[00:00:05.01] bachelor's degree at sanford and then moved on to do a phd at princeton in psychology neuroscience [00:00:13.16] [00:00:14.13] and then moved on to a post doctor or fellow position at the harvard medical school uh in 2019 [00:00:21.14] [00:00:21.14] when she accepted an assistant professor position in at the university of pennsylvania [00:00:27.03] [00:00:27.03] in philadelphia so she has been there for about a year and a half now uh having uh to deal with [00:00:35.05] [00:00:35.20] studying the lab during the pandemic um so hopefully that's been working [00:00:40.02] [00:00:40.02] okay for her i know he's working on memory and learning with a particular [00:00:46.06] [00:00:48.10] emphasis on consolidation during awake periods in sleep and she is also working on your network [00:00:55.09] [00:00:55.09] modeling so we're very pleased to have her and anna take it away great hi everybody thank you so [00:01:03.12] [00:01:03.12] much for having me thank you for the introduction dobie um so i want to start with um this really [00:01:12.15] [00:01:12.15] broad question which is how does the brain store new information and i want to define two ends of [00:01:19.16] [00:01:19.16] a representational spectrum from localist to distributed representations which i'm going to [00:01:25.12] [00:01:25.12] illustrate here in terms of their use in neural network models um so in a localist representation [00:01:31.16] [00:01:32.06] some input comes in from the environment there's some set of neurons or units that are going to [00:01:37.20] [00:01:38.13] represent that information across this hidden layer here um and when i say localist all i mean [00:01:43.22] [00:01:43.22] is that some new input that comes in regardless of how similar it might be to that first [00:01:48.06] [00:01:48.06] piece of information is going to be represented with a completely non-overlapping set of neurons [00:01:54.02] [00:01:54.02] in this hidden layer as opposed to a distributed representation where you allow there to be [00:01:59.11] [00:01:59.11] overlap across the different the different inputs that you see in the environment and [00:02:05.20] [00:02:05.20] this distributed kind of representation has been crucial to the success of neural network models [00:02:11.20] [00:02:11.20] in machine learning applications and i think also in um in in thinking about how these models apply [00:02:20.23] [00:02:20.23] to the brain and to to behavioral phenomena um but there are pros and cons to using this kind [00:02:26.23] [00:02:26.23] of representation uh maybe the most obvious pro is that if you have this kind of overlapping [00:02:34.21] [00:02:34.21] representation it makes generalization really easy and automatic so if you for example represent [00:02:41.11] [00:02:41.11] all of your memories of seeing all these different birds flying with this overlapping representation [00:02:47.18] [00:02:47.18] that makes it really easy to to infer properties of some new thing that you see [00:02:54.17] [00:02:55.07] that that has a lot of features in common with all these birds that you've seen flying before [00:03:00.15] [00:03:00.15] maybe you haven't seen this particular bird flying but um you'll be able to immediately understand [00:03:05.14] [00:03:05.14] that probably this bird also has that property of being able to fly so this kind of representation [00:03:10.15] [00:03:10.15] makes that kind of generalization really easy whereas a localist representation makes that [00:03:16.08] [00:03:16.08] extremely difficult because these representations are separate and so um so it's less easy to see [00:03:22.12] [00:03:22.12] those relationships between experiences on the other hand there are some kinds of memory tasks [00:03:27.22] [00:03:27.22] where this type of representation is really useful like if you're trying to remember that penguins [00:03:34.08] [00:03:34.08] uh swim but they don't fly whereas this very similar looking bird does fly in this case uh [00:03:40.15] [00:03:40.15] this kind of representation is useful because it has such a low interference so it allows you to [00:03:44.15] [00:03:45.12] think about those differences between um between these objects whereas in this case you get there's [00:03:52.00] [00:03:52.00] going to be much more interference and this interference issue is associated with the kind of [00:03:58.02] [00:03:58.02] behavioral difference between these two representations which is that [00:04:01.11] [00:04:01.11] um these distributed representations require information to be learned in an interleaved order [00:04:07.16] [00:04:07.16] the order presentation matters whereas that's not true for localist representation so just to give [00:04:13.05] [00:04:13.05] you a quick intuition for that if you haven't um thought about that before in uh an interleaved [00:04:19.18] [00:04:20.10] uh presentation situation for a local store presentation you can go back and forth between [00:04:25.16] [00:04:25.16] studying information um and that will that works fine if you present information in a blocked order [00:04:31.03] [00:04:31.03] so like one set of information before you move on to the second set that also works there's no [00:04:35.22] [00:04:36.13] there's no overlap in these representations so the order presentation doesn't it doesn't make [00:04:40.08] [00:04:40.08] any difference but in the distributed case um interleaved exposure works really well [00:04:46.19] [00:04:46.19] but blocked exposure has this issue where to the extent that the representations are overlapping [00:04:54.02] [00:04:54.02] for this second set of information relative to the first there's a tendency to overwrite um [00:05:00.10] [00:05:00.10] that information that you um that you initially learned so distributed representations are really [00:05:06.23] [00:05:06.23] great at finding and storing structure and data efficiently but this representation is highly [00:05:12.10] [00:05:12.10] susceptible to interference and this is the classic catastrophic interference problem that [00:05:19.03] [00:05:19.03] still plagues many neural network models most i would say um so what is what does the brain use [00:05:26.15] [00:05:27.14] well probably it uses um some combination of both of these kinds of representations and [00:05:34.08] [00:05:34.08] that was the proposal of the classic complementary learning systems theory which basically said well [00:05:40.00] [00:05:40.13] maybe the hippocampus uses um these these localist style representations and it does really [00:05:47.18] [00:05:47.18] rapid learning of individual episodes and it's useful to use this kind of uh representation [00:05:53.12] [00:05:53.12] because it keeps all those experiences separate to avoid interference and then offline maybe [00:05:58.15] [00:05:58.15] during sleep the hippie camps will replay these experiences in an interleaved order [00:06:03.12] [00:06:04.06] and that will allow the neocortex to slowly extract the statistics across these experiences [00:06:09.20] [00:06:09.20] and build up these these distributed representations that are so useful for [00:06:13.16] [00:06:13.16] kind of long-term storage of structured information so um so this i think is a is [00:06:21.16] [00:06:22.08] basically right and explains a lot of data but there's um a really important missing [00:06:27.05] [00:06:27.05] piece here which is that it doesn't explain how uh we can do generalization quickly before [00:06:34.00] [00:06:34.00] there's time for offline replay and sleep um so we know that we're very good at this in in [00:06:40.02] [00:06:40.02] um various different uh ways so like when you have to generalize something across the time scale of [00:06:47.14] [00:06:47.14] seconds or minutes or hours how do we do that how does that fit into this picture and we've proposed [00:06:52.17] [00:06:52.17] that actually the hippocampus is capable of this kind of rapid statistical learning as well um as [00:06:59.09] [00:06:59.09] this localist style learning that it has this kind of intermediate style representation [00:07:04.00] [00:07:04.00] where things are somewhat overlapping and there's a kind of intermediate learning role [00:07:07.14] [00:07:08.04] learning rate and that allows the hippocampus to both extract statistics and learn episodes [00:07:14.23] [00:07:15.18] so um i'm gonna tell you about some of the experiments that got us thinking about the [00:07:20.17] [00:07:20.17] hippocampus in this way um and i'll show you a model of the hippie campus that tries to explain [00:07:27.16] [00:07:27.16] how it's possible that these two different kinds of representations could coexist in one brain area [00:07:32.23] [00:07:32.23] because these things these things really are in tension computational tension so how is it that [00:07:37.11] [00:07:37.11] one structure can do both of these things um and then mostly today i'd like to focus on our [00:07:43.20] [00:07:43.20] recent work testing predictions of this model so um how how might we know that this is what's [00:07:50.06] [00:07:50.06] happening as opposed to some other um some other way of producing rapid generalization and then if [00:07:57.05] [00:07:57.05] i have time i'll tell you a little bit about um what we're thinking about in the sleep domain so [00:08:02.15] [00:08:02.15] once you've encoded these new representations then how do you um process that information [00:08:09.03] [00:08:09.03] offline and and kind of complete this process of building up these neocortical representations [00:08:15.18] [00:08:18.04] okay so the the experiments that got us um thinking about the hippocampus in this way [00:08:23.01] [00:08:23.01] are um our visual statistical learning experiments where participants see a stream of novel images [00:08:29.22] [00:08:29.22] presented one at a time and there's some hidden temporal structure um embedded in these streams [00:08:36.02] [00:08:36.02] so in this first experiment there were pairs of objects that always occurred together so the [00:08:43.05] [00:08:43.05] these objects are seen much more frequently back to back than two objects that happen to appear [00:08:48.12] [00:08:49.12] together in the transition from one pair to the next [00:08:52.04] [00:08:53.12] and we measured the representations of each of these individual [00:08:56.17] [00:08:57.09] objects in the hippocampus well in the whole brain um before and after this sequence exposure [00:09:04.00] [00:09:04.15] and um right so we're just picking up the patterns of activity across the voxels in [00:09:10.04] [00:09:10.04] the hippocampus and other areas that are evoked by each of these individ these individual images [00:09:15.18] [00:09:15.18] and then we can ask whether images that were paired in the sequence exposure are represented [00:09:20.21] [00:09:21.18] more similarly or differently than images that were not paired during the sequence exposure [00:09:27.03] [00:09:28.12] and what we find is that over the course of the sequence exposure there's a change in pattern [00:09:34.23] [00:09:34.23] similarity in the hippie campus where items that were paired together um tend to be represented [00:09:40.17] [00:09:40.17] more similarly and there's no change in items that were not paired together so this suggests that [00:09:47.03] [00:09:47.03] the hippocampus is sensitive to this kind of statistical learning [00:09:50.15] [00:09:51.05] and also that it's using this kind of overlapping representation of related items [00:09:58.15] [00:09:58.15] um that's maybe not what we're used to seeing in the hippocampus or what we would expect [00:10:03.03] [00:10:03.03] from an orthogonalized localist representation in the hippocamp [00:10:06.13] [00:10:08.17] uh we we think the hippie campus is not just sensitive to these things but it's really [00:10:13.05] [00:10:13.05] critical for this kind of statistical learning so here we tested a patient with [00:10:18.17] [00:10:19.11] bilateral hippocampal damage um on the same kind of paradigm and we found that she was [00:10:24.19] [00:10:24.19] unable to learn um this kind of structure across across not just shapes but also um scenes and uh [00:10:32.15] [00:10:32.15] tones and syllables so even auditory stimuli she was unable to pick up on this kind of um [00:10:38.02] [00:10:38.15] statistical structure and that's been replicated in some some additional uh hippocampal patients [00:10:46.06] [00:10:46.06] and it's not just the simple kind of pairwise structure it's also more complex structure we [00:10:52.04] [00:10:52.04] think the hippocampus can learn so this is an example of a temporal structure that's uh [00:10:59.03] [00:10:59.03] that's created from uh um from this graph with community structure that we think is kind of a [00:11:04.04] [00:11:04.04] more realistic form of temporal structure and this these sequences can't be parsed based on [00:11:09.14] [00:11:10.21] pairwise co-occurrence frequencies or simple differences in transition probabilities [00:11:15.14] [00:11:15.14] you have to be sensitive to this higher order structure and we find that people [00:11:19.14] [00:11:19.14] are sensitive to the structure and that the hippocampus again represents the structure [00:11:23.14] [00:11:23.14] in a similar way way where items in the same community end up being represented more similarly [00:11:30.08] [00:11:30.08] than items in different communities so we think that hippie campuses is very good at this um kind [00:11:36.23] [00:11:36.23] of picking up on structure over time and seems like it's using these overlapping representations [00:11:41.20] [00:11:43.18] okay so so it seems like the hippocampus is both separating related experiences i didn't [00:11:49.18] [00:11:49.18] show you data on that today but there's quite a lot of literature on that [00:11:52.19] [00:11:53.14] but also allowing related experiences to overlap as a function of their statistical structure [00:11:59.16] [00:12:00.08] um and these two kinds of representations are um they're different they're really in tension and [00:12:07.07] [00:12:07.07] how is it that the that the hippocampus could be doing both of these things so [00:12:11.18] [00:12:12.17] um to try to understand this we have been working with this neural network model of the hippocampus [00:12:19.03] [00:12:19.03] that has um that has properties that reflect what we know about the actual hippocampus [00:12:24.19] [00:12:24.19] it has internal uh cortex that provides input and output to the model and then we have three hidden [00:12:32.06] [00:12:32.06] layers that represent three of the subfields of the hippocampus dead date gyrus ca3 and ca1 [00:12:38.17] [00:12:40.00] um the if you haven't seen this kind of um this kind of depiction of a neural network model before [00:12:47.05] [00:12:47.05] the the height and color of these little boxes represents the firing rate of a neuron [00:12:52.12] [00:12:52.12] or or maybe a small population of neurons and these arrows represent connectivity between [00:12:58.12] [00:12:58.12] those neurons um the connectivity looks actually something more like this but we just use those big [00:13:05.03] [00:13:05.03] arrows to represent that that fuller connectivity so there are two pathways through the through this [00:13:13.16] [00:13:13.16] model and and the actual hippocampus there's the tri-synaptic pathway and the monosynaptic pathway [00:13:18.23] [00:13:18.23] and these pathways seem to have different properties the tri-synaptic pathway is really [00:13:24.06] [00:13:24.06] the pathway that implements those famous sparse localist-style representations that are the heart [00:13:32.15] [00:13:32.15] of the reason that we think the hippocampus is probably so good at episodic memory and the [00:13:37.03] [00:13:37.03] reason that it has these kinds of representations is that there's this really kind of specialized [00:13:41.18] [00:13:42.12] sparse fanning connectivity into dentate gyrus that will take inputs even if they're very [00:13:48.19] [00:13:48.19] overlapping and project them to non-overlapping populations of neurons in this area so you end [00:13:55.22] [00:13:55.22] up with this pattern separation process this direct pathway that goes from internal cortex [00:14:02.06] [00:14:02.06] to ca1 seems to have different properties um has less extreme pattern separation and that might [00:14:09.14] [00:14:09.14] be related to the fact that um the there's less extreme scarcity in the in its connections with [00:14:15.07] [00:14:15.07] the the regions that it that it connects to um and also it seems to have a slower learning rate [00:14:22.04] [00:14:22.04] um than the tri-synaptic pathway so this pathway on its own can do slow incremental learning [00:14:29.09] [00:14:29.09] but um so it's it's able to learn you know over the time scale of like tens of trials um uh [00:14:38.02] [00:14:38.02] whereas this right but whereas it wouldn't be able to learn on on one trial whereas the tri-synaptic [00:14:43.14] [00:14:43.14] pathway can do one-shot learning and is less less involved in kind of more incremental learning so [00:14:49.11] [00:14:49.11] um so rodents burden studies have been able to lesion these two different pathways and show that [00:14:54.08] [00:14:54.08] they're these different properties so um we're gonna present this like statistical structure to [00:15:02.12] [00:15:02.12] this model and just see what happens can it handle um parsing these sequences so the way we do this [00:15:09.07] [00:15:09.07] is really simple we just show the model um two items at a time where the current the current item [00:15:15.20] [00:15:15.20] has full activity and then like the immediately proceeding item will have some decayed activity [00:15:20.17] [00:15:20.17] we just move that window along and we ask can the model figure out from the statistics over time [00:15:26.21] [00:15:26.21] that there are these pairs in this case embedded in this structure and we can contrast that with [00:15:34.04] [00:15:34.04] like a kind of a classic episodic memory kind of simulation where we just ask the model to directly [00:15:40.19] [00:15:40.19] memorize that there are pairs of things that go together so memorize the a and b go together and [00:15:46.10] [00:15:46.10] just move on to the next pair so here we're not going to require the model to extract statistics [00:15:51.09] [00:15:51.09] over time but just see if it can quickly form these these bindings between these items [00:15:57.11] [00:15:59.16] and then after we train the model either with the statistical structure or the [00:16:03.16] [00:16:03.16] kind of classic episodic variant we want to test the model in some way that's analogous [00:16:10.04] [00:16:10.04] to what we uh what we did in our fmri experiments which is um very easy to do in this kind of model [00:16:16.21] [00:16:16.21] we're just going to show the model after after learning item individual items by themselves [00:16:21.20] [00:16:22.12] and then record the pattern of activity that's evoked across these different layers of the model [00:16:27.03] [00:16:27.03] so we can say okay this is the pattern of activity evoked by item a in region ca3 and we can do the [00:16:33.09] [00:16:33.09] same thing for all of the other items and then we can um do these same kinds of like correlation [00:16:41.11] [00:16:41.11] analyses that we do in the in the human data so we can ask you know are two items that were paired [00:16:47.07] [00:16:47.07] in the sequence represented more similarly than items that were not paired same kind of question [00:16:53.12] [00:16:54.17] so this is just to orient you to the figures i'll show you here so first i'm going to show [00:17:00.23] [00:17:00.23] you the results for the episodic variant of the model where these pairs were clearly demarcated [00:17:05.18] [00:17:05.18] and there was no need to extract any statistics from the sequences and what we see here is that [00:17:12.12] [00:17:12.12] the dentate gyrus and ca3 so the trisynaptic pathway are really good at learning this kind of [00:17:19.03] [00:17:19.03] structure which is which is what we would expect from um from known properties of this um of this [00:17:25.22] [00:17:26.19] pathway and also from previous modeling work um so basically these areas are are building up [00:17:33.05] [00:17:33.05] a conjunctive representation of each of these pairs and each item by itself will evoke that [00:17:37.14] [00:17:37.14] same conjunctive representation and so you get this really strong pair structure in these areas [00:17:42.06] [00:17:43.01] da1 does have sensitivity to the structure but it's much weaker basically it would [00:17:47.11] [00:17:47.11] just take more time for ca1 to build up strong uh representations of these pairs [00:17:52.15] [00:17:53.09] so this tells us what we already knew about this pathway which is that it's very good at quick [00:17:58.12] [00:17:58.12] episodic learning so what happens in this situation where the model has to extract the [00:18:04.13] [00:18:04.13] statistics from the structure from the sequence over time here we see a totally different [00:18:10.02] [00:18:10.21] pattern of representations in the model it's it's the opposite of what i just showed you where now [00:18:16.15] [00:18:16.15] c1 is really good at representing this pair structure whereas dentate and c3 are completely [00:18:23.14] [00:18:23.14] failing they are doing something interesting this checkerboard structure [00:18:28.08] [00:18:28.08] suggests that these um regions are picking up on both the transitions between pairs like bg [00:18:36.10] [00:18:37.03] and the actual real pairs like a b as opposed to something a pair like a and g which was never [00:18:44.04] [00:18:44.04] actually shown together in the sequence so this pathway is doing too good a job of memorizing [00:18:51.18] [00:18:51.18] everything that it's seen which includes this the pairs and the transitions between the pairs [00:18:56.02] [00:18:57.01] and that's that's exactly what you want from from an episodic memory system um but it's [00:19:01.12] [00:19:01.12] totally not useful for this learning problem where you're trying to extract the statistics [00:19:06.10] [00:19:06.10] over time so these different areas have different properties as a function of the learning problem [00:19:13.11] [00:19:13.11] and the structure of the input um and this shows us that ca1 can solve this task and it's the [00:19:19.09] [00:19:19.09] only part of the hippocampus it seems to to be able to solve this statistical learning problem [00:19:24.19] [00:19:27.12] um one one kind of interesting application of this model is in the domain of [00:19:35.12] [00:19:35.12] development because infants have very undeveloped hippocampi but they're uh but they seem to be very [00:19:42.02] [00:19:42.02] good at statistical learning and so for people who think that the hippocampus is important for [00:19:46.15] [00:19:46.15] statistical learning it's a little bit of a puzzle why infants would be so good at [00:19:50.19] [00:19:50.19] statistical learning if they if they don't have developed hippocampi um but it turns out that [00:19:55.20] [00:19:55.20] actually these two pathways develop at different different times and it's it may really be that [00:20:01.20] [00:20:01.20] it's the tri-synaptic pathway that's undeveloped in infants because the monosynaptic pathway seems [00:20:06.15] [00:20:06.15] to develop very early on and so we can ask okay in the model if we don't give it access to the [00:20:13.20] [00:20:13.20] trisynaptic pathway and has to just learn on this monosynaptic pathway what happens [00:20:19.11] [00:20:19.11] um and it turns out that it's completely fine in fact it actually does even a little bit better um [00:20:25.18] [00:20:25.18] without the trisynaptic pathway so so this suggests that the monosynaptic pathway [00:20:30.02] [00:20:30.02] can act on its own and it might it might provide some insight into what's happening [00:20:35.09] [00:20:35.09] for infants who don't yet have this tri-synaptic pathway episodic memory system set up yeah [00:20:40.19] [00:20:43.12] so um so i have um you know made the argument that this pathway has these properties that are well [00:20:52.06] [00:20:52.06] suited to extracting structure over time that the slow learning rate is really good at kind [00:20:56.15] [00:20:56.15] of picking up structure over many trials and the overlapping representations are useful for seeing [00:21:02.15] [00:21:02.15] that um seeing that there's shared structure across time and we think that it's not just [00:21:09.09] [00:21:09.09] temporal statistical learning that these kinds of properties will be really important for it's [00:21:13.16] [00:21:13.16] really any situation where you need to pick up on structure across across experiences and so [00:21:19.09] [00:21:19.09] we've been um applying this model to other kinds of domains one area we think uh that that this [00:21:29.18] [00:21:29.18] learning strategy should be really useful for is um category and concept learning um where you're [00:21:35.03] [00:21:35.03] trying to understand uh uh some new category uh by picking up on um structure of exemplars [00:21:42.21] [00:21:42.21] over time and we know that the hippocampus is uh plays an important role in this kind of learning [00:21:47.12] [00:21:48.17] so i'm going to show you a couple simulations um that um that jalan asuchevich has done applying [00:21:54.21] [00:21:54.21] this model to a category the category learning um domain so first she uh took the classic [00:22:02.15] [00:22:02.15] weather prediction task where you see a set of these abstract cards and you're supposed to um [00:22:07.22] [00:22:08.17] categorize into sunshine or rain and we know that amnesiacs can do this task but that [00:22:16.06] [00:22:16.06] they're impaired so we don't think that the heavy campus is the um the only region but capable of [00:22:22.21] [00:22:22.21] of doing of doing this kind of learning but it's certainly playing a role so we want to try to [00:22:27.16] [00:22:27.16] think about what that what that role might be so yelana finds that um that so here she's breaking [00:22:34.13] [00:22:34.13] it down by the intact model in green and then a version of the model that only has the [00:22:38.17] [00:22:38.17] monosynaptic pathway in orange and then a version that only has the tri-synaptic pathway and purple [00:22:43.14] [00:22:44.12] and she finds the model can do this categorization task if you take uh if you take away the triceptic [00:22:51.09] [00:22:51.09] pathway just use the monosynaptic pathway it seems that it's maybe even doing a little bit better [00:22:57.05] [00:22:57.05] like we saw in the statistical learning case and that it's really the monosynaptic pathway that's [00:23:01.01] [00:23:01.01] responsible for this the trisynaptic pathway totally fails on this this uh prediction task if [00:23:07.05] [00:23:07.05] you change the task and instead ask the model to recognize particular card combinations that it saw [00:23:13.14] [00:23:13.14] during exposure now it's a very different story now that tri-synaptic pathway kind of learning [00:23:18.08] [00:23:18.08] strategy becomes really useful um and it turns out that the that the tri-synaptic pathway only [00:23:25.01] [00:23:25.01] model does the best and um and the monosynaptic pathway by itself struggles a lot with this kind [00:23:31.05] [00:23:31.05] of uh task so both of these these um pathways are engaged by this task and are doing something but [00:23:38.17] [00:23:38.17] um but that the the actual category learning aspect of it is driven by the monosynaptic pathway [00:23:44.23] [00:23:46.15] he also applied the model to this um this object category learning paradigm where participants [00:23:52.02] [00:23:52.02] learn about these three categories of of these fake satellite objects objects in the same [00:23:57.18] [00:23:57.18] category share most of their features their parts but each object also has unique individuating [00:24:04.21] [00:24:04.21] parts and and participants learn about both of those things and we find um in in our [00:24:12.00] [00:24:12.00] uh imaging experiments that use this task that ca1 is sensitive to this category structure so their [00:24:18.12] [00:24:19.11] items from the same category are more are represented more similarly than items from [00:24:23.09] [00:24:23.09] different categories in ca1 but not in ca 3 and dentate and also that there's just more similarity [00:24:30.00] [00:24:30.00] overall in in ca1 which i think is consistent with this idea that it's using a different kind of um [00:24:35.20] [00:24:36.15] representational scheme so in the model um if you ask it to do a generalization task [00:24:44.02] [00:24:44.02] so you um you show it a new satellite it hasn't seen before and ask it to try to [00:24:48.13] [00:24:49.12] infer some uh some missing feature the um the intact model can do this well the tri synaptic [00:24:57.16] [00:24:57.16] pathway only model really struggles with this and interestingly the monosynaptic pathway only model [00:25:03.11] [00:25:03.11] does really really well at this task so this is the this is the most obvious example we've seen um [00:25:09.11] [00:25:10.00] uh what looks like a really strong trade-off between what's happening in these two pathways [00:25:14.19] [00:25:14.19] and a demonstration that in something like generalization the mono synaptic pathway is [00:25:19.01] [00:25:19.01] really um is doing all of the work but again if you if you change the task and ask the model what [00:25:25.09] [00:25:25.09] it you know um can you remember unique features of individual exemplars now the monosynaptic pathway [00:25:32.21] [00:25:32.21] fails at that and the tri-synaptic pathway um is completely responsible for that kind of memory so [00:25:38.10] [00:25:38.10] so again the tri-synaptic pathway is doing something it's sensitive to these objects and [00:25:42.06] [00:25:42.06] to the extent that you want to remember what's unique about them it's critical but it's not [00:25:48.12] [00:25:48.12] helping with this this generalization process that requires understanding the structure over time [00:25:55.20] [00:25:57.09] and if you look at the hidden layer representations of the model you'll see that ca1 [00:26:01.09] [00:26:01.09] represents this the structure really nicely which is consistent with what we see in the imaging data [00:26:06.13] [00:26:08.23] so i've i've shown you that the mod synaptic pathway is good at um this temporal statistical [00:26:15.14] [00:26:15.14] learning and also at uh category learning building up kind of new semantic memory [00:26:20.23] [00:26:20.23] um have we are we done have we solved the problem or are all that are there alternatives are there [00:26:27.01] [00:26:27.01] other ways um that we might uh think about how the hippie campus contributes to these tasks and [00:26:33.05] [00:26:33.05] how might we test between these alternatives um so so in order to address this question i am going to [00:26:42.17] [00:26:42.17] bring up one last task um where there's been a lot of theorizing and modeling about um about how the [00:26:49.20] [00:26:49.20] hippocampus might might contribute which is this associative inference task where you learn that [00:26:56.06] [00:26:56.06] two things a and b go together you learn that b and c go together and then there's some kind [00:27:00.15] [00:27:00.15] of test of whether you can do that transitive inference from a to c and we know that this task [00:27:09.11] [00:27:10.06] relies on the hippocampus from um from rodent lesion data and also from from human fmri data [00:27:16.02] [00:27:18.10] so how do you solve associative reference there's uh there's been different proposals so um one [00:27:24.10] [00:27:24.10] proposal uh was implemented in this remerge model which said well let's let's kind of preserve and [00:27:33.18] [00:27:33.18] focus on that localist kind of coding scheme in the tri-synaptic pathway and see if we can get [00:27:39.18] [00:27:40.17] inference to emerge from that coding scheme without assuming that there's anything else [00:27:45.22] [00:27:45.22] happening so the way that this works is that there's um there's a try synaptic pathway-like [00:27:52.10] [00:27:52.10] kind of layer that has these representations of each of these pairs of items so kind of [00:27:57.18] [00:27:57.18] conjunctive episodic representations of each pair and they're connected [00:28:02.02] [00:28:02.02] to these item representations so if you present item a by itself at test [00:28:07.14] [00:28:07.14] it will activate that a b memory that you formed during encoding which will remind you of the [00:28:14.06] [00:28:14.06] b item which will get you then to your bc memory and then you can solve the task and get to c so [00:28:20.10] [00:28:21.07] you without having stored the relationship between a and c at encoding as long as you [00:28:25.09] [00:28:25.09] have these these kind of recurrent connections set up correctly um you can you can use spreading [00:28:31.01] [00:28:31.01] activation to solve the problem at retrieval so that that's one strategy um that's very different [00:28:37.07] [00:28:37.07] than the strategy that i've been telling you about which is this kind of interleaved learning [00:28:41.11] [00:28:41.11] to build up distributed representations strategy where you have items um a and c they start off [00:28:47.12] [00:28:47.12] with maybe some amount of overlap in their representations and then as you go back and [00:28:52.17] [00:28:52.17] forth between studying a b and b c you you kind of merge these representations because that's [00:28:59.12] [00:28:59.12] what this uh this kind of learning algorithm will tend to want to do to the extent that things are [00:29:04.23] [00:29:04.23] similar in the environment that a and c share this b feature it's going to tend to push these [00:29:10.02] [00:29:10.02] representations together um but critically again this this the order of presentation is going to [00:29:17.01] [00:29:17.01] be important here um so interleaved uh learning allows this kind of representation to be built up [00:29:23.03] [00:29:23.03] whereas some other kind of order would not allow that so that's gonna be that's gonna be important [00:29:28.23] [00:29:29.22] um okay but if you have the injury leave learning and you build up these representations inference [00:29:35.01] [00:29:35.01] is very easy and automatic much more so than than in the in this recurrent kind of [00:29:40.10] [00:29:40.10] strategy because a and c are just showing sharing overlapping representations that makes this um [00:29:45.22] [00:29:45.22] this transitive inference really trivial there's one other um important theory of [00:29:53.12] [00:29:53.12] uh how you solve associated reference which is distinct from these other two [00:29:57.03] [00:29:57.03] and it's called integrative encoding and it says that you study a and b and then when you see b [00:30:03.14] [00:30:03.14] and c it reminds you of the a b experience and you form this kind of conjoined um abc representation [00:30:13.01] [00:30:13.22] that allows you to do the ac lenka test i'm not going to go into the details of this so much [00:30:19.09] [00:30:20.12] but for the purposes of the experiments i'm about to show you it makes the same predictions as [00:30:25.12] [00:30:25.12] the remerge model um which is basically that it's not sensitive to the order of presentation of [00:30:31.11] [00:30:31.11] information whereas this one is okay so um so i'm gonna focus on the the uh remerge kind of [00:30:40.00] [00:30:40.00] strategy and the distributed um representation strategy and both of these strategies are [00:30:46.12] [00:30:47.09] available to this neural network model that i've shown you and that's because it has both of these [00:30:52.15] [00:30:52.15] pathways so um so if you if you don't allow the model to use its recurrent computation um [00:31:01.12] [00:31:01.12] that's fine it can solve the model it can solve the problem in ca1 because it has these [00:31:06.00] [00:31:06.00] representations of a and c that are overlapping so can solve the problem that way but even if [00:31:11.05] [00:31:11.05] you're just using the trisynaptic pathway if you allow the model to use recurrence at retrieval [00:31:15.12] [00:31:16.06] it can also solve the problem that way in the same way that remedge does so it's kind of like [00:31:20.17] [00:31:20.17] the trisometic pathway is um is implementing remerge but then there's this other pathway [00:31:26.00] [00:31:26.00] that can implement this other strategy okay so why do we have two strategies when might you [00:31:30.12] [00:31:30.12] want to use one strategy or the other well this just this strategy should be really [00:31:36.02] [00:31:36.02] fast and efficient um which could be really useful in some kinds of situations but it's going to be [00:31:41.18] [00:31:41.18] susceptible to interference whereas this strategy should be slower it's more effortful i think of [00:31:48.10] [00:31:48.10] it as being a little more explicit you have to like think through um these different experiences [00:31:53.07] [00:31:53.07] that you had in order to solve the inference but it should be resistant to interference um [00:32:00.00] [00:32:00.00] because that localist representation is gonna um is going to uh fight interference problems [00:32:05.22] [00:32:05.22] for you so um so there's differences in the behaviors we expect from these two strategies [00:32:12.00] [00:32:13.16] so uh here's a paradigm that we've developed to try to tease apart these two different strategies [00:32:20.08] [00:32:20.08] um so this is a line of experiments led by my grad student z joe and marley tandock and daria singh [00:32:27.07] [00:32:27.07] in the lab are also involved in this project um so here we're going to do an associative inference [00:32:33.03] [00:32:33.03] task where we present triads um either in an interleaved or in a blocked order [00:32:41.01] [00:32:41.01] so so here there's a um here a b pair and then a b c a matching bc pair in the interleaved case we'll [00:32:49.07] [00:32:49.07] go back and forth between studying the a b and the b c and in the blocked case we'll study all of the [00:32:55.12] [00:32:55.12] abs before any of the bcs but everything is mixed together in these two conditions and then after [00:33:04.04] [00:33:04.04] all that exposure we do two kinds of tests uh the first test is a speeded recognition test where we [00:33:11.16] [00:33:11.16] show participants two objects um and we ask them to quickly make a judgment about whether these [00:33:19.01] [00:33:19.01] two objects were actually presented together um during the exposure phase and so the answer in [00:33:26.02] [00:33:26.02] this case would be no because you know even though they were indirectly related objects they were not [00:33:32.08] [00:33:32.08] actually seen together so um so the answer is no in the explicit inference test so this is a [00:33:39.11] [00:33:40.08] this is like the standard associated inference kind of test we show participants and object and [00:33:45.14] [00:33:45.14] we we make it very explicit we say one of these objects was indirectly associated with this first [00:33:51.01] [00:33:51.01] object via some other object can you remember which of those objects it was um and then we [00:33:57.16] [00:33:57.16] give them plenty of time to make that judgment so we don't expect there to be any difference between [00:34:03.12] [00:34:03.12] the interleaved and the blocked conditions in the explicit inference test because we think that [00:34:08.17] [00:34:09.14] either of these strategies could solve this task and be available and there's no there's no reason [00:34:14.17] [00:34:14.17] to think that that the order presentation is going to make a big difference for that but in the speed [00:34:20.12] [00:34:20.12] and recognition test um we think we predict a difference between between interleaved and blocked [00:34:26.04] [00:34:26.04] conditions and the reason is that if you've built up a distributed overlapping representation of a [00:34:32.21] [00:34:32.21] and c over the course of this exposure then it should be confusing to see a trial like this [00:34:40.04] [00:34:40.04] where you where you're asked to say that you did not see these two things together because if you [00:34:45.12] [00:34:45.12] have overlapping representations of a and c are gonna you're gonna get you're gonna get [00:34:50.17] [00:34:50.17] confused and think maybe you did see them together so we predict you'll either be slower or uh to say [00:34:56.08] [00:34:56.08] that you did not see these things together or that you might even false alarm and say that you did [00:35:01.09] [00:35:01.09] see these things when you didn't if you're using a localist representation which we think you're [00:35:06.00] [00:35:06.13] more likely to be using in the blocked case then you're less likely to have that kind of problem [00:35:12.00] [00:35:14.06] so first these are the data for the the explicit like standard associated inference task [00:35:19.12] [00:35:19.12] um accuracy on this task is higher on this axis for the interleaved um and lower for the [00:35:28.10] [00:35:28.10] blocked condition and we find that there's no difference between these two conditions and it's [00:35:33.09] [00:35:33.09] not that people can't do this they can do this task very well um there's just either just equally [00:35:38.06] [00:35:38.06] good on the in these two different conditions um but in the speeded recognition case we find [00:35:44.08] [00:35:44.08] that participants are slower to reject those interleaved ac's that are confusing [00:35:50.12] [00:35:50.12] relative to blocked acs and they false alarm more so they're more likely to false alarm [00:35:56.15] [00:35:56.15] and think that they saw these acs that they never actually saw together [00:36:01.09] [00:36:04.00] uh we this was an m trick experiment so we decided we would just [00:36:07.22] [00:36:08.23] pre-register it and run run the same thing with more subjects and we find the same results where [00:36:13.22] [00:36:13.22] there's no difference in the explicit um uh test but you're slower for the interleaved condition in [00:36:22.23] [00:36:22.23] this speeded test and you're also more likely to false alarm so we think this is a reliable pattern [00:36:28.12] [00:36:30.10] so so that was an example of a situation where the overlapping representations are are [00:36:36.23] [00:36:38.00] are hurting behavior right because they're making it more confusing to um to make a judgment about [00:36:44.06] [00:36:44.06] whether you've whether you've seen these uh pairs of objects together or not but um we'd [00:36:49.14] [00:36:49.14] like to highlight situations where distributed representations might be useful um and so the [00:36:54.06] [00:36:54.06] next couple experiments are showing you know these are the kinds of things that we think that [00:36:59.07] [00:36:59.07] the building of this representation might do for you that would be that would be useful so in this [00:37:04.19] [00:37:04.19] experiment we did the same um same exposure either interleaved or blocked presentation of these pairs [00:37:10.17] [00:37:10.17] um and then we did a generalization kind of test so we asked uh or we first taught people [00:37:16.12] [00:37:16.12] about novel properties of some of the objects so we might um say that this turns out that this [00:37:22.10] [00:37:22.10] object fleeces whatever that means and then we ask people which uh which other objects do they think [00:37:29.09] [00:37:29.09] also have that property and we test to what extent they'll generalize that property to the indirectly [00:37:35.07] [00:37:35.07] related items as a function of the order of exposure and we find that the that they're [00:37:42.06] [00:37:42.06] they're more likely to do this generalization um for the interleaved acs which is consistent with [00:37:49.03] [00:37:49.03] this idea that you're building up this distributed representation that's so useful for generalization [00:37:54.12] [00:37:55.20] um the if you break this down separately into the interleaved and blocked conditions [00:38:00.06] [00:38:00.06] it looks like this is working for the interleaved case but not at all for the blocked [00:38:04.02] [00:38:04.02] case where there's no they're not doing this generalization at all across these indirect pairs [00:38:08.15] [00:38:10.15] um and then in this last experiment we thought well are there scenarios where [00:38:15.11] [00:38:15.11] we can make this effect um even bigger like are there scenarios where there's no hope that the [00:38:22.06] [00:38:22.06] local style representation could ever um solve this task for you basically and so we thought [00:38:28.08] [00:38:28.08] well what if we take this the same paradigm and instead of asking people to explicitly remember [00:38:34.08] [00:38:34.08] memorize that these two pairs of these two objects go together in a pair [00:38:38.00] [00:38:38.00] we'll turn this into a statistical learning experiment and um and ask people to [00:38:43.20] [00:38:44.17] both kind of parse the sequence and find where the pairs are and and understand these indirect [00:38:50.17] [00:38:50.17] relationships and this is the kind of situation that we think um distributed representations are [00:38:56.15] [00:38:56.15] really crucial for it's unclear that localist representations would can solve this task at all [00:39:02.08] [00:39:02.08] as you saw in the in the neural network model in the tri-synaptic pathway and here indeed we find [00:39:09.18] [00:39:09.18] that even if we do this explicit inference test um the kind of standard uh associated inference test [00:39:16.15] [00:39:17.11] the accuracy for interleaved acs in this paradigm is much better than for blocked um [00:39:24.06] [00:39:24.06] and there's no evidence that people can learn the structure now in the blocked case [00:39:28.06] [00:39:31.18] okay so um so interleaved excuse me [00:39:35.09] [00:39:39.09] interleaved learning seems to benefit rapid inference we think this is [00:39:42.21] [00:39:42.21] consistent with the idea that the hippocampus might contain these distributed representations [00:39:48.10] [00:39:49.03] that complement they're separate from these um pattern-separated localist representations that [00:39:54.23] [00:39:54.23] we still think are very important for um avoiding interference in in episodic memory okay i'm happy [00:40:02.17] [00:40:02.17] to like pause and take any questions at this juncture if if that's if that's useful um [00:40:09.20] [00:40:10.21] otherwise i'll um just spend a few minutes on on our sleep work [00:40:16.23] [00:40:18.23] okay i don't see anything yeah there's no question in the chat right now this is dobby [00:40:23.20] [00:40:23.20] there's no questions in the chat so maybe you can just keep going okay i will power power on okay um [00:40:29.07] [00:40:30.06] so i'm gonna tell you a little bit about what we're thinking about the next step here so [00:40:34.06] [00:40:34.06] um what happens after these new regularities that you're learning in the hippocampus are encoded how [00:40:40.02] [00:40:40.02] do you what does this transformation look like so if you have no more exposure to information [00:40:45.14] [00:40:45.14] in the environment it seems like there's still a lot of learning and consolidation that's happening [00:40:50.06] [00:40:50.06] how does that happen um and one one kind of prediction of this framework this way [00:40:57.05] [00:40:57.05] of thinking about things is that over over time and in particular over sleep that there should [00:41:04.08] [00:41:04.08] be this emergence of even more overlapping distributed representations in neocortex [00:41:09.16] [00:41:10.17] even without further exposure to information in the environment and so we ran a behavioral [00:41:19.07] [00:41:19.07] experiment using these satellite stimuli that i showed you earlier these object category learning [00:41:24.10] [00:41:24.10] um stimuli where participants are learning about three categories of these objects where objects in [00:41:30.23] [00:41:30.23] the same category share most of their features but there's also unique individuating features [00:41:35.18] [00:41:36.12] and what we find is that over the course of um of 12 hours that includes either sleep or wake [00:41:43.18] [00:41:44.19] unique features of the individual satellites will be um maintained over a night of sleep [00:41:52.02] [00:41:52.02] but forgotten over a period of the same amount of time spent a week and this is [00:41:57.22] [00:41:58.12] consistent with a lot of existing literature on kind of arbitrary pairings and the ability of [00:42:04.19] [00:42:04.19] sleep to prevent um the forgetting of that kind of information but in the case of shared features so [00:42:11.07] [00:42:11.07] the features that are shared across members of a category we find that there's actually this above [00:42:16.06] [00:42:16.06] baseline improvement in your ability to remember these features across the night of sleep [00:42:21.03] [00:42:21.03] which suggests that sleep is promoting somehow a better understanding of the shared structure even [00:42:26.08] [00:42:26.08] without having more exposure to that structure in the environment and so we think this is consistent [00:42:32.15] [00:42:32.15] with this idea that if you're building up this offline um distributed reputation a neocortex that [00:42:39.14] [00:42:39.14] it it's going to highlight the shared structure for you and and make you understand that structure [00:42:45.16] [00:42:45.16] even better um we also found that um we ran a fmri experiment with this paradigm looking for [00:42:54.10] [00:42:54.23] um offline reactivation of these um of these satellites to see how it relates to changes [00:43:01.16] [00:43:01.16] over time um we're now doing um experiments in with mg trying to actually look at sleep replay [00:43:08.12] [00:43:08.12] but for this for this first experiment it was a fmri experiment where we where we stuck with awake [00:43:13.22] [00:43:14.17] replay and we found that awake reactivation of these objects in the hippocampus [00:43:19.07] [00:43:20.04] predicts improvement in your ability to remember them after sleep so we think that what we were [00:43:25.18] [00:43:25.18] measuring in these awake periods is kind of representative or maybe influences what [00:43:30.08] [00:43:30.08] continues to happen um offline during sleep and that sleep is is then especially useful for um for [00:43:38.06] [00:43:38.06] these uh these these um memories so um so with the the reactivation that we observe results in [00:43:47.11] [00:43:48.06] more behavioral improvement over time if if that time includes sleep so how um how does [00:43:57.12] [00:43:57.12] replay actually shape cortical representations this feels like kind of a mysterious thing and [00:44:04.12] [00:44:04.12] it's really a difficult computational problem because usually our models rely on input and [00:44:12.04] [00:44:12.04] feedback from the environment in order to build up useful representations so how do you um how is [00:44:17.12] [00:44:17.12] it possible that just this like offline learning based on already existing representations could do [00:44:22.15] [00:44:22.15] something so useful and so here we're again doing uh neural network modeling where we have a model [00:44:31.03] [00:44:31.03] that can sleep it has a hippocampus that has a cortex and we're going to try to understand what's [00:44:36.02] [00:44:36.02] happening in interactions between hippocampus and cortex to produce these changes offline [00:44:41.09] [00:44:43.03] so um learning during sleep like i said is a very tricky computational problem we've been developing [00:44:49.09] [00:44:49.09] kind of new learning schemes that might allow um this to work like learning from your existing [00:44:55.09] [00:44:55.09] representations to do something useful so here's a here's a video of the model um sleeping it can it [00:45:02.21] [00:45:02.21] can move from from memory to memory each of these different um like strong yellow looking states is [00:45:08.12] [00:45:08.12] a particular satellite and we get the model to move so it it behaves completely autonomously [00:45:15.11] [00:45:15.11] we set it running with some random initial um uh activity and then it just runs completely on its [00:45:22.10] [00:45:22.10] own and it moves from memory to memory uh because we have some short-term synaptic depression that [00:45:28.00] [00:45:28.00] forces uh co-active uh pairs of units to kind of um tire out and um and move to the next memory [00:45:35.12] [00:45:36.04] um and then we use we have them the model track its own stability and when states are very stable [00:45:42.04] [00:45:42.21] they get they get marked as good states or plus states for learning so like this state here [00:45:48.10] [00:45:48.10] that's a stable state so it gets marked as a good state and then we have oscillations um that are [00:45:55.09] [00:45:55.09] perturbing these states and helping to reveal kind of aspects of these memories that need some help [00:46:03.11] [00:46:04.08] and oscillations are very prominent um perturbation in the sleeping brain and so [00:46:09.01] [00:46:09.01] we can kind of make use of that and the way this works is that as also as the um as we oscillate [00:46:15.01] [00:46:15.01] levels of inhibition in the model that that controls the the amount of overall activity [00:46:19.16] [00:46:19.16] that's possible and when we raise the inhibition above baseline um that reveals parts of a memory [00:46:27.05] [00:46:27.05] that are weaker and in need of strengthening and when we lower inhibition below baseline it reveals [00:46:34.10] [00:46:34.10] who is connected to this memory and potentially interfering with it by having activation spread [00:46:39.22] [00:46:39.22] farther than normal and then we can contrast our stable states with these [00:46:45.18] [00:46:46.12] perturbed states from the oscillations and then we can do error-driven learning in this case we're [00:46:51.05] [00:46:51.05] going to do contrast of heavy learning which is going to allow us to do some something useful [00:46:57.16] [00:46:57.16] from these existing representations we also have sleep stages in the model we know that [00:47:03.09] [00:47:03.09] the hippocampus and cortex are especially uh well connected and communicating during slow wave sleep [00:47:09.07] [00:47:09.07] and then during rapid eye movement sleep seems like these systems are relatively decoupled and we [00:47:14.21] [00:47:14.21] kind of let the neocortex run by itself during uh rapid eye movement sleep period and what we find [00:47:21.20] [00:47:21.20] is that over the course of the model's sleep it builds up these neocortical representations where [00:47:27.14] [00:47:27.14] objects from the same category that the shared parts of those representations are enhanced [00:47:33.03] [00:47:34.02] but also the unique parts of the representations are are preserved um and so we're very excited [00:47:39.14] [00:47:39.14] about that because it matches up with what we see in our behavioral data where you're able [00:47:43.05] [00:47:43.05] to remember unique properties but there's this um there's this enhancement of the shared structure [00:47:48.04] [00:47:50.12] um and in my last minute here i'll just um tell you about what we're up to now or as as soon as [00:47:57.01] [00:47:57.01] they let us back in the lab um we have uh this targeted memory reactivation study that we're [00:48:03.20] [00:48:03.20] running where participants are again learning about these three categories of these objects [00:48:10.06] [00:48:10.06] but we're actually going to have people listen to the names of the objects as they're learning about [00:48:16.12] [00:48:16.12] them so they'll hear these names um and then during sleep we're going to play these names [00:48:25.03] [00:48:25.03] um at at the peaks of these oscillations which is this the slow oscillation which is the time that [00:48:31.09] [00:48:31.09] we know that replay is kind of likely to happen um with the idea that we can try to have some control [00:48:38.13] [00:48:38.13] over what participants are replaying and this is this targeted memory reactivation technique is [00:48:43.12] [00:48:43.12] sounds kind of science fictiony if you haven't heard of it but [00:48:46.04] [00:48:47.05] but has been used successfully in many many sleep studies now so it's a really exciting way to have [00:48:52.12] [00:48:52.12] um some some causal control over what's happening in human sleep and we're trying out different [00:49:00.06] [00:49:00.06] manipulations but one of the things we're really excited about is is playing these um hues in [00:49:07.05] [00:49:07.05] either interleaved or blocked order if we think that replay is offline replay is most useful when [00:49:14.08] [00:49:14.08] information is interleaved well can we show that you really extract the shared structure best if [00:49:19.01] [00:49:19.01] we encourage that replay to be interleaved okay so that's just a sense of what we're up to now [00:49:25.05] [00:49:25.20] um and let me just sum up overall so so i've told you that we think the hippocampus might [00:49:31.05] [00:49:31.05] contain these distributed representations um that are complementing the the well-known localist [00:49:36.12] [00:49:36.12] representations but more broadly we think this is part of a broader kind of continuum where during [00:49:42.19] [00:49:42.19] sleep the hippocampus is helping the neocortex to further enhance this shared structure with even [00:49:48.17] [00:49:48.17] more distributed representations that are kind of being built more slowly over time over the course [00:49:54.21] [00:49:54.21] of of long-term consolidation okay i'll stop there and take any questions if you have them thank you [00:50:02.00] [00:50:04.17] all right thank you anna if you if some of you want to unmute and [00:50:09.09] [00:50:09.09] join me in thanking anna for her talk i'm sorry for this crazy flashing of my background [00:50:16.04] [00:50:22.04] yeah it's i think for most of us [00:50:24.10] [00:50:25.14] your picture is sort of small so the background didn't okay okay all right so um [00:50:32.00] [00:50:33.01] if uh if anybody wants to ask a question you can just unmute and maybe just ask it directly [00:50:39.20] [00:50:47.09] all right uh while everybody's thinking about questions maybe i can ask my my question um so [00:50:53.03] [00:50:53.03] that was that was super cool really like a ton of interesting stuff so thank you so much for that [00:50:58.17] [00:50:58.17] um i have a question about some of the um stuff that you talked about at the very end so you're [00:51:05.09] [00:51:05.09] talking about that model of sleep and you're saying that um some states that the stable get [00:51:13.05] [00:51:13.18] marked as good and some states that they're unstable or also having it marked as is bad [00:51:19.05] [00:51:19.22] can you talk a little bit more about that like mechanistically how how is this who is doing the [00:51:24.19] [00:51:24.19] marking how is that marked um and how is that used later yeah yeah so um so the way it works is that [00:51:34.10] [00:51:34.10] um when you first fall into an attractor um the pattern of activity is very stable so it makes [00:51:42.02] [00:51:42.02] it so that we can just track this one variable which is how much is activity changing from one [00:51:47.14] [00:51:47.14] time point to the next um and i you know we don't have a we don't have like a particular [00:51:53.07] [00:51:54.08] uh candidate mechanism in mind for like what what exactly we think would be implementing [00:52:00.10] [00:52:00.10] that but there's i think there's a few options for how neurons can be sensitive to how much [00:52:05.11] [00:52:05.11] change there is on a short time scale like that so we assume that something like that is possible [00:52:10.00] [00:52:10.21] and then we just say okay so as you fall into an attractor things are very stable we assume [00:52:15.20] [00:52:15.20] that you can mark that that's true um and then as synaptic expression starts to take [00:52:20.15] [00:52:20.15] hold and the attractor becomes less stable the oscillations which are always present um will [00:52:26.02] [00:52:26.17] will will become stronger basically um just because of that action of the synaptic depression [00:52:33.14] [00:52:33.14] so oscillations get bigger and so what we can do with this very simple thing where we just compare [00:52:38.04] [00:52:38.04] that initial state to the rest of what happens until things drop below some some you know some [00:52:46.19] [00:52:46.19] threshold that we say okay we're probably not replaying a memory anymore um and so just by [00:52:51.12] [00:52:51.12] tracking that one variable that's enough to say we can now contrast the stable state to everything [00:52:57.07] [00:52:57.07] else which includes both sides of the oscillation right includes both the showing the revealing the [00:53:04.10] [00:53:04.10] weak parts that need strengthening and revealing the competitors that need weakening that all [00:53:10.10] [00:53:10.10] gets baked into the same like bad minus state and the math works out so that we can just contrast [00:53:16.13] [00:53:16.13] that with the stable state and that's all you need so you just have to be able to track stability [00:53:20.10] [00:53:20.10] and then compute that contrast um and then that's enough to get the learning working in [00:53:25.22] [00:53:25.22] the directions that you need awesome um and to the extent which now you have a model of [00:53:33.01] [00:53:33.01] like the different stages of memory do you think you can play with like what would happen if you [00:53:39.14] [00:53:39.14] just interfere with rem sleep or what will happen if you just interfere with some other [00:53:44.00] [00:53:44.00] part of the sleeve hey yeah yes definitely the model can probably do and like make predictions [00:53:50.02] [00:53:51.09] yeah we're very interested in that one of the things that we're interested in is [00:53:54.08] [00:53:55.03] the idea that cycling between slow wave and ram might be important um so that's yeah it's hard [00:54:01.12] [00:54:01.12] it's hard to have um to have experimental control over like a sleep stage cycling or the pro having [00:54:09.16] [00:54:09.16] you know having or not having particular sleep stages but um but yeah we can explore that in the [00:54:14.02] [00:54:14.02] model and one idea that we've been thinking about is that maybe cycling between slow wave and ram [00:54:20.10] [00:54:20.10] over the course of one night is really useful for integrating new information into existing [00:54:26.12] [00:54:26.12] knowledge structures so if you assume that um that's that cortex contains kind of your [00:54:31.11] [00:54:31.11] long-term memory of everything whereas the campus might be more biased towards things that have [00:54:35.01] [00:54:35.01] happened more recently then um then effectively what you're doing by going back and forth between [00:54:40.10] [00:54:40.10] slow wave and ram is thinking about things that happen more recently and then thinking about [00:54:45.14] [00:54:46.12] the rest of what you know right and that that's another kind of interleaving process [00:54:51.05] [00:54:51.05] that might be really useful for integrating new information into existing knowledge and so we can [00:54:56.04] [00:54:56.04] build models where we can build simulations that um that either have that interleaving or don't [00:55:01.03] [00:55:01.03] have that interleaving and um and it does indeed seem to be the case that in that kind of situation [00:55:06.21] [00:55:06.21] where you need to integrate new knowledge into existing knowledge um that that going that [00:55:12.13] [00:55:12.13] sleep cycling is really important so yes we're very interested in that super cool so let me [00:55:20.21] [00:55:24.13] provide the opportunity for somebody asked us somebody else to ask questions [00:55:28.04] [00:55:28.23] this is uh mike hunter uh thank you for the talks uh really interesting to see what's going on here [00:55:36.00] [00:55:36.21] i'm going to ask what's probably a super ignorant question so forgive it [00:55:40.08] [00:55:41.07] uh so you talked about uh kind of monosynaptic pathways and trisynaptic pathways is there [00:55:47.20] [00:55:49.07] any reason to limit to those options now you can have four or five or 132 synaptic pathways [00:56:00.13] [00:56:00.13] if you want you know the model will work the same if less efficiently what is the [00:56:07.09] [00:56:08.13] justification for this kind of limited set [00:56:11.14] [00:56:13.22] uh yeah so i i actually don't care about the number of synapses in particular um [00:56:20.08] [00:56:20.08] we i when i use those terms i'm i i just use them as a way of referring to two different pathways in [00:56:26.02] [00:56:26.02] the pro in the hippocampus that have different properties so um so it could have been that the [00:56:31.14] [00:56:31.14] hippocampus had three synapses on the pathway that had a slower learning rate and and more [00:56:37.07] [00:56:37.07] overlapping representations but that's not the way it was built and you know so we built the model [00:56:41.01] [00:56:41.01] to correspond to the the known anatomy of the hippocampus so the only thing that i care about [00:56:46.10] [00:56:46.10] is those two properties how much overlap is there and what's the learning rate and those are the [00:56:52.04] [00:56:52.04] two properties that that vary across those two pathways and you know there might be reasons we [00:56:57.16] [00:56:57.16] could talk about for why the hippie campus has set up three synapses on its on the sparse high [00:57:03.14] [00:57:03.14] learning rate pathway there might be a need for multiple transformations of the representation [00:57:09.05] [00:57:09.05] there that's not as important for statistical learning kind of strategy um but yeah i don't i [00:57:15.01] [00:57:15.01] don't mean to suggest that the number of synapses is important to the theory it's not thank you [00:57:25.20] [00:57:26.13] i have a question too um by the way it was a super cool talk and i i really enjoyed it and thank you [00:57:33.20] [00:57:33.20] so much um so my question is um so it seems like the hippocampus has these two different pathways [00:57:41.16] [00:57:41.16] of learning like doing statistical learning and encoding things um episodic memories [00:57:48.23] [00:57:48.23] and i i was wondering whether those two systems can work together or um is there any kind of [00:57:58.02] [00:57:58.02] like priority that one happens first than the others yeah yeah that's a really great question [00:58:05.07] [00:58:05.07] something we're thinking about a lot right now um so i showed you that there was look what looked to [00:58:10.12] [00:58:10.12] be trade-offs between these two pathways and that brings up a really interesting question which is [00:58:16.10] [00:58:17.07] might we have control over the action of these two pathways um we know that there's fluctuations in [00:58:24.08] [00:58:24.08] the um in the strength of the two pathways as a function of um of theta oscillations and as [00:58:32.04] [00:58:32.04] a function of the amount of acetylcholine in the hippocampus and so it's plausible that actually [00:58:38.08] [00:58:38.08] you know depending on what you're trying to do you might be able to root information information more [00:58:42.15] [00:58:42.15] strongly through one pathway or the other which would be very useful if you're you know if you're [00:58:46.15] [00:58:46.15] in a situation where you're trying to generalize or you're in a situation where you really want to [00:58:50.02] [00:58:50.02] focus on the specifics of what you learned it'd be great if you could have some control over that and [00:58:55.12] [00:58:55.12] so we're interested in the possibility that medial prefrontal cortex which has direct connections to [00:59:00.19] [00:59:00.19] to ca1 might be able to like help root these um connections through these two pathways [00:59:06.19] [00:59:07.09] um but you asked if there was like a a difference in different phases of learning is that what you [00:59:12.15] [00:59:12.15] said between the two pathways yeah uh yeah i mean we we don't we haven't seen we we [00:59:20.10] [00:59:20.10] haven't implemented that in the model and i don't think we've seen evidence [00:59:24.23] [00:59:26.02] for that in data that i can think of other than the fact that there's a developmental [00:59:31.03] [00:59:31.16] difference right where you seem to start off with more a reliance on statistical [00:59:36.15] [00:59:36.15] learning and then over time add up in the episodic memory but that's that's different [00:59:40.19] [00:59:40.19] that's over the course of development so yeah we haven't we haven't explored all right thank you [00:59:46.10] [00:59:54.19] all right okay one last question thank you um so i have a question about the um sleeper effect [01:00:04.02] [01:00:04.02] uh so that's related to the last part of your presentation um i got quite interested with that [01:00:11.11] [01:00:11.11] sleeper effect because it almost seems like um people didn't pay attention their brain actually [01:00:18.04] [01:00:18.04] at rest but actually during that process they learned something which kind of like associated [01:00:24.06] [01:00:24.06] with the last part of your presentation about you know the um when they are shared um structure [01:00:32.23] [01:00:32.23] people can um replace the car better through sleep when they wake up so i kind of wonder whether [01:00:40.15] [01:00:41.16] they two things are actually correlated or um it's just like uh illusion are you asking whether we [01:00:52.19] [01:00:52.19] know that the reap that replay during sleep is like is responsible for the memory changes that [01:00:58.04] [01:00:58.04] we see yeah yeah um i would say the evidence is pretty strong at this point for that um [01:01:06.08] [01:01:06.08] the strongest evidence probably comes from rodent work where they actually [01:01:10.04] [01:01:10.04] can well they see extremely clear evidence of replay um which is of course harder for us [01:01:16.06] [01:01:16.06] with our methods and humans and you can disrupt that replay you can like specifically see it [01:01:22.15] [01:01:22.15] and and stop it in rodents and find that that um that that impacts memory consolidation so [01:01:28.23] [01:01:28.23] um so i would say we have a we have a pretty we have pretty strong evidence at this point [01:01:34.13] [01:01:34.13] that um that replay is causally involved in the memory consolidation changes that we see [01:01:38.17] [01:01:44.13] all right great well thank you again anna that was that was wonderful [01:01:48.13] [01:01:48.13] and we have some more meetings for you with faculty and students uh this afternoon [01:01:53.09] [01:01:55.03] wonderful well thank you again and uh hopefully uh we'll see you around during [01:02:01.18] [01:02:01.18] non pandemic yeah yes yeah perhaps me come back when in prison someday great thanks everybody [01:02:18.15]