fine it's like she's actively maintaining it.
09:26:36 If you come in and say, Hey, Lonnie, Sue, you interrupt her and distract her.
09:26:38 Then it's gone. So short term. Memory is like active maintenance.
09:26:42 That's the easiest way to distinguish it. Whereas so you could in a computer metaphor, you could think about RAM being short-term memory and like a hard drive being on. From. That's another way of thinking.
09:26:57 So long-term memory. This is a longer time scale memory.
09:27:01 Traditionally it was divided into declarative and non declarative.
09:27:05 So this is an old framework from Squire and others, but it's carried lot of sway.
09:27:11 I think this is the standard view in the world about these things.
09:27:16 So what you'll notice, and I'm glad justice is not here, because he might agree with this.
09:27:19 There's a distinction between conscious memory and unconscious memory explicit memory, implicit memory.
09:27:27 So that's what declarative versus non-declarative refers to your ability to consciously report on a memory versus not.
09:27:33 And so the sort of declarative, explicit memory is mainly what I'm going to talk about in the context of episodic memory.
09:27:42 So memory for an event like for the barbecue last night would be an example of a declarative, explicit, conscious memory, and that in this fuse, dependent on the medial temporal lobe, and especially on the hippocampus so that's this kind of Monty sue
09:27:57 example, and then a non declarative forms are sometimes, when people think they're studying in the context of statistical learning.
09:28:08 So, for example, many people in this room have done work on something that's more like procedural learning.
09:28:13 So you could think about 0 reaction time task learning. If that that means something to you.
09:28:17 So motor sequence learning as an example of procedural learning.
09:28:21 It's described as skills in habits. So these are things you do.
09:28:24 But that might be kind of hard to describe, like you know how to ride a bike.
09:28:29 But you can't really report how you ride a bike.
09:28:33 And then each of these different types of non-declaredive memory depend on a in this framework, on a specific system.
09:28:42 So the procedural memory depends on the stratum right, and I don't know.
09:28:46 Certainly! Motor conditioning depends on the cerebellum, and so on.
09:28:51 Priming here refers to your facilitated sensory processing of a repeated stimulus so if you see the same picture twice in a row, you'll be faster, better better recognising.
09:29:02 So these reflect different kinds of memory. This view is a very simplified view.
09:29:08 This is a one to one mapping between behaviors and brain systems.
09:29:12 And this really came out of that lesion literature because you could find people who were missing a hippocampus and had a deficit only in this, and he could find people with damage to visual cortex who only had a deficit in this or people would damage the
09:29:26 Basel ganglia, which only had a deficit in that.
09:29:29 The problem is that that doesn't that that's kind of an oversimplification because of the reliance on lesions.
09:29:35 So it is the case that these are critical structures for each of these kinds of memory but it's not a one to one mapping.
09:29:43 And so that's what I'll get to later on is it's actually a many to many mapping that's a lot more interesting.
09:29:46 Yeah, in the example, the patient that you just showed up. She remembers fans right.
09:29:55 Yeah, that's right. So this is, you know, in the sort of tolling literature this would be like episodic versus semantic memory would be events versus facts.
09:30:06 Semantic memory can be fine without hippocampus, although it's still dependent on the temporal.
09:30:15 With a lot of with a lot of training. Yes.
09:30:20 This is primarily coming from lesion studies which can be.
09:30:25 Correct. So we look at what people have tried to review the politics from animals.
09:30:35 Affect fine, motor skills. But then they find that I'm mostly that pretty easy.
09:30:39 So yes, how does this apply for this whole memory system? The case for you to have damage?
09:30:48 Yeah. Show. The third of Ntl, Dad is at 2 years old.
09:30:56 The foundation of justice. Tool. Yeah, that's a great question.
09:31:00 The short answer is for things like hippocampus.
09:31:04 They'd be screwed. There's no substitute for the hippocampus for cortical damage.
09:31:07 They can recover, so it's really beautiful work. By what was Newport and others in Perinatal strokes.
09:31:11 Of babies who are born missing entire hemisphere.
09:31:15 I can still acquire language, and still develop, and sometimes don't even know that they're missing hemisphere by adulthood.
09:31:22 So the earlier the damage the better, the more that the damage was cortical versus on the specific subcortical circuits, the more they can recover.
09:31:30 So there has been some developmental work on kids who are born with oxygen deprivation during delivery.
09:31:39 So epoxia, and they tend to suffer like relatively small.
09:31:43 You needle temporal lobeesions and also have behavioral deficits as they grow up in episodic memory. Be a general.
09:31:52 The earlier the more recovery and just ask that as well. Sometimes the deficits that you see killing the sup portable and on cortical regions, and even a little bit unintuitive.
09:32:03 So one finding that about it for me is, it looks at Perinatal stroke, and the locations of the parental stroke, and how it is associated with visual at the I believe that what they hypothesized is that if you have a stroke in the
09:32:19 eccentric. That was not what actually strokes.
09:32:25 And you know I wouldn't. Why is that? You see that?
09:32:30 So she's not as a single study, but you know there are circumstances.
09:32:34 It's just kind of when you think about development. We have a particular relationship in adulthood and maturity between cortic religions or.
09:32:45 It can be. Yeah, I think here and then here.
09:32:56 Yeah, minus second rule is still 0 years. Austria.
09:33:07 Basically functional I'm in screw for memory. Good.
09:33:08 Yeah, this is actually funny enough. This is a question that came up with my dissertation, because that's talking about comparable any memory systems this morning and whatnot.
09:33:19 And they were like, Well, have can you explain that that in bits you don't have, you know, highilaterals they can't buy, you know.
09:33:26 That's since nations learned language, and I was like.
09:33:31 Also mean, I'm not aware of a single case of a baby born with complete bilateral picample loss, because to have that kind of damage during birth.
09:33:39 You probably wouldn't survive. It's very rare.
09:33:41 That's why patients like Lonnie Sers are rare to have that extent of hippocampal damage, but not have really broad or fatal damages.
09:33:48 Rare, severe, encephalitic cases that lead to complete bilateral hippocampal loss.
09:33:54 A third of the time lead to death. A third of the time lead to selective damage like this, and the other third of a time we did broad damage.
09:34:02 So the cases that exist in infancy are pretty restricted damage, where I would expect some residual memory function.
09:34:09 What? Just maybe momentum. But you know what we mean by learning language can be incredible.
09:34:14 Right so to say, that someone has, like a capacity, you know, you really want to probe that a lot of right.
09:34:24 And so a lot of time, just a lot of times. So at such a high level, it's sort of hard from a neurotic perspective.
09:34:32 These are these corroborate at all. But your point is fair.
09:34:39 That it's mailing memory deficit. Yeah, yeah. But that's one of the only papers, and that's very limited damage.
09:34:43 To be honest. Yeah, yeah, master. So as a disclaimer, I think whatever criticisms me have about this particular thing in this particular classification diagram, I think.
09:34:57 It's still useful. So now the first approximation, it's right.
09:35:04 Yeah. So in this case, I would like to raise another criticism that I'd be missing because really, this side that makes 2 claims right.
09:35:13 And so far the discussion has been centering on one of those, which is, How do we pair these memory systems to biological substrates?
09:35:20 Right to brain awareness. And sure, I think that it's going to be hard to draw a bottom.
09:35:26 Okay. Awesome.
09:35:30 Classification of memory systems, that that that is relevant.
09:35:35 I as a feminist, I have the biggest problem with that part, because, you know, for one, it's not clear whether these are really different versions or just different paradigms that I agree with happen to develop to test.
09:35:47 Memory, is that really a memory system? Yeah. So I'm going to come back to that in the third part, because I think you know, these are tabs.
09:36:01 Essentially, these are constructs. They're not. Yeah. They're not computation.
09:36:03 So the kind of memory that I tend to care a lot about, which is kind of general knowledge about how the world works, which humans as well as animals, will have.
09:36:13 And yes, humans can maybe part of it, partly partially explain it in Indones it's a partially high basis.
09:36:21 A lot of our knowledge about the world isn't explicitly accessible.
09:36:27 And that's definitely going to be true for and where is that?
09:36:31 Where is my knowledge that houses have windows and doors even, I mean yes, I can explicitly explain them.
09:36:37 But there's a lot of knowledge like well, and even if you could explain it when you use it, you don't walk around you don't walk up to that door. Say, oh, this is a door.
09:36:47 Okay. So I remember, I need to push this you're preaching the choir.
09:36:54 What I'll say the first part of talk I'm going to talk about.
09:36:56 The biological systems. In the third part, I'm going to talk about this issue is, you know, let's start from these brain systems.
09:37:05 Ask, what kind of computations they support, and how that relates to different aspects of behavior.
09:37:08 That's what I mean by many to one is the same stuff.
09:37:12 The whole memory systems issues. I don't personally believe that there are boundaries.
09:37:15 I don't think these are the right boundaries. I think there is maybe one more.
09:37:19 Maybe I'll press ahead. This is what I want to ask the version about this. So what I'm going to share, I mean, I wouldn't normally do this as didactically.
09:37:32 But I didn't just want to talk about my own work, and so I will walk through.
09:37:35 You know, one theoretical account of how statistical learning has been treated in the memory.
09:37:40 Literature, so this is what's called a complimentary learning systems. Framework. J.
09:37:45 Mcclellan, and many chemical Norman, Randy, O'reilly, and many others have develop these ideas.
09:37:51 And basically they propose that there's a fundamental distinction between 2 kinds of learning.
09:37:55 So it's not all those memory systems, but 2 kinds of learning, I will say, for people who are more computational oriented.
09:38:02 These theories are implemented in relatively simple neural network models.
09:38:05 So you can actually play with them. So the basic distinction is that the hippocampus is important for rapid learning of specific events and the key words that are rapid and specific.
09:38:15 Ok, and the idea is that even 2 related memories will be stored in non-overlapping units or populations in the, and that this is intended to try to reduce interference between memories right?
09:38:31 So if I wanna remember you talk on Monday next week, separately from a talk on Tuesday, even though it's in the same room.
09:38:41 There's the same people in the audience maybe there's similar questions in the hippocampus.
09:38:45 It's important to store those traces separately, so that when you go to try to remember what happened on Monday, get interference and report what happened on Tuesday.
09:38:53 Okay, so that's one type of learning in this. In this model, the other is a slower or more gradual kind of learning of generalities rather than specifics, generalities.
09:39:06 And this is viewed as opposed to the hippocampus, as being cortical.
09:39:09 Broadly defined. So in these papers they talk about neocortex versus hippocampus, and the neocortex.
09:39:17 So this is association, cortex. However, you want to refer to it.
09:39:20 The idea is that related experiments are stored in over the idea is that related experiences are stored in overlapping ways rather than so.
09:39:27 In addition to remembering specific presentations next week, I might want to learn about your personalities, which requires over time extracting what kinds of questions you ask and what food you eat, and that kind of thing, or am I might want to figure out where people sit in the room over days in order to extract those
09:39:45 commonalities, that theory, at least says you want to store this experiences in an overlapping way, so that the common features of those experiences get reinforced.
09:39:53 So what you end up remembering is not a specific event, but rather the common features or the regularities across those events.
09:40:02 So this is not my theory. This has been around a long time.
09:40:07 Magnum opus. On this theory is from 1,995. It's an amazing paper.
09:40:12 The Professor. Simulations. I encourage you to read it if you haven't read it.
09:40:15 Yes, spatial memory, so episodic memory one should learning, rapid learning.
09:40:21 But special memory. It's some people believe that.
09:40:27 What it does. Yeah, well, space is a really important part of episodic memory.
09:40:34 But I think one of the the open questions is in the field is a lot of animal work is using overtly spatial tasks.
09:40:39 Or navigation tasks, and there's not agreement about the relationship between spatial coding place cells, head direction cells, grid cells and so on.
09:40:49 These are fundamental spatial features that are encoded in these systems, and that these more general episodic memories, of memory abilities in humans.
09:40:59 I think of it, that spatial information as a fundamental part of the context in which you experience advanced most episodic memories are grounded in space that's like the simple way I would think about this, that that kind of spatial representation is really important for forming specific episodic memories.
09:41:18 But there's not a good cross species. Resolution of the relationship between space and episodic memory.
09:41:23 There are people who believe that a human episodeic memory is entirely spatial, like well, there's no advance of that space.
09:41:39 So I think we can ignore or time. And so that's the thing is that people focus on the spatial component.
09:41:43 But now people are starting to notice. Oh, there's a temple component.
09:41:46 And they did in the wrong literature, too. I combined the beautiful work on that from from this I would conclude that your impassion or your evolutionists, like hippocampus, is going one shot rapid learning.
09:42:01 That's what this theory, that's the role of the hippocampus that this theory focuses on I don't think it excludes other functions of the hippocampus okay, so that's yeah.
09:42:11 I think other people would say that they should also pass facial memories.
09:42:14 Not one shot station you can have spatial memories that are one shot you can have facial memories that are not.
09:42:20 You'll see that I'm going to end up where you're going.
09:42:25 But sorry. Can every space without memory? I think maybe not, but I haven't thought about it. Could you?
09:42:30 Would you need to have space if hippocampus is only encoding space in the context of memory, which is what we're putting forward as well.
09:42:39 Ok, it's hard to know. You need to have space elsewhere if it but I can't think of a way that you would navigate space without memory.
09:42:46 Lots of different spaces, but you can over learn one space right? But that might still be the second one where you're looking at things sort of your integration.
09:42:58 So there are cortical representations of space, for sure recipiental cortex posterior singular pecunias, parapitamal cortex absolutely in addition, in the hippocampus, there are variables that track spatial features that are not specific to memories like
09:43:14 for example, there are cells in the bi campus that track.
09:43:16 How fast you're moving, and that generalises across contexts.
09:43:19 Spatial representations of the hippocampus of places, though, are specific.
09:43:25 So place cells in the hippocampus in this room would have a different field than that same cell on a different so it's hard to say, I think your story memories all the time.
09:43:35 So I don't want to say that, like those spatial representations and the P.
09:43:39 Gets aren't part of the input for encoding episodic memories all the time.
09:43:42 It's a hard question I'm having a few people from there. But I think I've just also codes for lots of nonstatial things as well.
09:43:49 It's not just to these phases, at least in humans and monkeys.
09:43:53 Odor and time. Yeah, there we I completely agree. Yeah, absolutely.
09:44:07 I agree with all this, I'm mainly representing these older perspectives to get everybody on the same page.
09:44:14 But so just give you a concrete example. This is a spatial example.
09:44:16 This is my prawled my face every morning parking in the park in my parking lot.
09:44:21 The idea is that you want to be able to store where I parked yesterday separately from where I park today.
09:44:28 This is like a canonical example of an episodic memory.
09:44:30 However, think about these experiences. This isn't trivial, these look like they're far apart, but I'm in the same car and listening to the same radio station.
09:44:39 I have my same anxiety. It's the same time of day, I might be looking at the same landmarks right?
09:44:46 There might be the other same cars around me. I'm going to the same intersections I'm making the same terms.
09:44:51 The actual extended events leading to where I park in the Pry model are like 99% identical.
09:44:56 And just at the end I park in a different place. There's nice evidence over the last few years that if you look at patterns of activity in the hippocampus with Fmri, these 2 experiences over time that are 99% identical will be stored in an orthogonal way in terms of the
09:45:10 similarity, the patterns of activity evoked in my hippocampus sometimes.
09:45:15 So it. This is the core idea of what the hippocampus does.
09:45:20 Is this taking related inputs and storing separate memories.
09:45:25 It's often called pattern separation. This is studied in animals and studied in humans.
09:45:29 So that's the core computation, or one of the core computations.
09:45:34 The campus. That's important for storing related experiences separately is to avoid interference in memory.
09:45:39 Search. That's the basic idea about this. Yeah, so this makes sense retrieval, like, wait after you push your car.
09:45:47 It makes sense. All these memories completely open away. Goes to remember where your car is.
09:45:53 However, during that experience, if you come to remember ways, I put my foot in the car.
09:46:01 Yeah, whether whether I need my keys now. It doesn't make sense to school that memory right?
09:46:06 The getting hurt out right. So I think all of these 2 different kinds of learning and others are all happening in parallel, and they all guide behavior.
09:46:15 So this gives you access to, you know, separate representations of the location, but you might also have general knowledge about where you keep your phone.
09:46:23 And so on. So don't think this precludes extracting those kinds of patterns.
09:46:28 In fact, that's the sort of second piece of this.
09:46:31 This is what I think of a statistical learning like in this parking case.
09:46:34 It's kind of funny. But for the example you gave like across experiences, if I want to predict what's going to happen in the future, I actually may not want to rely on the specific details of a past experience, right?
09:46:46 If I'm going to park in this lot, I may not want to go back to the exact spot I parked in yesterday, but rather pick up on a kind of spatial distribution of where there tend to be open parking spaces so if your goal is sort of future oriented adaptive behavior
09:46:59 it's often useful. To aggregate across noisy experiences.
09:47:03 And so I think, that could give rise to something about knowledge about where you store things in your car or routes that are best to take out of the parking lot or other aspects.
09:47:13 And the idea is that these are stored also the traditional view is that this is stored in cortex I'm going to talk about my thought that this is stored in the hippocampus.
09:47:22 This kind of more general knowledge, so I'll go through that a bit more.
09:47:26 But the idea here is that because your aggregating across experiences the learning needs to be more gradual, to avoid catastrophic interference.
09:47:33 So you have a really fast learning rate, and your overlying memories are going to wipe out earlier.
09:47:37 So this has to be a slower learning process, and instead of separating things like you said, you wanna integrate them.
09:47:44 So this requires more overlapping representations. I'm gonna go through a little bit more about sort of how people think about this in the context of the hippocampus but that's the basic complimentary learning systems idea that there's this dissociation between hippocampus does
09:47:59 pattern separated episodic incomeoding and the cortex does this more gradual integrative learning.
09:48:06 Okay, I'm gonna skip over this for the sake of time.
09:48:09 I'm just going to show you. I was just going to say that what I mean by statistical learning here is not limited to the tasks that people have used in the cognitive psychology literature to study statistical learning hopefully aaron introduced what those are already so I don't need to say
09:48:23 it again. But obviously I think everybody in this room thinks of these questions more broadly than the kind of narrow tasks that have been used before, even though I'm going to show you data from some pretty narrow tasks.
09:48:33 So when we did when we started doing statistical learning studies which in the complementary learning systems framework is a cortical process.
09:48:42 Your ability to pick up on regularities across experiences. We started finding hippocampal involvement.
09:48:49 So this is at odds with the complimary learning system, too.
09:48:53 I'm not going to go through any of these studies in detail, but if you look in the kind of early Fmri literature where people were doing statistical learning experiments during functional MRI, they would often find learning related signals in the hippocampus that are related to
09:49:11 extracting, regularities. So these were studies I did. This is like following up these studies actually have a similar design to Joseph's work.
09:49:19 Early work, that 2,002 papers, Jp. The Lmc. Paper, 2,002.
09:49:26 So sequential statistical learning between transition probabilities between shapes.
09:49:31 So that's what these kinds of tasks were.
09:49:33 And we found that if you're exposed to a sequence that has regularities in which shapes tend to follow each other, that led to increased hippocampal activation compared to sequences that were random these are a bunch of other tasks, I'm not going to go through them.
09:49:45 But just if you're familiar, this is artificial grammar, learning.
09:49:47 So this is picking up on kind of finite state grammar that produces strings of characters.
09:49:53 This is a contextual, queuing task. We're doing visual search, and there's regular patterns and where to look for things.
09:50:00 Some of these other ones are cereal reaction time sort of more motor sequence, learning tasks.
09:50:05 I think maybe these guys, this is more of an associative learning type, sequential associative learning task.
09:50:12 And this is a more general study of looking at mutual information and sequences.
09:50:18 These are all cases where their regularities, embedded in a sequence or in a spatial configuration that lead to increased hippocampal activity.
09:50:24 Yeah. Certainly have no debate on this I think there's a large number of studies that where you see hippocampal activation and task, so that I think would agree are spilling.
09:50:36 Take tasks. I just want to commentary for you to agree with or not agree with.
09:50:42 I think there's also you. First of all, school needs many different things.
09:50:44 Right, you know, for table. And so it all you. They also have found basic angle activation.
09:50:50 And that was more for the frequency of which which items were being presented, but also seemed very relevant to this exploring aspect of it. Yeah, I agree.
09:50:59 And it's some tasks where there's image.
09:51:01 You don't see any of the Campbell activation right?
09:51:02 So with like adjacent learning, not choice. There's many different areas.
09:51:09 Yes, certainly, I don't disagree that I would.
09:51:13 I would say that there are certainly other brain regions that contribute this, and you see cortical effects mainly I'm emphasizing this because this is not predicted by these sort of classic memory theories.
09:51:27 Yes.
09:51:31 Yeah, I think that's more noise enough fromri data than anything else.
09:51:35 But we tend to find stronger effects, and these are radiological convention says the right hemisphere tend to find bigger effects than right hippocampus and left hippocampus.
09:51:43 I'll show you more data on that later. So there are what depends on error.
09:51:49 What you believe. Why is it?
09:52:05 Sorry. It's just a synthesis. Well, these aren't tasks.
09:52:06 Worry, that when I'm saying that I'm saying, Ok, Lonnie Sue, I want you to remember this is so and so.
09:52:12 Artist. I'm trying to teach it. These are not.
09:52:14 These are incidal learning tasks. These are in some cases implicit learning, task where we have a cover task and there are embedded regularities without any intention, without any awareness.
09:52:23 So that's the unusual part of this. Okay for tax I don't disagree with you, but that's certainly not the view in the literature, because if you have the kind of pattern separation I was talking about repeated experiences over which he could extract regularities restored
09:52:40 separately. I don't disagree with you, but that's the view, at least in kind, of coming from Era, where is really a dominant model.
09:52:51 The idea is that implicit learning was intact in, you know I don't know what your take on.
09:52:55 It was going in this era was the idea that Cisco Lynn was falling much more in linuge of those sorts of tasks right where it's implicit.
09:53:03 You're doing some kind of like motor rotation tactics which is outside of conscious awareness.
09:53:06 And that would be intact. This is a very canonical, implicit learning test.
09:53:13 If you go to that tree you would put it in a different place okay, I'll keep pressing forward.
09:53:20 Yes, I'll try to keep Christmas. I was hoping there would be enough controversy.
09:53:28 This is good, so if, rather than representing the campus perspective, yeah, which reaches must have, it's, it's a basic angle of another Virginia.
09:53:44 Yeah, what would you have? Well, I don't have time to give that talk, but I could give that talk.
09:53:50 And there are people basically getting music stratum is a key structure for a lot of sequence.
09:53:54 Learning tasks and people in this room have done that kind of work also sensory systems.
09:54:00 So you see effects of learning, visual regularities in visual cortex auditory regulars, and auditory cortex, frontal regions, especially when you're talking about speech sequences are involved in learning patterns.
09:54:10 So there are other systems that drive this as well. Yeah, okay, I just want to give.
09:54:17 So this is related to your comment. So I came at this.
09:54:19 I didn't start studying hippocampus.
09:54:22 In fact, I was more on the vision science side of the world where people think about higherarchies of features, and one of the really striking things in studying vision is that people sort of stop stop thinking about the visual system adding for your temporal cortex.
09:54:37 But there's really strong projections into the hippocampus.
09:54:39 So if you look at the cert classic, you know, film and vanesscent diagram, the very sort of top of the ventral stream is the hippocampus.
09:54:47 And so I've been thinking about this all long as the hippocampus as sort of an extension in the invariance that builds across the visual processing through.
09:54:58 I'm not the only person thinks there's several other groups.
09:55:00 Morgan Brands, and many others who think about this idea that really the medal temporal cortex, especially but perhaps also the hippocampus, have visual functions.
09:55:10 The key difference is the invariance that builds up as you go from early visual cortex.
09:55:15 So edges and contours up to object identity and category.
09:55:19 The key transition into the epicentus is that the coding visual coding in the Epicampus?
09:55:24 Our hypothesis was. It's not based on visual features.
09:55:27 Or abstract visual features. It's based on spatial-temporal features.
09:55:32 So now I'm not giving the memory perspective from a vision perspective, you know.
09:55:37 If you, if I showed you 2 people's faces you might get a similar pattern of activity in fusion form gyrs from your temporal cortex.
09:55:44 If I showed you a different cups or containers versus different pieces of furniture, they would be stored differently, and similar to other members of the same category.
09:55:55 My thinking about this is that if you look at these medial temporal regions, what determines whether 2 objects or 2 items or 2 environments are coded similarly differently, is based on whether they co occur in space and time or not so if I'm always in my office the pattern of activity
09:56:13 evoked by a picture of me versus a picture of my office will be pretty similar, whereas a picture of my face in a different room would be less similar, or instead of 2 copies, cups being similar in the hippocampus.
09:56:28 It would be this coffee cup may be on on my desk if you always see a coffee cup on my desk, so not the identity of objects in terms of how we think about invariant object representations, but rather it's based on cooccurrent statistics, between
09:56:42 objects. So it's not that this is cylindrical, and this is a surface.
09:56:46 It's rather that these 2 things always appear next to each other.
09:56:48 And so if you take invariant identity of objects and spatial environment, the hippocampus binds those 2 together through repeated Co.
09:56:58 Occurrence. That that's the sort of way that people think about this in the vision side of things.
09:57:03 Yeah, basically saying that the gaps. But then, is this something that is for the campus?
09:57:11 I'm going to break down different circuits in a minute.
09:57:14 Yes, yes, context. But also, if this was like how I can bomb, or you know, Cohen or others, they would talk about relations, and that's separate from context or context.
09:57:26 You can think of us are the set of things in the environment.
09:57:29 But really can be sort of direct association between 2 objects, or between an object in a and a place.
09:57:36 So relations are this, or building blocks of a context so I'll talk a bit more about that new Joseph.
09:57:45 And then it seems like more emphasizing state space and time, and objects to what and what their core prices happen, because nobody would say that co-crease happens only right? Right? Sure.
09:57:59 So there it's rather than contexts which you can define in contexts and energy, different scales that the ingredients that you're using look, that's right.
09:58:10 And so right. So the only unique thing up here is that you can do that across things that have no visual similarity.
09:58:14 You get, you know, learning of categories here by linking together things that have similar visual features, right?
09:58:24 And so at like higher levels of it, you can get really generic at it, like furniture or vehicles.
09:58:30 But as you go farther and farther down here, you get similar coating of very similar hues of color.
09:58:35 So here the only unique thing is that you can have totally arbitrary links.
09:58:42 I think that's the one feature of the hippocampus.
09:58:44 You can see 2 things together in some environment for the first time in length, whereas these other more cortical regions rely on some, you know, more consistent feature, coding either semantic features or visual features.
09:58:59 Yeah, yes, I was just thinking in the context of I was trying to think, even based on Harris's question of like things like spatial memories of rooms, things like that.
09:59:10 I tried to imagine a completely new one, and I was finding it pretty hard.
09:59:15 Imagine a nice step. Our architect might be better at it, or something, so I was trying to take.
09:59:19 It says, like all that you domains, in which at the least, that, like pretty much all of the innerabilities simulator, imagine it sort of just pieced together by memory versus actually be able to generate it completely, take that as a semantic or whatever yeah, no it's an interesting so
09:59:38 you're asking are, is the input here pre-existing representations.
09:59:48 Is that I mean some. You could see a new person that you've never seen before.
09:59:58 You have strong priors in your system that allow you to recognize that as a face.
10:00:02 But the particular arrangement of Fe might be unique, and the idea is that provides input to the hippocampus.
10:00:10 So you can bind that face to some other specific thing.
10:00:12 But yeah, I think you're reusing your whole sensory architecture to generate the input for these processes.
10:00:19 Yeah. Okay. Yes.
10:00:28 Didn't agree that if you think you're saying, I'm also wondering to how you think about, you know the hypnotic campus doesn't get input from vision right?
10:00:35 It gets input from many, many other. But together you see that as well as the community. There's a canvas in the context of it.
10:00:43 That is, it's not only that you get multisensory convergence in other places, of course, but it is a property that the campus.
10:00:48 That's why smells can evoke really evocative episodic memory.
10:00:52 So there is multimodal convergence, and that's a kind of relation so there's something like multi sensory numbers.
10:00:59 Usually, in your view, entry work, but you feel like that's more uncommon in a hippocampus or yeah, I think it's yeah.
10:01:07 It's very common to have multimodal memories.
10:01:11 Yeah, I guess when good the hippocampus is one of a number of brain systems that seems to be doing a similar thing at the same time.
10:01:22 So you mentioned a page of ganglia, the Sarah films another structure, prefrontal cortex might be doing some parallel components of this and so you know, I think one complicated thing is that for 8 tabs all these brain structures are going to be contributing to a component of that
10:01:43 market, and with their time scales. Yup, just to kind of comment on.
10:01:51 So I agreed that super campus is pretty especially in terms of conversions of different parts of the brain.
10:01:56 Yeah, and different. Spenser, yeah, and inputs angularities in?
10:02:01 But singularity, prefrontal cortex has a raised.
10:02:06 It's especially in terms of this conversion. Yes. How do you think about this, too?
10:02:11 Yeah, yeah, so mean, I'm not going to really have time to go into it today.
10:02:16 But there, there's a version of this kind of idea for prefrontal cortex where it's not about refronting specific events, but rather representing schemas so like semantics over events.
10:02:29 If you go to a new restaurant, you kind of know what to do because you've been in other restaurants and those kinds of scripts or schemas seem to be coded in prefrontal cortex more than and the hippocampus this is a
10:02:41 consolidation process that happens from hippocampal episodes to more generalized schematic knowledge and medial prefrontal cortex and lateral prefrontal cortex.
10:02:49 That's the yeah. We can talk more about. But that's the basic way I think about the relationship.
10:02:55 Yes, so questions you're talking about one long term correct from the animal literature there's a lot about, I guess the function of the hypocrisy.
10:03:07 The long-term storage that this role for really storing it or being essential for retrieving it, may actually be a little bit when you put up the skis.
10:03:17 Would you propose actually humans, that the hippocampus is also the really the long-term storage space for the exact population? Space-time?
10:03:27 Or is it more? It's important to represent that. Calculate that.
10:03:31 Yeah. But then, really, if you wait after month, even if it's a one-time episodic memory, you, it's would you still locate an initial company?
10:03:42 Yeah, there's different views on this. Some people. This is back to the sort of complete retrograde amnesia idea.
10:03:47 So the reason why people thought retrograde amnesia was graded was that over time things became epicemically independent.
10:03:53 Or trade. Memory traces became epicentently independent. There's other theories, though, that says that even if one commanders are stored in cortex, they're still linked through the hippocampus, so you need the campus to retrieve them.
10:04:05 Right? So that's why you get complete, complete retrograde amnesia with selective hippocampal damage is that you can't access the memories that are stored in cortex.
10:04:15 So I think that's possible. In addition to that, I want to say that the hippocampus is certain functional capabilities that are relevant for many tasks, including pattern separation for encoding.
10:04:27 But other things as well. So I'll get towards that at the end.
10:04:31 But yeah, it's a great question. Okay? Well, I'm gonna jump forward Ana is gonna be here next week.
10:04:35 So I'm not gonna spend a lot of I don't think she's speaking, but you should speak to her because she's amazing.
10:04:41 She's a former graduate student. I'm just to show you the kind of data I'm not just like making up this idea.
10:04:46 We showed people sequences of these fractal patterns.
10:04:49 This is like the fisher houses in 2,002.
10:04:54 So there's a sequence with transition probabilities.
10:04:56 We use fractal patterns because hippocampus likes slightly more complex stimuli, but let's say A was always followed by B in the sequence.
10:05:04 Basically what we showed in this experiment is that from before to after exposure to these sequences, if you just showed individual objects on their own objects that had been paired during learning, come to evoke more similar patterns of activity in the hippocampus so it's plotted here is the
10:05:21 increase in the similarity of the activity pattern evoked by objects that had a high transition probability during learning versus objects that had a 0 transition probability during learning.
10:05:33 And, interestingly, there's a non monotonic effect that objects that have a weak relationship and a sequence end up differentiating from one another in hippocampus.
10:05:41 And we've seen this across many studies now. So this would be objects that add 100% transition probability objects that had a 0 transition probability and objects that had a 33% transition probability there's really something interesting non monotonic going on here, focus on it too much but
10:05:56 that's we have a grant right now on non-monotonic plasticity that might drive these kinds of hippocampal differentiation.
10:06:04 So this is taking the pattern of femor activity in a hippocampal region of interest before and after statistical learning, and looking at how the pattern evoked by each object on its own is modified by learning regularities, so this is an example that relation coding I was just
10:06:20 referring. Is there a correlation between this pattern and actual performance in this paper we found a weak relationship to behaviour in a post-test?
10:06:30 So this is in the more standard kind of statistical learning, design, and other tasks.
10:06:33 We have more online measures of learning that have a tighter relationship.
10:06:37 But it's never really strong, I mean, first of all, the behaviour in these tasks isn't great.
10:06:41 To begin with it. And so there's not a lot of reliability to to work with, but we find some weak relationships with behavior.
10:06:49 What we really, what I've always wanted is if it's true that A and B evoke more similar patterns of activity, is it the case that you might confuse A for B, or that if you see a the extent to which this hippocampal representation of b is activated relates
10:07:09 to how quickly you make a judgment about B. So we find some relationships like this, or predictive effects.
10:07:14 Yeah, do you know how many instances do you need of this violation of the sequence to shift from being totally from from one category to being decorated?
10:07:27 So if they happen always at the same time, then they have a similar representation.
10:07:31 Yeah, they can show one variation in this case. We didn't introduce violations is a great question.
10:07:37 But here these were always together. There were violations in this case, because a third of the time they appeared together 2 thirds so. But this is a pre-post design.
10:07:48 So we actually in this experiment. So we we don't get like a, you know, we can't look at the dynamics.
10:07:53 So we we don't get like a you know. We can't look at the dynamics, but it's be really interesting. Let's say you start with seeing these things together, and then you get a violation.
10:07:59 Does it switch? You know immediately? Yeah, this, this I forget. This is a few dozen repetitions of the pairs.
10:08:06 Long term exposure. It all happens within a session.
10:08:09 So yeah, it'd be interesting.
10:08:14 To. I thought the main thing about these tasks is what you see.
10:08:21 Oh, I mean, this is directional. So this one is always followed by that one in the sequence. Yeah.
10:08:28 And I'm just showing you the overall similarity that patterns have opened.
10:08:31 You present them separately. But there actually is an asymmetry.
10:08:34 If you look at the pattern, if you show a so, if A is always followed by B.
10:08:39 After learning, you show a it evokes B. More strongly than if you show B.
10:08:43 And look at how much it evokes a. So there's a temporary symmetry where you got shift in the representation of the directions of the exposure during the task.
10:08:52 But in a that in behavior you can reverse the order and get people to recognize pairs or triplets.
10:08:58 We published that a long time ago. There's a lot of invariants over direction and time in behaviour as well, so it's not about the second.
10:09:07 Not only about that right associations. Yeah, so this is grouped across your subjects.
10:09:15 Is this relationship sort of exist within a subject, or is this driven by this is a way these stats are all within subjects, are all within subjects.
10:09:23 Pair t-tests, or repeated measures in opioid-tests, or repeated measures in Ops. So this 133 that was within subjects, just in case it's like people who are not embedded in the human statistic learning literature might find these kind of pair
10:09:36 designs less interesting, I think, of regularities as existing in the world, and much more complex structures.
10:09:44 So this is a community structure graph and we basically did the same kind of experiment.
10:09:48 So this is a graph that has some interesting properties. But imagine you assign a fractal to each node on the graph, and then you do a random walk across this graph, and every time you visit a node, you insert that fractal into the sequence.
10:10:03 So subjects are shown a continuous sequence of fractals, but the order is governed by a random walk over the graph, and the question is, does the hippocampus come to encode the underlying generative structure that produced the sequence that people are exposed to and so this graph has
10:10:20 some interesting properties here. What drove these effects is transition, probability.
10:10:24 A is always followed by B. Arrays, never followed by B.
10:10:26 Here the transition probabilities are are, you know, indicated by the line.
10:10:34 So you can see each node has 4 edges, and so each of those edges is equiprobable.
10:10:39 So in this case the transition from this node that's part of the green community.
10:10:43 To this node. That's part of the purple community is 25%, whereas the transition probability between this node and this node within the same community is also 25%.
10:10:53 So the transition probabilities, the kinds of transition probabilities on the left actually are uninformative about the pairs or the relationships in this design.
10:11:03 What's different between the green and the purple nodes?
10:11:06 Is that all of the green notes here have overlapping sets of transitions, whereas the purple nodes have overlapping sets of transitions.
10:11:14 But it's not the actual transitions within the sequence themselves.
10:11:17 The first order transitions. Just to say that this principle we wanted to hold more generally that the hippocampus might come to encode the objects assigned to the green nodes as more similar to each other, because they tend to have overlapping transitions than the objects assigned to the
10:11:34 purple or to the orange nose. Is this design?
10:11:38 Make sense. I know it's kind of abstract.
10:11:39 But yeah, final designer is to design something that leads ambiguous, whether it's the communities or the secularity transitions that are in this case.
10:11:50 Yes, it's just a walk on the graph. And in this particular study we didn't know what they were relying about.
10:11:57 Others have followed up on this, and tried to unpack different statistics from the graph in the sequence.
10:12:03 Kind of meta-cestion. Yeah, when you were introducing complementary memory systems cortex, you're talking about the important role that the campus has for rapid learning and separating things.
10:12:14 Yeah, here it's both separating things and connecting.
10:12:16 I don't know. If this is something where you're going, but and here, right this is separated.
10:12:23 Separating. But it's also separation of episodic memory, legend like pushing one memory apart from other memories. But this kind of statistical learning has something like that which is these differentiation of learn differentiation versus integration.
10:12:37 Yeah, integration. Differentiation. Yeah, so this is this non monotonic plasticity stuff that we are working on.
10:12:46 Maybe we'll save that, for after that's a much longer thing.
10:12:49 Short version of it is, there's some evidence in many different animals and models, and so on, that if you have 2 representations and one sort of moderately active, while the other one's strongly active that that can weaken that can lead to synaptic depression and the connections between those representations
10:13:14 and so I think of that in this sort of weak transition probability case A brings B to mind a little bit, but not strongly, and that puts it into the zone where you can get kind of synaptic depression where you could eventually sort of decreased the connections between those and so
10:13:26 there's some evidence for this and other studies that I could tell you about.
10:13:32 It's something like the reconsolidation literature, if you're familiar, that.
10:13:34 So if you remind people of something then it makes it opens it up to be extinguished, or so it could be the idea of focusing on that property that computers campus is not so much about pattern separation per se.
10:13:49 But it has a pattern now. Yeah. And I don't think that's the same pattern separation but it seems like it's not unidirectional.
10:13:58 Right, and it's a bigger question as to why it could be similar to other aspects of we're thinking about it simplistically it's always being one way.
10:14:05 That. Really it's a general principle of what you always do.
10:14:07 Yeah, this happens in Ally Preston, who has done a lot of work trying the cortex also does this.
10:14:13 Some, like the same cortic region, can show like integration or differentiation effects and associative inference tasks, for example.
10:14:19 So it's not just hippocampus here.
10:14:21 It's all the negative prime, in fact, right? Negative priming. Think no.
10:14:23 Think there's many examples of this? Yeah, that might also be related to the finding, since a one in the monster.
10:14:28 Yeah, exactly. Inspiration. Some people think of integration as completion and separation is differentiation.
10:14:38 Yeah, absolutely. Just to show you that if you take the patterns of activity of voc by an fmri evoked by each of those nodes on their own, and you calculate the similarity of those patterns, and then you take that correlation, matrix you know do multi-dimensional
10:14:55 scaling the distances between the patterns ofoked by the nodes, conformed to the underlying graph structure.
10:15:02 And so this is just one visualisation of that to show it.
10:15:06 But basically, you can see that if you think of each point here as a point in a high dimensional space where each dimension is the activity of all of the boxes in the hippocampus, the distance between representations that were part of the same community is smaller than the distance between representations that are sort of closer to the
10:15:26 middle. So subjects have no idea about this graph. You don't tell them anything about the graph.
10:15:32 We at the end we asked them draw the graph, and they have no idea what we're talking about.
10:15:36 But the hippocampus is representing the graph.
10:15:37 So we've done other things. But this is one example of more complex regularities.
10:15:43 I don't just mean pairs are triple. Yeah.
10:15:46 And you mentioned this earlier that so there are other studies that that they have that is not adjacent, a dependencies justice, right, and they don't find involved in that.
10:16:03 Well, there's a lot of reasons why you get null effects, as one thing I'll say, you know, doing hippocampal imaging is requires certain sequences and certain ways of segmenting my hippocampus.
10:16:16 And I'm not sure that all the studies that report no effects in the hippocampus do the really careful hippocampal imaging.
10:16:26 That's neat, like. Theoretically, you don't believe that there is any reason no, I don't believe.
10:16:32 I think there maybe could be kinds of. There certainly are kinds of learning that are not epicemically dependent.
10:16:43 No, not with this kind of a design. I'd predict that it would be similar.
10:16:48 The critical thing. Feature of these designs is that there A and B have no, they are totally arbitrary.
10:16:55 The fact that they're paired is only a result of what you're seeing right now.
10:16:59 If A and B were like 2 people who, you know, are friends and you put in a test like, I don't think that's it's really learning arbitrary associations for the first time that I think is Hippocampal.
10:17:12 Okay, this Joseph asked before about the patient.
10:17:15 So I'll show some data on that. Yeah, very nice.
10:17:24 Is that an exclusive statement? Did you have the chance to look at a visual object area? Yeah, you see similar.
10:17:31 Things in visual object areas. Yes, you don't see the differentiation effects, but the integration effects you see in yeah, what drives what you know.
10:17:37 We don't know from this. But yeah, it's not only enough.
10:17:39 No, no, yeah, although that tends to be the common denominator across many studies.
10:17:45 So this is Lonn Sue's data. One patient like N equals one.
10:17:50 So all the caveats in the world. But that's the nature of this kind of work.
10:17:56 But what this is, a standard statistical learning, basically your design.
10:17:58 Joseph was with shapes. That's the original one more of the saffron style design, with syllables, with scenes.
10:18:06 Tim, who couldn't be here today to Brady, has a nice paper 2,000.
10:18:12 On scene, statistical learning and tones, because we don't know if this is language, it's a different auditory.
10:18:20 So why don't you, Susan? In the X, and then these are age and education match controls.
10:18:26 So she fails to in the standards. Statistic learning, type, design.
10:18:31 Where afterwards you have the subject. Do a tool alternative, force choice for patterns that conform to the sequence versus not she's not able.
10:18:39 We have positive controls in this study, where we show her.
10:18:43 Instead of asking her to judge familiarity of triplets, we have her judge.
10:18:47 The familiarity of individual items from the sequence versus novel items that weren't in the sequence.
10:18:53 And she can do that fine so she can discriminate old versus new items.
10:18:55 But she can't discriminate patterns among familiar items versus random combinations of this is with triplets.
10:19:03 The same thing works with pairs. We only had one patient.
10:19:07 Yeah, 50% is chance. Yes, correct. Is there any significance to a test? If you're so far below shared some total experiments?
10:19:20 So yeah, chuck it up to noise, because these are randomly group, because you know, you can just end up with things that spuriously seem like they go together.
10:19:28 That's like, probably what's going on here. But yeah.
10:19:32 And so you see, like sometimes controls are below chance, and if you run these tasks and healthy young adults, you get some people below chance as well.
10:19:38 Yeah, so the classic, you know, idea would be that if each of those 50 times you saw her, you repeated this experiment, then you know it would be learned, maybe not by the campus, but with the same structure we haven't done that that would be a painful experiment.
10:19:57 Yes, yeah, that would be a fear. What's that? At least you wouldn't feel the pain right?
10:20:02 She wouldn't remember exactly. Yeah, I feel the pain. So you mentioned the Simon that you see some of these effects in the visual system.
10:20:12 Yes, right? So we've done the brain imaging to see whether or not you have yeah, we haven't done that.
10:20:23 I could get to this question, of course, which interaction not with statistical learning.
10:20:29 But it's a great idea. I also just throw up. This is one patient.
10:20:32 So like there's a huge caveat. She has other damage right?
10:20:34 So let's not way too heavily. But Melissa Duff's lab basically Re ran this, this is Ana's work, so she can tell you more about it.
10:20:46 But we gave them all of our materials. They rerun it in, I think, like a set of in this case.
10:20:52 These are 4 hippocampal patients, but they also have patients with Mtl. Damage.
10:20:55 So it's a much bigger sample of pain. These are only patients, and they are all like at or around chance and lower than controls.
10:21:02 In this study. So this is a mix of patience with partial versus complete hippocampal damage, lateralized damage versus not, and so on.
10:21:11 And it's sort of specific to hippocampal damage.
10:21:14 So you get impairments in this bigger set of patients.
10:21:17 Yeah. So the comparison between being able to identify these pictures, but not.
10:21:26 What's the?
10:21:29 Yeah, that's what this is. That's what these lines are.
10:21:33 These are age and education. Match controls. This in controls.
10:21:40 Yes, yeah. And she also has antip files, I mean, so maybe she just does.
10:21:46 Well, the easier stuff. This is a general like criticism of all neuroscience stuff is that, if somebody has limited capabilities, maybe it's just that the task is more difficult for them.
10:21:58 And so it's not in this paper we've done other things that are demanding, like working memory tasks and so on, where she doesn't show an impairment.
10:22:06 So. Yes, it's almost impossible to fully so that is certainly more difficult to do.
10:22:11 The this kind of discrimination than item discrimination.
10:22:14 And I think that's a legitimate alternative.
10:22:15 Explanation for a lot of patient findings. I think, in this case, especially in this paper, where these patients aren't as all as severe but still impaired.
10:22:25 I think I'm a little more confident in that.
10:22:27 But yeah, it's a fair point. Yes, the affair that this association cares about is the familiarity test.
10:22:33 Versus the reaction time effects of the sequential. Yeah, yeah, sure.
10:22:41 So let me show this because I think this is another way of getting at this.
10:22:43 So this is work from Lucia Maloney. Who's that?
10:22:49 The Max Planck for empirical aesthetics.
10:22:52 In Frankfurt doing intracranial recordings with the Statistical Learning design.
10:22:59 So this is a frequency tagging type design. I don't know if this has come up in group yet.
10:23:04 Other people might talk about. But imagine you are doing a saffron type task, if that means anything to you, you hear an auditory stream of syllables, the colors here are just for visualization purposes where the syllable 2 is always called by p and p is always followed by rho but
10:23:20 what happens after row is the beginning of another. Tricyllabic word.
10:23:24 So in this case, B. So the transition probabilities within word are 100% between words.
10:23:29 I think in this design are 33%. So it's kind of similar to what I showed before.
10:23:35 The critical point is that if you present this, and if you generate these auditory sequences carefully, you can make it so that the syllables all onset at the same relative interval with respect to each other, so there's a frequency with which the syllables on set and so that's what we'll
10:23:50 call the syllable rate, and the case where you have tricyllabic words.
10:23:53 There's also a frequency at which the words on set.
10:23:56 So that's a tr syllabic boundary.
10:23:59 In that case it's going to be one third. The frequency of the individual syllable.
10:24:04 So, if this is, you know, 4 hertz that the syllables appear.
10:24:08 The word rate is going to be, you know, like a third of that one and one of the third hurts.
10:24:14 Critically. These word boundaries don't exist in the acoustics right?
10:24:22 It's a learn frequency. It's only by learning the transition probabilities that there's anything different about the transition from Rho to B than between B and dot.
10:24:31 And so others have done this in scalp, eg.
10:24:37 With Meg and other techniques. Using these kinds of stationary designs to look at learning by assessing how much power coherence do you get at the frequency of the learned boundaries versus the frequency of the sensory boundaries the control condition here is to present the same syllables
10:24:55 but in a random order, so the only meaningful structure there is at the rate of the syllables.
10:25:01 There's no word structure. And so this sheet.
10:25:03 The seed did this across whole brain. Bunch of intracranial epiloguey patients, broad coverage.
10:25:07 This isn't just hippocampal. This is broad cortical recordings.
10:25:11 And just to show you how nice it are. I mean, Aaron is asking about reaction time data.
10:25:17 This is another way to get an online measure of statistical learning so it's plotted.
10:25:20 Here is coherence is a function of frequency. So you can see in both the structure and the random sequences.
10:25:25 You get high, coherence at the syllable rate that's in the acoustics.
10:25:29 You can. That's the prosody of the soul.
10:25:32 So that's that's trivial. However, in the structured stream you also get a bump in coherence.
10:25:37 At the word frequency, that doesn't exist in the random case.
10:25:43 This one, this is the pair frequency. So, yeah, so this is 2 syllables.
10:25:49 And so this is really interesting to me. Is this a result of learning the pairs?
10:25:54 And then chaining them together. Yes, it could also just be non-susoidal responses that could be.
10:26:04 Don't know from that. It almost looks like the syllable frequency tracking is also stronger in structure.
10:26:08 I don't think that's consistent. I think the means are pretty similar.
10:26:13 I think the variance is maybe a bit higher in the structured.
10:26:17 That would be interesting, too. I mean the idea that you may even been in training wars, but I mean, that could be a attention could be could be more.
10:26:26 There is some evidence for that. Yep, just to show you I like.
10:26:29 Asked Lucida. Look at the hippocampal depth electrodes in this design, because these are cortical electrodes.
10:26:36 But most of these patients also have depth electrodes in the hippocampus.
10:26:39 And so what she did is she took the spectrogram of activity in these hippocampal depth, electrodes elicited by each of the syllables, and did the same kind of multi-dimensional scaling I just showed you for that graph structure and you see the same kind.
10:26:52 Of grouping of syllables that are part of the same words.
10:26:54 In other words, the spectrogram, the average spectrogram, have activity evoked by the syllables is more similar for syllables that were part of the same learned word, so the hypothesis that hasn't been tested.
10:27:07 Yet is. It's this kind of hippocampal learning that's driving the cortical synchrony to the boundaries.
10:27:12 So what we're doing with Licia now? And also at Yale is to stimulate.
10:27:17 The hippocampus. During this task, and see if you can knock out the word synchronic in cortex.
10:27:25 That's the hypothesis that we're testing.
10:27:28 We've run 2 people, and so far that works. But that's just 2 people.
10:27:31 Yes, were the subjects able to report the words yeah, they heard in these sequences well, in that sort of 2 alternative force, choice, discrimination task.
10:27:42 After learning? Is there any that during States? Sure, in fact, if you hear, I should have played it if you just listen to these sequences, you start hearing the words.
10:27:53 It's very striking. And then they really studies.
10:27:58 They present like little babies with these sequences for 2 min.
10:28:00 This is a very so you're asking sort of is it something about them recognising the words, or how they're attending to the syllables? And there's a bit of a chicken and egg problem.
10:28:10 They can only do that differently. In this case. In this case, as a result of learning.
10:28:13 And so your question is like, is this the learning itself? Or is this some way that the learning is affecting?
10:28:18 How they're attending to, or perceiving the sequences.
10:28:23 And you can't say from this, but if you, if you knocked out the hippocampus and it eliminated this, I think it would be consistent with our interpretation.
10:28:31 Yeah, I'm going to ask my usual question. Please.
10:28:35 The relationship between these old signatures and behavior I don't remember in this paper.
10:28:38 That's a good question. Yeah, hey? Many mixed feelings about that question, because I'm not sure that you would expect.
10:28:50 I'm not sure what the relationship is between hippocampal representations that's that's like an open question.
10:28:58 Because they're filtered through. Sensory systems or motor behavior.
10:29:03 And so you'd want it to be the case. We often find relationships if you want to claim that there's a causal that people come up with developing these representations is close of the darn.
10:29:14 Well, if you want to claim it's causal to behavior I can claim it's causal learning without showing that's correlated behavior.
10:29:19 Behavior comes off from alert, from having learned something. I agree here that we are looking at here can only be subjects, can only perform the task by our measure performance if they have learned the representative.
10:29:32 I agree with that. I'm talking about so causation, I guess what I'm saying is, yeah, right?
10:29:38 But you can get learning in the brain. There's plenty of learning in the brain that doesn't.
10:29:42 And so I'm also okay with a case where you stimulate that the campus and knocks off frequency tagging to boundaries.
10:29:48 Saying that the campus is necessary for this learning effect. We're a slightly different question of whether this or that correlates with their discrimination performance.
10:29:59 My only pushback is that those behavioral tests are also not great, so I don't talk a lot up to a null effect in predicting a bad behavioral measure.
10:30:06 That's my days, but we always look at that, because that's also my interest in trying to like explain behavior.
10:30:16 Is there another question? Ok, 'm, just going to show this, and this is the end of this section. And then we're already at 1030. Wow! This time flies.
10:30:26 So I will give you a choice about what we do next.
10:30:28 After I show this, this is Ana's work again.
10:30:31 This is a nice paper. This is sort of the next generation of complementary learning systems.
10:30:37 Theory. It's a Philosophical Transactions paper in the.
10:30:43 There's a special issue on statistical learning in 2017, and some of you may be part of.
10:30:49 And so this is a modeling paper with a bunch of simulations of the tasks that I showed you.
10:30:54 The pair coding task, the community structure, task, associative inference, type tasks.
10:30:59 This is a simplified version of a hippocampal neural network.
10:31:03 So this is actually taken from a paper by Randy O'reilly.
10:31:08 Cat's it all. If you're familiar with that, basically, the squares represent different layers of the neural network, and also roughly different subfields of the hippocampus this is a subset of the hippocampus is not a totally biologically, accurate but
10:31:26 the idea is that entorana cortex provides the input to the hippocampus and receives the output.
10:31:31 This line is what's called big loop recurrence.
10:31:35 So the output of the hippocampus feeds back into the input so Darsh Kumaran and Jay Mcclellan.
10:31:40 Others have a model of so big loop recurrence that can explain some of these results, too, that came out after this or around the same time.
10:31:47 But the basic idea here is that within the hippocampus there are at least well more than 2.
10:31:54 But there are 2 key pathways here. One is fermentoronto cortex into dente gyres to see 3 to c, one.
10:32:01 This is the traditional memory, episodic memory pathway that people talk about with Dente Gyrus and Ca, 3.
10:32:10 Having relatively sparse activity which allows similar inputs and a Toronto cortex to get mapped to somewhat orthogonalized representations.
10:32:19 So if you have really high inhibition in these layers, even correlated interiorinal inputs can produce different patterns of activity.
10:32:27 So it's the sparsity and inhibition of these layers that drives simulations of pattern separation in this modeling framework.
10:32:34 So there's this pathway that triesnapped pathway.
10:32:38 So 1, 2, 3, and then there's this pathway.
10:32:41 The temporal monic pathway that goes directly from Mentoron cortex to Ca, one.
10:32:45 And there's some evidence that the weights here are less sparse, and so that correlated inputs and a Toronto cortex resulted in correlated patterns in Ca, one.
10:32:57 And so what honest shows in this modeling paper is that the trsynaptic pathway shows properties of episodic memory that is storing related experiences to distinctly, whereas the monosynaptic pattern shows something more like statistical learning where it will encode related
10:33:17 experiences more similarly and it'll extract transition probabilities and other kinds of predictive relationships.
10:33:23 And so per ship in the field is, instead of saying, hippocampus, does episodic memory cortex does statistical learning.
10:33:33 It's a bit more nuanced within the hippocampus.
10:33:36 You can get both, at least relatively like short time scale.
10:33:41 Statistical learning seems to be able to happen in the model along this monosynthic pathway and I'm not going to go into the details of this.
10:33:47 And hopefully get a chance to talk with her. This uses a contrastive heavy learning training algorithm that when you're Riley developed for sort of simultaneous trying to predict the interiorinal output from the interiorinal input and toggling between whether you derive the answer from the
10:34:07 trisnaptic pathway, or the monosynaptic pathway so you're toggling between training these 2 pathways.
10:34:12 I won't go into details. But in the paper she does do some really nice experiments where you lesion one pathway, and you look at what kinds of memories are learned.
10:34:20 Or lesion, the other pathway, and she shows you and get episodic memory like effects by lesioning the monozypic pathway.
10:34:27 And you can get statistic learning effects, even if you lesion the trice.
10:34:31 And what I mean by that both is what gets output from the model and the similarity of the representations in the model sort of conformed to what we see in Fmri data.
10:34:41 Okay, so let's see, where are we?
10:34:48 I'm you know. That's the end of Section one, and we're an hour and a half in.
10:34:52 So I'm tempted just to summarize, for now and then maybe we can decide what to do to summarize what I think's going on here is that statistical learning, at least as we study it in human in these kinds of tasks might be an intermediate form of learning
10:35:07 between the canonical complimentary learning systems, which is sort of rapid one shot, episodic encoding, and this much slower kind of statistical learning that's incomparable learning systems that required days or weeks and consolidation over a long time period in order to set up
10:35:25 Cortical Knowledge, Cortical regularities. So the kinds of learning that happened in statistical learning tasks in humans can happen over the course of minutes to hours.
10:35:34 They don't require sleep. They don't require consolidation, although that is an effect.
10:35:39 And we're going to talk about that later, hopefully. So that's how I think of it is this sort of a third learning system.
10:35:45 If you will, and, unlike the traditional view where cortex is doing all of the work of extracting regularities against scores, things discretely in the cortex, you might think of this manussynic pathway as doing some of the initial integration, across memories that
10:36:02 then gets Consolidated. And so that's kind of one of the ideas we're testing now is to not only get consolidation of individual episodes but do you get consolidation of structured knowledge and that's hard to test but that's something we're
10:36:17 working on. So I called, remember now, but most of the use a particle, and probably the most popular version of the statistical learning is sequences. Essentially yeah.
10:36:33 So statistical patterns across time, whereas even the one of the first examples you gave about the concerting of the podium.
10:36:42 Yeah, that's about licensing facial. Yes. And so, as we know, there's a lot of I'm wondering how much of this kind of interim summary you think, also applies to that sort of statistical learning, or is there something specific about sequences here.
10:36:58 I think there's something special about sequences and at least it's more amenable for testing predictive consequences of these kinds of representations.
10:37:07 And there's less work of this type on spatial regularities.
10:37:14 There are a couple of studies that have done more of the like.
10:37:17 2,001 Fischer Naslun type design that you guys have done.
10:37:21 She arrays. There's a couple studies on that. There's more extensive work on contextual queuing, which is not quite the same, but a similar type flavor where there's regular arrangements of distractors during visual search and there's both imaging and patient
10:37:36 studies showing that that's hippocampally dependent.
10:37:39 If you'd asked me a priori. Is this more temporal or spatial?
10:37:43 I said, this is definitely going to be more spatial, but all of our work was more focused on the temporal dimension so I would expect that this would hold I don't think anybody, as far as I know, maybe we're on our others know, has done a version.
10:37:56 Of this, where you do a spatial, statistical learning, then you do the pre post representational change, type analysis.
10:38:03 I don't think that's been done. As far as I know.
10:38:08 This is probably a silly question, but it's like is spatial navigation.
10:38:14 Yeah. Oh, I'm glad you asked that. I was forced to skip something that I wanted to show.
10:38:20 All right. So there's a beautiful paper. If you haven't read it by Richards is the first author and Paul Franklin is the senior author.
10:38:28 It's a nature neuroscience paper from 2013 or 14, where basically they do a Morris water maze.
10:38:37 With rodents, and they have. So there's a hidden platform in the and submerged, and so the room swims around, finds a platform, and they get a new platform location each day.
10:38:48 This is a model we were talking about spatial representation.
10:38:51 This is a model for studying is often used as a model for studying episodic memory in rooms, because you put the rowing back in the in the pool, and you see where they swim.
10:39:01 Do they go back to the platform they found. Right? That's a specific memory for a certain location given the landmarks.
10:39:08 And what this study did is across days. They sampled the platform locations from an underlying spatial distribution over the pool, and the first experiment was like a Gaussian distribution over the pool.
10:39:20 So in every day it was a different location platform was sampled from that distribution, and what they found is after month the rodents stopped going back to prior platform locations, and they started going back to the mean of the distribution.
10:39:37 So this is one way that people have studied spatial navigation in a statistical learning context, my student cat has published a series of papers now doing this in humans in VR, so that's what these papers are basically, where humans navigate.
10:39:53 And a virtual environment, a circular virtual environment, to find a hidden reward, and the reward is sampled from a spatial distribution.
10:40:00 What she finds behaviorally is that over time, within a session, so in rodents it's 30 days in humans.
10:40:06 Within an hour they start navigating to the like. The mean of the distribution, or, in the case of a bimodal distribution, to the 2 modes of the distribution.
10:40:14 This is a way of studying spatial, statistical learning and a navigational context, and she's got some nice intrananial data here, too.
10:40:22 I won't go through this. But basically she finds that if you take the patterns of hippocampal, electrophysiology when you put the person back in the maze before they have to navigate, you see that they sort of activate the mean of the distribution, of the prior
10:40:37 locations. But if you test them again longitudinally like in this case, 3 weeks later, they no longer do that, and instead, the mean of the distribution is represented in medial prefrontal electrodes rather than hippocampus, so this this is part of what I was saying, on that last slide that I
10:40:55 think hippocampus does the initial work of extracting regularities in humans, but that over time it might become more of a cortical or frontal type process.
10:41:03 Yeah, one sorry war. I fascinated that train rotors, these.
10:41:11 For them. It seems to take a long time for them to varieties, and one said, This kind of really explains a fast flowing over, yeah.
10:41:24 But then, if you have this behavior where it takes a moment.
10:41:33 For me has really struggled to find I agree. I have the same struggle.
10:41:40 It's actually very hard to come up with tasks that are matched that lead to the same time scale learning across species.
10:41:49 None runs or monkeys couldn't do any of these human tasks without extensive training.
10:41:53 And so then, are you studying the same thing? I don't know.
10:41:56 Yeah, it's a fair question. I don't have the answer to it.
10:42:01 Yeah, so just on that point, I think in part, it's also, I mean, what's the nature of the instruction that you give to humans?
10:42:10 And so I think the instructional capacity is much higher so we have some data from our network, even a very simple object recognition task.
10:42:21 You get 0 instructions of units, they perform almost just like, yeah, I love that.
10:42:27 That's a that's a good point. I think the verbal explicit instruction, even if it's a very efficient way of transmitting task.
10:42:35 Constraints, and now maybe sort of evolved for that.
10:42:40 But we just can't do that so let's add Nimkins into this, though. Right? You're not doing verbal instruction for infants.
10:42:45 But it wouldn't. This. It could be bulk.
10:42:49 It could be both. Yeah. But I think in lot of the cases, a lot of the time to get some training, just basically learning the generalized.
10:43:00 Because that you do have, like the exact change their skin. That's they can.
10:43:07 They can change the distribution very quickly. Right? So I will talk more about it.
10:43:12 Good. Okay. I look forward to it. Also in 2 facilities, a water based as Edmonds has that be allowed out?
10:43:23 So you have a spatial working.
10:43:26 Basically.
10:43:31 The floor is a glass that we are back to generate.
10:43:34 So then we can basically present mistake. And there is a target on this.
10:43:44 They would target these like videos, the sticks moves around and the angle should draw and poking the remembered location of this J.
10:43:51 The training takes few months right? You get you retained them to do the spatial work, but then, once they do that, we also see this kind of a contraction mean of the distribution.
10:44:03 But we didn't do that. I'm pretty sure that if you start changing the distribution they will add that once they know the whole general structure of the task, just the statistics, you know, I agree that the question, I think, is just like when is the hippocampus important in that process right is it
10:44:22 at that end stage when you change the distribution, may adapt their behavior.
10:44:25 Was it in that months where they're learning a task?
10:44:29 I don't know, but maybe that could explain why the hippocampal effects in animal seem to take longer to acquire.
10:44:34 If it's part of the task learning itself. Ok, so 1045 I didn't say any about development.
10:44:43 Maybe we should take a break, though. 15 min and come back at 11 and then I'll just say 5 min about development, and then we'll switch to a panel discussion.
10:44:51 But it's already very lively, is 15 min enough, Joseph.
10:44:55 Us, people, do people want water? Break? Are you tired of sitting in there 15 min?
10:45:01 Ok, come back at 11, please.
11:03:51 I'm not going to ramble on for too much longer, rather than to just introduce 2 other ideas before we open it up to more discussion.
11:03:58 Already has already been a lot discussion. First thing I wanted to say is, if you, jumping way ahead in a developmental part.
11:04:10 But if you think about that model, I just presented that maybe there's more of an episodic-like learning effect in the Trisynaptic pathway through Dentigy virus and C 3. And the hippocampus is a more statistical type learning effect and the monosynaptic
11:04:23 pathway to Ca. One. This might account for a really interesting developmental sort of mystery that's been in my head since we started this work, which is the infants, seem really good at statistical learning.
11:04:36 That was some of the earliest human studies were an infants, and I have continued since then, but infants are also really seemingly really bad at episodic memory.
11:04:48 So there's a phenomenon called infantilenesia that you may have heard of, which is that as adults and older children we remember almost nothing from the first few years of life, and sometimes people say they don't remember stuff from when their baby it's pretty much we don't know for sure but it's
11:05:03 unlikely, so it seems to be most. If you do, retrospective surveys.
11:05:07 The earliest memories people have later in life are start around 4, 5, 6, 7 years.
11:05:12 So were able to learn a lot as babies were learned. Language.
11:05:18 We learn how to walk, we learn relationships that are the names of objects, but we don't have specific episodic memories from that period that persistence.
11:05:29 So this is a mystery. How do we learn so much?
11:05:31 But not in code. Memories. And I think this sort of pathway idea, this is what we're testing helps address this.
11:05:41 So if you look at this is in Macaques.
11:05:42 If you look, if you do tracing studies to look at the development of these pathways and macaques, it's hard to map map this onto human.
11:05:50 So I don't know what the homology is, but in an infant macaque you actually don't see.
11:05:54 You can't trace from a Toronto cortex along either of these pathways and juvenile macaques.
11:06:00 You see an emergency connection between and Toronto cortex in Ca.
11:06:05 One the monocyclic pathway. And it's not until adult macaques where you see the full trsynaptic pathway through Dentate C.
11:06:11 3 to c. One. So this suggests that there might be a differential developmental trajectory more statistical versus more of episodic type. Learning.
11:06:23 And there's a lot of normative reasons why this might be the case.
11:06:26 I'm not going to go into. But just to say, I think these kind of circuit models could lend themselves interesting developmental predictions.
11:06:32 Yeah. What do you mean by you don't see the projections? Yeah.
11:06:37 And these are trace of studies and dissections in Macaques.
11:06:42 This fascinating, because you would expect that one trial journey would be necessary.
11:06:46 You would expect that. That's why I'm interested in this. Yeah, yeah.
11:06:49 So one way to think about that is that episodic memories are the input to statistical learning.
11:06:55 This pathway idea is a slightly different account that the one shot learning is really separate from the more integrative learning.
11:07:00 Yeah, yeah, is there? So coming from from vision, do people think about critical periods related to episodic memory and or statistical learning?
11:07:11 I don't know that people think about statistical learning from a critical periods, perspective.
11:07:16 You certainly see it throughout the lifespan. Episodic memory.
11:07:19 People have talked about critical periods, more the critical period wouldn't necessarily be that you can't see it throughout the lifespan. It's just that there's more in certain periods of minutes, right? So I guess I don't think there's evidence, right?
11:07:32 Are there more plastic in? I don't. It's very hard to know whether you're getting the same amount of learning or speed of learning, and for instructional reasons for stimulus, complexity, reasons so I think it's hard to know you tend to find very similar effects.
11:07:46 There are age effects on statistical learning. Maybe Marner knows this work much better than I do, so you see, better statistical learning performance over childhood, older children.
11:07:57 But again, it's hard to know that they're better understanding the test instructions, or the more attentive during learning.
11:08:01 There's a lot of reasons for that think about in terms of it's not a framework that's used, the way that it is an accessory.
11:08:11 That's right. I've never heard a discussion of critical periods. That's right. I've never heard a discussion of critical periods for statistical learning.
11:08:15 But in the episodic memory literature there's absolutely the thought that sort of underage.
11:08:20 4 or 5. There's a qualitatively different.
11:08:23 Give capabilities for episodic encoding. There's different accounts of itatomyia.
11:08:29 One account is that babies don't store episodic memories.
11:08:32 I don't believe that there's other counts that are more retrieval, based that they store episodic memories, but because of cortical development and other brain changes later in life.
11:08:41 You can't access those memories. That's sort of where I come down on that more.
11:08:45 I have data show that we want to go into now.
11:08:49 But yeah, so I think this might be a good time to kind of bring up different types of learning, because sure, you know the most, you know, conventional kind of argument of us is still learning, and the infant, you know, often is for pattern to representations is that you know how do you know what a basis.
11:09:09 Or a dog, or a cabinet things of that sort, and that some of this you know, could be semantic.
11:09:15 But a lot of it, you know, is basically advertising the visual.
11:09:19 Yeah, and so you's, that's probably cortical exactly right? So probably a representational component of cysticleording that's happening early on, for sure.
11:09:30 Then a lot of what you're talking about is more of a association.
11:09:35 Most of basically, how are events replaces or things of that bound together?
11:09:42 And then the last part is this for me, that has been brought up in a lot of other talks are basically relationships between the experiences and actions and so like, when you get to the foraging that's a little bit closer to this, other concept that you know might also involve different brain structures that
11:10:01 might have also different development role. And so I do think that finding a way to either differentiate these things or make arguments that know their whole the same, you know.
11:10:14 So I don't think they're the same. I will say that that's associative thing has been studied, you know, even in newborns, both at all.
11:10:20 You know, and up. So the second type that you talk about is what I'm referring to here.
11:10:26 I don't think you need the hippocampus to do the perceptual narrowing and cortex that drives, you know, human face preferences, and then the longer or the other third type linking to actions.
11:10:40 I don't know. I don't know of a lot of work in that space.
11:10:44 Yeah, so maybe just to say, like what we have done in infants.
11:10:50 It's very hard to study the infant hippocampus, because you can't do it with normal infant neuroscience methods like, eg.
11:10:58 Or nears so what we've been developing is ways of doing fmri and babies where we can image the hippocampus.
11:11:03 So I can show you like just that. Maybe the task that we use for that I'm not going to go into any details about this.
11:11:10 But basically, this is like what it looks like. This is based on the Kirkum design.
11:11:20 Natasha Kirkham did a study where they looked at 2 month old, 5 month old and 8 month old.
11:11:24 And they did basically the 2,002, the fisherman has to a temporal statistical learning design.
11:11:29 So you present looming stimuli where there's a transition probability structure.
11:11:36 So, you know, it's very, very simple. Statistical learning and babies.
11:11:41 But behaviorally they show orienting preferences.
11:11:44 After this kind of exposure to patterns versus random things.
11:11:48 In this case we just compare brain activity to sequences with structure versus sequences, where orders random, but where the individual items are equally familiar.
11:11:58 And so in infants we see stronger responses to the structured sequences in the hippocampus than to the random sequences.
11:12:07 Am I gonna go through all this? Basically, we see a learning effect.
11:12:10 If you break down. This is a small amount of data it's like, you know, up to 6 min per kid. The scanner.
11:12:16 So from an fmri standpoint. This is not much, but if you look at over time, the difference in aggregation in the hippocampus structured versus random sequences, you see a difference emerge by the second half of exposure, so that's some evidence it's not
11:12:31 age dependent. So this is to the earlier point, like we see it.
11:12:36 This is the distribution we see, you know, equally kind of weak effects, but across the age span.
11:12:43 So this is a claim that maybe early on there's some hippocampal involvement, statistical learning.
11:12:50 And then the episodic memory piece of this is also a very simple task, but here we don't see early evidence of episodic memory or episodic encoding.
11:12:59 So this is a task borrowed from the adult literature.
11:13:02 Is his subsequent memory design. So you see individual stimuli. Now.
11:13:06 So it's not associations, objects, places, faces, and then we test episodic memory in the babies by having them Orient in a visual pair of comparison.
11:13:18 We measure their eye movements. So I think you'll see a test trial here in a second.
11:13:24 That!
11:13:29 Okay, so it's a test trial. You look at how much they orient to the picture they've seen before.
11:13:34 As a measure of some kind of memory. It's one shot running, and they only saw the picture once and we measure the amount of looking at the familiar picture.
11:13:43 The psychedelic background is to keep them watching the screen.
11:13:46 Otherwise, it's context. There's other reasons, but it's an attention grabber, because otherwise they fuss out pretty quickly.
11:13:54 You're equating oneship memory memory. Yeah. And you know, this is like an animal problem, too.
11:14:01 How do you know a baby is having this retrieving, episodic memory?
11:14:04 It's really hard. I don't know what the answer to that is.
11:14:07 Yeah, level of complexity. For sure, what we want to get to is the kind of recollection that you were me.
11:14:15 II remember the barbecue last night, I said, excess person. Aaron gave me a sausage like that's what I want to talk about.
11:14:21 This is very hard to, in pre-verbal babies, and so it was a burger while you're getting a burden temperature.
11:14:31 I think the lesson here is I ate too much because I had a hot dog, a sausage, and a burger.
11:14:34 But yes, that's where we want to get it. Yeah, no.
11:14:39 I realize that the problematic. But is there any evidence in adults, for example, that this type of memory and episode the more complex episodic memory equally I wouldn't say they're equal.
11:14:48 But this task in adults does drive at the counties not exclusively.
11:14:51 You can get behaviour in this task, based just on entail cortex, and you would have defects on the subjects with that, and correct although they could still do it, they would be impaired compared to that.
11:15:03 Anyways, I just going to show it. We're out of time.
11:15:06 But basically. This is what this is during a scan. This is a baby, and we're tracking their eye movements.
11:15:12 And then we measure proportional looking at the familiar stimulus, and then.
11:15:17 It's kind of that. Looks like it was.
11:15:21 Pacifier in there. I'm holding his hand. It's it's weird data collection.
11:15:26 We've scanned like 300 babies now.
11:15:29 So if anyone wants to do, let me know. Okay, yeah.
11:15:36 Can't write the same card. Not an external memory.
11:15:39 Specifically use the fractal images.
11:15:45 Well, what would you look at? So they're just seeing one fractal at a time.
11:15:50 So what would you want them to? Don't use faces and landscapes?
11:15:56 Oh, you mean, do it with fractals, well, each picture there is trial unique whereas in the statistical learning design they see the same fractals over and over in different patterns, you could do one trial unique fractals right?
11:16:08 And that would be a cleaner comparison, I agree, part of the reason we use faces, objects, and scenes is that there's evidence that infants have category selective visual areas.
11:16:17 And so we're wanted to look at whether there are cortical memory effects as well but yeah, that would be a nice comparison.
11:16:23 Anyways, I'm just going to show one key result here.
11:16:27 Which is in this task. So we're in the statistic learningcast, we found equally robust effects in young babies and older babies.
11:16:35 In this task the effect is entirely within the older babies.
11:16:39 So these data are consistent with a differential developmental trajectory.
11:16:42 At least within infancy, between a very simple, paired, statistic, learning task and a very simple item.
11:16:51 Recognition, subsequent memory task. So this is the kind of thing we're working on.
11:16:54 Now, this suggests to me that babies are at least on this very short-term thing.
11:16:59 You see a picture? And then 2 min later, you do a visual pair of comparison.
11:17:02 Last night or a year ago, but they have some ability to do one shot encoding of pictures that involves the hippocampus, and this is selective to the epicemus.
11:17:11 If you look at these babies, there's no other predictive signals in the brain that predictive signals in the brain that predicts subsequent working in this task.
11:17:18 So that's the kind of thing we're working on now.
11:17:20 Yeah. So I'm curious about the development in the monkeys that we showed.
11:17:26 So I'm not really up to speed in terms of the periods and monkeys.
11:17:31 How correlated is that with their movement, ability, or speed, I mean, some of these things kind of make sense.
11:17:37 If you cannot remove that much, if you don't need to form these kinds of structures, and once you can, then yeah, no, it's a great question.
11:17:47 It's hard to know how the monkey development stages map to human development.
11:17:52 So is a juvenile macaque a one-year-old baby.
11:17:56 I don't know in terms of movement, though the statistics, the sensory statistics, change dramatically with motor development.
11:18:03 You know, this is your point. You're like lying on your back, you know, for a few months you're looking at the ceiling or maybe people leaning over you, and then you're crawling, and you're looking at the floor and you're walking, and you're looking at kind
11:18:14 of low, and then you grow taller. So the natural statistics of experience changed dramatically.
11:18:20 And so I think that's a really interesting space I don't know that anybody's looked at the nature of what babies can learn at different like in terms of sensory statistics, at different stages of motor development.
11:18:33 Maybe Lauren knows something yeah, it would be really interesting.
11:18:35 I think you are asking a slightly deeper question. Does the neural development track of learning track the neural development of motorbi?
11:18:46 And I think that's also a really interesting question. Yeah, Ok, the last thing I'll say, partly because someone during the break brought it up you know, I was gonna at the end sort of go back and and address this original taxonomy that Monte raised an appropriate chronology about this is a really
11:19:07 nice review paper that I'm not going to do justice to that.
11:19:10 I courage you check out the source, starts from this and says, Ok, given everything we know.
11:19:13 Is this the right way are these the right boxes, boundaries, and I won't go through the arguments.
11:19:23 But basically, there's plenty of evidence that these aren't quite the right boundaries.
11:19:29 And there's some logical reasons for that. Based on the neurocycle data.
11:19:33 And then there's also a lot of imaging data behavioral work that came out after this taxonomy was, you know, proposed.
11:19:41 And so the sort of takeaway from that is that you see, this paper argues for a different sort of orientation, which is rather than thinking about task-based constructs like an episodic memory task, or a statistical learning task, or a priming task, they argue.
11:20:04 For i'm not sure this is better, but they argue that there are a set of basic processes or computations are implemented in the circuits for these memory systems, and that different sort of cognitive psychological test load on those computations different so if you set up a task where you only
11:20:19 show something once right, you're tapping rapid, encoding right?
11:20:25 If you have a task that requires arbitrary.
11:20:26 Associations that might also tap into the ability of the epic campus.
11:20:31 So they're trying to sort of shift away from using these terms and thinking about what are these terms?
11:20:37 What are the common underlying processes that are implemented by these brain systems, so that's kind of how I think about it.
11:20:44 And for the hippocampus. I think there's several candidates here.
11:20:47 I haven't. I haven't gone through this, but we talk about relational binding.
11:20:51 Be able to link are things that are arbitrary, related.
11:20:53 We talked about convergence zone for different multimodal inputs, talked about the ability to orthogonalize similar inputs came up before the ability to take a partial queue like I could say, what did you have for lunch yesterday and you can retrieve a lot of information that I didn't
11:21:08 supply you're not just recognizing what I'm saying.
11:21:11 You're filling in missing information. This is recollective ability.
11:21:14 There's also an idea that there's something akin to a prediction error, computation in the hippocampus, so that the fact that the monosynaptic trisynaptic pathway both converge on c one has been prosed by lis men and grace
11:21:27 and others to do. Sort of a differencing between expectation and experience.
11:21:32 So if that trisaptic pathway does something like pattern set pattern completion, I say, once you retrieve some stuff, the monosynaptic pathway provides more current sensory information, and so Ca, one can compute a difference between your expectation based on memory and your actual experience it's not a reward
11:21:49 prediction error. It's more of a sensory prediction.
11:21:52 Error with an so just say these are some of the computations that people studied episodic memory tasks depend on a lot of these, but other tasks do too.
11:22:00 So you can set up, perceptual tasks that require pattern separation.
11:22:04 You can set up attentional tasks that benefit from relational binding.
11:22:09 And so a lot of the modern work shows hippocampal involvement in much broader set of cognitive psychological functions than the sort of narrow view of a one to one mapping between episodic memory and hypocemple function yeah, it seems as though, from the previous slide that the 2 critical
11:22:25 dimensions that this review paper outlines this speed sort of rapid, slow, and then the rigidity of flexibility.
11:22:35 And so do you think that now, if you fast forward and think, Ok, someone's going to look at a taxonomy that looks at those sort of more, perhaps dimensionalize.
11:22:47 That's the word I was thinking. Do you envision that?
11:22:50 Ok, now, we're going to poke. Holes in sort of what are we doing by slow? Yeah.
11:22:55 Or is the goal to kind of identify the precise timeline of what slow means for rigid associate.
11:23:01 So I love that paper. I think it's fantastic, but it doesn't do this.
11:23:05 What you're doing, punting on this. If you're relabeling the dimensions and stuff, I think of it more like this.
11:23:13 These are all things that you can model. These are all things where there are neural and behavioral assays, and that they don't depend on as much on psychological constructs.
11:23:22 So speed to me is not a binary thing. It's not going to be slow versus fast.
11:23:27 We talked before about different time scales you do get hippocampal involvement in working memory.
11:23:32 I've sort of skipped over that like if you do a working memory task you see, hippocampal activity.
11:23:37 Right. So what is that? So? Yeah, I think timescale is continuous.
11:23:43 But I think it's still very helpful to really think about it.
11:23:48 I mean a lot of, let's say, qualitative social science, and put everything in a 2 by 2.
11:23:52 Right, right right? And so the the important thing there makes very challenging is, can you fill every box in the box?
11:23:59 That seems unclear to be able to fill in that previous framework is slow.
11:24:03 And flexible right. And so, at least in that diagram that there's not in the framework of slow, fast, rigid, flexible.
11:24:14 Can you get a slow, flexible yeah. So I don't think she I don't think Katery Hanky talks about that in this paper, but I think of slowflex to me, slow, flexible is generalizable.
11:24:23 Now. So if I learn world knowledge that allows me to make predictions in a new scenario that's flexible, that doesn't have to be the same context, I can use it. So I think there are examples of slow, flexible.
11:24:35 But yeah, you could go through that exercise for sure. I can't follow a dichotomy of slow and fast and flexible. You can follow this.
11:24:45 You don't understand, or you dispute it. I just can't understand how it can be, can't get behind it, because it, you know, it's very relative it doesn't get to, you know, these types of computations or things that need to be solved by this so that I violate it
11:25:05 by all sorts of camera examples. I agree. So the reason I listed it.
11:25:08 This isn't a paper she just says, rather encoding of associations, but under the hood you could say, Ok, that could depend on a subset of these things that we know.
11:25:17 So you could go through this exercise for every memory system. There are other people in this room are more experts than me.
11:25:23 What are some interesting properties of the basal ganglia?
11:25:25 Well, it's sort of timing dependent. It's more integrated with the motor system.
11:25:30 There's like certain properties of all of those. Quote unquote memory systems, and some of them overlap.
11:25:36 So there are many tasks where this is. I didn't go through it, but where you don't get a one-to-one mapping it's kind of obvious to people who do this work statisticalarentheses. You don't just get hypocemple learning effect.
11:25:46 Episodic memory task. You don't just get hippocampal effects, but that might be because of functional overlap across systems in these underlying computations.
11:25:55 So I mean kind of unavoidable that time. Stay within working.
11:26:02 When you look at like like trace learning task in time, matters whether the campus becomes involved involved or not.
11:26:13 If it's shorter than 5 s, or longer than 10 s things are very different in if there's a gap or no gap, denial that these timescapes are important in the real brain, no, but I think that if you try to define things based on time, scale, you're
11:26:30 ultimately going to get confused because there's a lot of different brain structures that have time scales that are relevant and they're heavily overlapping as opposed to.
11:26:38 If you basically describe things in terms of types of computations or functions, and you relate those to typical or atypical timescales then I think that can be very useful.
11:26:49 Yeah, I think you guys are green. But you're basically saying, you want to do a task decomposition.
11:26:54 Say, what are the what? What do you need to accomplish this task?
11:26:57 What can brain region do? And then you could generate a prediction of which systems should be involved.
11:27:01 I love that trace conditioning, delayed conditioning example, because that's a perfect case.
11:27:05 Where you get toggling between memory systems based on a very simple variable, which is the amount of time, and whether there's a gap right?
11:27:14 And I do have data. It's not published. But in a delayed comparison task broadened auditory task.
11:27:22 So tactile. So it depends on the working delay between the first stimulus and the second string lose if it's shorter than 4 s, you don't see that memory, the canvas you make that that interval longer than second, or 6 again, suddenly you can decode your memory
11:27:39 from the mechanics. I think the key thing is that you're talking about a different type of timing than the other fastboat economy.
11:27:46 Because when we're talked about learning rate, and now your basically talking about activation periods and local memories and so that's the exact reason why say fast and slow isn't useful because fast and slow for to a whole bunch of different things, that it's all.
11:28:04 But I totally agree with you that understanding time scales associated with computations is necessary to edit.
11:28:12 Yeah. So they're different apps of time, which is maybe one is the amount of time in a task or the delay between encoding and test.
11:28:20 Then there's the time scale over which the learning happens itself in a longer term.
11:28:25 Memory sense, and so that might be different. Senses of time yeah, maybe I'll just jump to stop talking now.
11:28:33 I'll just pull up the discussion here. This is, I sort of wrangled some folks who have some knowledge about different aspects of this, although many people in the room do as well, and sent them these questions last night, so they've had a lot of time to come up with very thoughtful, answers
11:28:55 but I just to put these up these are some interesting, some questions. I don't know if there's a questions that I think about.
11:28:59 Sometimes. So I just just walk through them, and maybe we have about an hour before 1230.
11:29:08 And so it's already been a very active discussion. Why are learning and memory studied separately?
11:29:14 I don't think like most people in this room would call themselves memory researchers.
11:29:19 It's not long- memory researchers. Just tomorrow morning builds up.
11:29:22 Longterm memories. Why do they? Different developmental trajectories?
11:29:27 We already talked about that a little bit in relation to motor development.
11:29:30 But is there a normative story there? Is it more useful to learn regularities early in life before you start remembering specific empathy what's the interaction between the 2?
11:29:39 How, and why do statistical learning and episodic memory cooperate or compete?
11:29:46 I think I didn't go through. But we have some data that statistical learning actually prevents you from encoding episodic memories.
11:29:53 So if you're like, if you're learning a predictive relationship if you're generating a prediction based on statistical learning, you're less likely to encode that experience in memory, there's different ways to think about that question are there different types of statistical learning whatever that means by
11:30:09 memory systems. So I focus on hippocampal-like, statistical.
11:30:14 Or maybe it's sequential associate irregularities.
11:30:17 That is hippocampal. Maybe spatial is separate, or maybe more probabilistic regularities are stridal, whereas it became really good at these deterministic kinds of relationships, different ways of thinking about that and then I think you know, maybe this this is something I think a lot.
11:30:34 About. There's a lot of concepts in the memory literature that aren't normally incorporated into statistical research, but could be like, for example, on consolidation.
11:30:43 So the classic case. So yeah, don't know. Maybe I'll call on each of the 4 people and then you could pick a question that you chiropractors are there chairs?