Numenta Numenta Research Meetings, 9 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Max Bennett: An Attempt to Model The Neocortical Microcircuit in Sensory Neocortex - Sept 9, 2020

Description

This week, we invited Max Bennett to discuss his recently published model of cortical columns, sequences with precise time scales, and working memory. His work builds on and extends our past work in several interesting directions. Max explains his unusual background, and then discusses the key elements of his paper.

Link to his paper: https://www.frontiersin.org/articles/10.3389/fncir.2020.00040/full#h1
- - - - -
Numenta is leading the new era of machine intelligence. Our deep experience in theoretical neuroscience research has led to tremendous discoveries on how the brain works. We have developed a cohesive theory, core software technology, and numerous software applications all based on principles of the neocortex. Our innovative work delivers breakthrough capabilities and demonstrates that a computing approach based on biological learning principles can do things that today’s programmed computers cannot do.

Subscribe to our Weekly News Digest for the latest news about neuroscience and artificial intelligence:
https://tinyurl.com/NumentaNewsDigest

Subscribe to our Monthly Newsletter for the latest Numenta updates:
https://tinyurl.com/NumentaNewsletter

Our Social Media:
https://twitter.com/Numenta
https://www.facebook.com/OfficialNumenta
https://www.linkedin.com/company/numenta

Our Open Source Resources:
https://github.com/numenta
https://discourse.numenta.org/

Our Website:
https://numenta.com/

A

All right we're good.

B

A

Okay, so uh really happy to have max bennett here uh talking to us about himself and his a new paper he just published. uh That extends uh some of the work we've done and takes it into some new areas. So I'll let them talk about it.

C

Cool, so let me share my screen. Can everyone see my screen? Yep? Yes, yes, so a little background of me uh my day, job has nothing to do with neuroscience whatsoever. um I started a company called blue core eight years ago, which does personalization for retailers, so we apply machine learning to help companies like nike or under armour, make personalization decisions on their websites.

C

So all the product recommendations you see all the emails you get from brands like that. We are powering a lot of the decisioning behind that it's been a crazy eight years, we're now 240 employees on three continents about a third of the top hundred retailers in the us use our products. um So it's been very, very fun uh working on the entrepreneurial stuff uh two years ago, um just out of intellectual interest.

C

I stumbled on jeff your 2009 paper and then your folks's 2016 paper um and I just kind of became very inspired by the approach which was sort of trying to decipher the underlying computational roles or algorithms based on the observed biological circuitry.

C

So my passion project for last two years has simply been working on this reading, a ton and then someone sort of challenged me to take some of the ideas I was working on and try and publish it. um So that's sort of the history. It's not my my background by any means, but it is something I'm very interested in. um So I'm very excited to be here. I'm humbled to be here um very open to feedback and thoughts, and hopefully some of these ideas will be interesting or useful.

B

So had you done any any neuroscience reading before what you were the two years ago.

C

Nope, it's all been two years of.

B

You've read a lot in last year, so it's impressive. You know the the all the references and the how much you've picked.

C

Up thanks well, I was inspired to see you know, I think, there's a lot of people that feel boundaries, that you need some sort of degrees to do things, and you know people who pave the path of doing this without those degrees make it easier for people like me to feel safe, and I think I can do it. So I appreciate that cool.

A

Should we dive in.

C

Yep that'd be great okay, so I'm gonna try and keep the motivation uh short I'll, keep it less than 60 seconds I'll, mostly do it for people watching on youtube. I assume I'm going to be preaching to the choir on a lot of the motivation stuff, so I'll try and skip past it as fast as I can um so high level. The brain can do a bunch of things that people have not figured out how to do in state-of-the-art machine learning um the five that are particularly interesting to me um online or continuous learning.

C

uh You know the brain is clearly far superior at that than any of the state of our algorithms. Far better generalization. One-Shot learning is something that we're way better at than state about our vision, algorithms or any alternatives, sensory motor learning. We still don't have robots that can do the dishes. That's too complex a task which is kind of silly there's, clearly amazing, sort of economic benefits we would get if we could even just have robots that would do tasks like that um that efficiency.

C

Obviously the brain works on the the energy uh of a light bulb. So in terms of trying to understand what the brain's doing, obviously, there's a lot of interest in doing that. The neocortex is where I would say, like 90 of the sort of smart brains are working on, and it makes sense why? Because it seems to be where all the interesting stuff happens. So uh my paper is similar to what you folks are working on trying to decipher what is the uh neocortex doing.

C

um Obviously, in 1978 vernon mount castle came up with this great idea. Over the last 40 years. There's been a lot of support for this idea, which is there seems to be a repetition of the same canonical micro circuit and the reason this is. You know such an exciting concept. At least this is true in sensory cortex there's some debate over frontal cortex, there's certain different neural types and all that stuff which we'll not talk about here, but the reason.

A

I don't think there's a single statement in neuroscience: that's not controversial.

C

Yeah yeah, I agree that almost everything.

A

Is complete seems to be controversial.

C

Totally exciting totally you're absolutely right, um but I would say frontal cortex seems to be more controversial, like I'm very aware of neuron types like von economo neurons that are present in frontal cortex that aren't present sensory cortex. There are no portions of sensory cortex that are missing layer. Four, there are lots of egg granular areas of frontal cortex, but yes, there are people that disagree that even there is a canonical micro circuit.

B

Well that you know that doesn't mean that mount castle is wrong. It just means that there's variations on the theme- and I think even thus, people like ourselves, who are big fans of mount castle's idea- will admit that there are variations on a theme, but there seems to be a lot in common, even when there are differences. So that's.

C

B

C

Totally um cool, so the the amazing idea here is: it simplifies the task of trying to understand trillions of synapses in the cortex just trying to decipher the micro circuit. um There's a bunch of people working on this.

C

My goal is not to review them, but I think there are some really interesting concepts from predictive coding and adaptive resonance theory that may or may not be applicable to htm and sort of you can think about parts of my paper as trying to cherry pick ideas from these other models of what's going on in the canonical microcircuit, and see how we can augment html to include portions of that.

C

The thing that seems to be agreed upon the theme between all of these is that the macro column in some sense seems to be an unsupervised prediction machine now: there's lots of sort of disagreements on the special features in them, but there seems to be broad agreement that it's somehow predicting its input.

C

um So the model I put together tries to propose how one simple circuit, uh somewhat simple circuit, can do all six of these things um and it builds on top of a lot of the folks that you folks have been working on as well as add some additional things, so I'll sort of dive right into it. So the first thing I did, which was time consuming, was to map all of the neurobiologically available information on what the actual connectivity is.

C

The question marks here are inferred, so, for example, there's lots of evidence that higher order thalamus projects to layer 5a, I have not found any evidence of specifically which neurons it projects to. So I try to be transparent about um the observed connectivity versus the inferred connectivity what's interesting is there seems to be six main types of excitatory neurons.

C

There are more so this is not comprehensive, um but there are six main types of excitatory neurons that seem to have consistent genetic markers, so the neurobiologists can regularly find, for example, these l5 ars neurons throughout all of century cortex, and they seem to always have this the same genetic marker and they seem to always project to the strain as an example.

C

So I took that also tried to incorporate the data on the relative number. So, as you said, jeff you're 100 right there are variations in the relative ratios of these different neuron types in different areas of sensory cortex.

C

There seems to be a consistent theme, though, that the two most numerous types of neurons are these l23 pyramidal neurons and these lc6a corticothalmic neurons. So I tried in our sort of conceptual model here, to incorporate the relative counts.

C

The most infrequently found neurons seem to be these l5 neurons, and perhaps the relative numbers of these neuron types are instructive into the computations being performed. So in the model I'm going to walk through I'm going to use this sort of conceptual model. Of course, there are more mini columns and I'm presenting here I'm sort of simplifying I'm going to incorporate activated neurons by denoting them by more filled in colors um and uh you'll, see where I sort of take it. So, let's.

A

Talk about sequence,.

C

Prediction so there's two canonical computations I'm going to invoke over and over and over again, um so I think it's useful just to make sure we're aligned on how this work one is directly from your folks's 2016 idea. I think, what's cool is this concept I think applies outside of just layer, two three. um So I'm gonna apply the same idea that if you get dendritic spikes and you put neurons in a sub threshold, depolarized state when they're in a winner take all network, and then you activate them with proximal input.

C

You will get a different, sparse pattern and I think that concept might apply in other components of the micro circuit.

B

But we, just if you don't mind just jumping in we make that assumption too. We assume that pretty much all parameter neurons would have this property, um although we did not have not exploited it yet, but we think this idea there's two different types of representations uh that can occur, this sort of predicted state and we think we agree with you. We think everywhere. You see, I think everywhere you see stellite cells and pyramidal cells. This principle applies.

C

Yep and what's cool also, you there's a lot of evidence that also happens in thermocortical relay cells, where modulatory distal input puts them in these sub-threshold predictive states which make them burst instead of uh atomically fired, which I think also might have impact um cool. So one is this idea. The other idea I'm going to regularly invoke is this concept of pattern separation, which is the idea that uh somehow networks of neurons are capable of decorating overlapping input um and there's been a lot of people that have modeled how this can happen.

C

But the idea is, if you have a sub pattern, so let's say this represents some object x. If you have a sub pattern, that's fully overlapping.

C

You can have neurons it's connected to decorate that output, in other words, separate the pattern. So there's a lot of people that have proposed how this might happen. It can happen if you're sparsely projecting to these neurons, as well as to uh inhibitory interneurons in a random way. You can get pattern separation.

C

I don't describe the exact mechanism, but I look for situations where it seems to be the case that one um network of neurons projects sparsely to another network and somewhat equally, to inhibitory neurons, as well as excitatory neurons as evidence that something like pattern, separation might be happening.

B

I mean, wouldn't you argue that it's already separated in that uh sparse representation in the labor two three cells as you're, showing there I mean. Wouldn't you say, that's already separated those. If I look at the actual neurons that are active, not the many columns they're separate patterns, um so um I'm just trying to get your language here, because I think what you're actually saying you want to be able to recognize it elsewhere or something like that.

C

Yes, so the key thing um I'm trying to get to is: how does the fact that this sub-pattern y might mean something completely different than this overlapping subpattern z? How is that communicated from the macro column to other places in the brain.

B

I mean, if I just if I just project those cells they're completely different set of cells, and so I would I would uh that would be obvious. There.

C

Too right, um yes, you could 100 do that which which for sure might happen- l23, cortico, cortical, neurons for sure yeah, but I guess what I think is interesting. Is I'm following the connectivity to these these neurons within the macro column, that projects to subcortical areas such as these layer, five v neurons or what are they doing? So? It's not the case that uh this network cannot perform the type of computations you're saying, but I think there might be other layers that are taking that pattern and somehow synthesizing it and passing out.

C

Does that make sense.

B

Yeah yeah- I just I think I know where you're going in this for the regular paper, but I just want to make point that that that there is these are already separated um and layer. Two three cells do project, but not subcortically right. They project other columns so um but you're trying to get to a subcortical projection.

A

Yep, maybe one advantage of this lower representation. Is it's much lower dimensionality? So there are the number of axons you have to project to communicate that information could be smaller and.

B

Also, I think the efficiency.

A

Argument there.

B

A

B

I also think, where you're going matches you're going to argue that these red cells, here these layer, 5 cells, are going to become a form of stable pattern over the sequence. Yes, which is, which is what we call temporal pooling, which we've also modeled, but not in specifically that layer themselves. Yeah.

C

Yeah you're already you're already heading the curve. Okay, so I uh I follow the connectivity that you folks do. I propose that layer, four cell light cells perform simply coincidence, detection a lot of evidence of this. um If you just record lgn neurons their response responsive to center, surround receptive fields in your eye layer, four neurons are responsive to lines of specific orientations and similar locations.

C

Of course, that conceptually is just coincidence. Detection you see this in other cortical areas as well.

C

Similarly, to how you folks model, I use pretty much the same exact idea that these layer, core stellite cells, project uh in a mini column, um so an object, a whether it be you know, pattern of somatosensory input, a pattern of visual input. uh Auditory input will be represented by a population of many columns being activated so now. The interesting thing is time.

B

Can I go back to keep aside? I don't want to delay things with you. um It's interesting. You chose to represent layer 4 as a single cell and in some sense we do our what we call our spatial pooling, which is this activation of minicoms. But we assume it's an inhibitory cell. That does that we assume it's the uh the bipolar or double decay cells that actually enforce that mitochondria become active.

B

I don't know if you know that or not, but but because you know there are a lot of layer, four cells, yeah and um there's quite a few actually in in many parts of the cortex, so you've sort of abstracted those way to to what we would call the spatial cooler, but um but we actually think that role is is performed by different cell uh inhibitory cells. But it's just.

A

Pointing out the.

B

Differences here, because you've kind of abstracted away any sort of a very what we consider a meaningful role of layer, four got it that makes sense, but it's okay. We make lots of assumptions.

C

No, it's good, it's good, um okay, so times this is where I think things get really interesting and I love the general concept of htm, which is that what all these other models don't incorporate, which is true of most predictive coding models, as well as adaptive resonance theory. Is this idea of time so intuitively you know. If I look at this picture my mind immediately projects forward in time, I can almost see her kicking the ball.

C

A more potent example of this that I love doing for people is one shot, sequence learning. So I'm going to play three notes: it's not going to be magical music! Then I'm going to play the same two notes, not the third one and your brain will instantaneously in your head, be able to hear the three. The three notes.

C

It's like it's trivial, it's a joke, but it's actually fascinating that your brain just did that you one shot learned an exact sequence. Interestingly, also your brain can one shot. Do complete, can disambiguate between fully overlapping sequences. So I'm going to play uh two sequences of notes. The middle three notes are completely exactly the same. The only difference is the first note and the last room. Then I'm going to play one of these two sequences and omit the last note, and I guarantee every single one of you will easily be able to tell which.

C

C

It's obviously a high note, so how did your brain just do that it was you weren't, even thinking about it, and your brain did that. So I'm like super fascinated in how that was so trivial and maybe that's somehow insightful into one-shot learning.

C

So the three features of this one-shot sequence- learning I'm really interested in is the one-shot learning of sequences, disambiguating overlapping sequences and one-shot prediction of the next element. um So, let's start by just rapidly inputting abc I'm going to assume there's no delay between them. That's something I'm going to sort of extend later, so you input this object a then right afterwards, you input this object b and then right after you input this object c. How does the macro column learn this sequence?

C

So the first thing that I think is happening all right. You know hypothesizes happening. Is these layer 5a neurons, do pattern separation and enable this sort of recursive uh projections between layer, 2, 3 and layer 5a. The connectivity is highly consistent with that, but there's no, you know definitive evidence for it, but what happened is if you had pattern separation. This way, one input a would create a pattern, separated representation, layer, 5 a and then what would happen is this would project back in sort of a random way and depolarize some random subset of them.

C

Now, if you input, b, you're going to get sort of the pattern, uh sort of the sparse pattern that you folks were proposing was coming from layer two three via this layer, five, a sort of random connectivity, and if you play that forward same thing, sort of happens, this representation of c ends up actually being also a code for the sequence, abc so same general concept that you folks have been proposing.

C

What's different, though, or perhaps different, maybe I you know don't fully. I didn't read everything is that now you can one shot replay this so so there's a variety of ways that kappa is capping through just sort of short-term plasticity. It could happen through gamma oscillation. So if you know a originally is input for free, a sufficient amount of time that you get gamma oscillations, you can show that you very quickly will get short-term potentiation.

C

But the interesting thing is inputting this. Just once you could potentially get replay and there's a lot of evidence that within the cortex you can actually see replay within layer, 2 3.

B

I don't think that's different than our network. Our network could also do this in one shot replay. I think what you've done is sort of we've. As you point out, we've we've assumed that the connections between layer, two three cells or these layers in the upper mini columns, are learning the transitions and you're you're assigned to a different set of cells and projecting back for different purposes.

B

I think they both do the same thing in that regard. Cool.

A

I mean like these requirements.

B

You're specifying max these are also requirements. We had.

C

B

Worked on our circuitry, so you got it familiar to us cool.

C

So then, the the sort of second idea here, which is what you were getting at before, is that these layer, 5b neurons, if they're doing pattern separation, will have a lower dimensionality output that is unique to any one of these sparse patterns, even if they're, highly overlapping and the proposal exactly. As you said, jeff is, after learning, I'll explain how this happens later. Suspend your disbelief for a second.

C

After learning the pattern separated output in layer 5b that represents the c from abc meaning, the sequence abc gets mapped to the whole sequence, so it becomes a common output for the entire sequence. So the next time you input a it's actually going to output the code for abc, and if this occurs, if I can convince you that that something like this is happening, then the macro column has one shot generated: the unique code for the entire sequence, um so I'll, suspend or just believe for a second I'll walk through how I think that happens.

C

But assuming this happens, you can see very easily how, even though these sequences are highly overlapping and the brain they're represented completely differently, because the the first input is different, you get different layer, 5b output and you get different sparse patterns very similar to how you focus model things.

C

But I don't think that this, if you might disagree with me on this- but I don't think this really demonstrates uh everything you need on prediction, because you can say out loud. I know that the next note is going to sound like blah you can. You can hum it, which means that around your brain, the language area etc needs to somehow get this output, which is the prediction that c is coming next, so so what I'm looking for is at least some subcortical output.

C

That is passing the prediction of the next element. um That is a code that represents that right.

B

So I don't know if you want feedback, because you're going thanks, yeah yeah, please I'm not sure so. We've we make this. um We break this down into two things. We call them active prediction uh and active prediction is a cell that actually is firing which isn't a predictive state which you're going to argue. Is these larger six cells in a moment?

B

But uh but we think that most of prediction in your brain you're not even aware you're just you're, just not even you in some sense, you have to attend to it to know to make that to hear it in your head, yeah the vast majority of times when you're dealing with the world you're. Not you don't hear these things. I don't imagine what my finger is going to feel when I touch something when I'm in the elevator listen to the elevator music. I'm not!

B

I don't think I'm hearing the next note, but I do know if it's wrong. We've made this distinction between there's really got to be two types of predictions going on. The vast majority of in our minds occur silently inside of neurons and uh they don't generate any kind of output, but you know when they're wrong and then, if you attend to something- and I say- oh, what's the next note or think about this or then there has to be an active set of cells. We did not address that in any of our papers.

B

Now we assume that it has to exist, so I think you need both of them.

C

Got it makes sense? So yes, so in that context, then what I'm looking for is the active prediction neurons, which I'm proposing happen uh in these layer. Six, a corticothalic neurons.

B

Which I like that idea by the way, I think because they're simple cells and and so I've always you know we always have to say where is those active predictions, though.

C

Right right, cool yeah and the connectivity is really interesting because these neurons have their apical dendrites in layer four, which is exactly where the input um that these layer, four simple cells, um are detecting coincidences. So when you, when an object, is represented in a macro column, the idea would be that these cell light cells actually fire, but a pattern of layer. Six, a corticothalomic neurons are in a predictive state because you get dendritic spikes.

C

Now, if that's the case, you also see that these layer, five b neurons, also project to these layer, six, a quarter of a thalamic neurons, and it's the case that we can regularly see that these layer, six a corticothalamic grounds, uh potentially activate inhibitory, neurons and act and inhibit each other. It seems to be a winner, take all network. So if you merge these things together, what you will get for every element in a sequence is a unique sequence element code.

C

B

I have a question about this max and maybe maybe I missed something here, but it seemed like there was a problem here. uh One of the problems is that you can have lots of uh the same element repeated multiple times in the sequence, and sometimes they repeat it multiple times.

B

You know you can have the same note repeated three times in a row and there has to be a unique representation for each one of those, uh because you then on the fourth note predict something different, and so it seemed if you're, using your stable layer, five code as to sort of the disambiguation of the input, it's not sufficient to to say. Oh, this is the second b in the sequence. This is the third b in the sequence, and so on. Did I miss that, or is that something you haven't addressed yet.

C

No you're, absolutely right. That is, that is an issue and I think the um the way I've been thinking about that is, I think there might be. There has to be something special going on um when you, when you re-represent something again um so so, for example, the idea of saying, b four times.

C

um If I go b, b, b c, I think there must, I think, there's something involving the striatum happening, because the striatum seems to be very, very involved with things like beat based timing, so I think there's some sort of refreshing going on yeah.

B

Maybe, although I used to often use the example, it doesn't have to be timing either. I use the example of sentence. I say there are too many tutus in this sentence to count yeah and you have the you. Have the sound two appearing multiple times and this.

A

B

Has to be represented differently. You know. I just want to point out that, where, if you actually allow the connections between the sparse representations and layer, two three, the as we proposed, if you allow those to learn transitions- and it may be in addition to your label at the layer, 5 stable output, because that's like that's the label of the sequence yeah, but if you allow those uh neurons, the sparse representations and the mini columns to connect one another, they solve that problem.

B

You you just end up with a unique representation for the repeated element each time and I think it's a little bit it'd be a bit of a stretch to say: oh, I'm going to rely on the story item for that, because I mean I think this is ubiquitous property everywhere in the brain. Yes, we do it in motor sequences. We do it everything it's in language, it's in um it's almost everywhere.

B

So I just want to point out that I think one of the things you lost by by you're going to you argue the layer of two three connections are really sort of forming stable patterns and not learning the sequence themselves. And when did you lose that ability to repeat uh elements and sequence inside yeah? That has to be addressed? No.

C

You're right, that's definitely a limitation. I appreciate the feedback um cool, so the idea at least, is for non-overlapping elements within a sequence. um You can get unique codes for this, and the idea, interestingly, is if you follow the connectivity of these layer. 6A ct neurons. You see something that seems at least highly consistent with the notion of prediction. So there's evidence that came out six years ago that they put these layer.

C

Four stellite cells into predictive states, they synapse onto mod and to distal dendrites and layer four, and they synapse on distal dendrites on the relay neurons that go to the same exact macro column and they've also shown that these relay neurons in the thalamus only versed when they get such sub-threshold input. So there's some interesting things that at least seem consistent with the idea that these layer, six, a neurons, are somehow predicting the next element.

A

Which sorry, which paper is that that show the bursting in relay cells? Is that the claim.

C

At all, it's later all, okay, okay- I didn't I didn't have that here, but I I in the next section I I cite it: okay, um cool, so here's just the general idea so subcortically these are passing predictions of the next element um cool. So here's an idea of how these three things happen, where I think things start getting more interesting is when you start asking about working memory, so a key flawed assumption.

C

At least my model made thus far is that there was no delay between the input of sequences and the reason that was essential is because, in order for the sparse pattern of b, for example, to show up it needed to be input within the time window that it was sub threshold depolarized from these layer, five, a neurons projecting onto it. So if I input a pause five seconds and then input b, pause five seconds input c.

C

Clearly, a human can learn that sequence, but the network I proposed here would not so the question is: how do we sort of cascade these things across time and that's sort of where I'm gonna propose an idea of how working memory might happen.

B

Can I just again I I'm I'm trying to be concerned for the people on our side of the fence here, mexico.

A

B

So from our work, I just want to make it clear that max is trying to deal with a different problem. We've dealt with timing in uh the timing of a sequence as in like the the tempo and the of a melody and the relative timing of notes in a melody, but we haven't even thought about this match. You didn't deal with that, you're doing a different problem, which is a working memory problem, um and so I just want to just parse it out.

B

These are two separate timing issues yep one then you're dealing with sort of a working memory, type of thing, not the melody.

A

So in this example, then, the delay is almost something you want to ignore right. You don't really care whether it's a one second delay or a five second delay or four. Second, it could be. It's.

C

Just the fact that you're, not in my exams,.

A

Melody where the specific timing really matters, yeah right.

B

And in the melody it's been shown that the specific timing can you could do it up to about 800 milliseconds? So beyond that you can't really keep it. You can't learn the duration of a note. You have to start counting interesting, so there's a there's an inch there's a built-in mechanism which we hypothesize about what it is that basically, can do this very specific, relative timing up to 800 milliseconds, and not only do you learn the relative time.

B

You can speed it up and slow it down, which is like you know you can play a melody back quicker or slower. um So that's another requirement, but you're doing something completely different here. This is working though.

C

Cool okay, so the problem statement which we just stipulated is how does human? How do humans learn sequences when inputs are delayed, so there's three models: I'm not going to review all of them, but just for folks that are interested in alternative ways. People have tried to model working memory, there's three main models out there.

C

One is attractor network models, the idea that representations maintain themselves for the delay period, there's bi-stability models, which is the idea that the neurons don't fire, but they remain in some sort of on state or upstate during the delay period, and then the third is that somehow working memory is maintained in short-term weight changes, so the model I propose is very much in the third category.

C

The reason why the third category has benefits is the first to struggle to understand. How would it be the case that you could store multiple elements simultaneously, because attractor states are very sensitive uh to parameter changes, it's hard to imagine how uh an attractor state would maintain multiple representations simultaneously.

C

um Also, how could you maintain something working memory without disrupting uh ongoing sensory processing, so attractor network models struggle to exam sort of explain how that would happen. So as a proposal of how this happens just in synaptic ways?

C

Okay, so here's our learning paradigm input a pause b, pause c. So the proposal here is that the origin for this comes from the hippocampus.

C

So we know that the hippocampus on the theta uh cycle will output a consistent place code, I'm going to call an episode code since I'm not referring to a location in space, but I'm referring to the same exact computational thing I mean we know that it projects extensively to the sort of medial, prefrontal cortex, and what we also know, interestingly, is that the acc and the medial, prefrontal cortex seems to project over many many areas of sensory cortex, specifically to these layer. 6A neurons.

C

It's also the case, interestingly, that ca1 projects specifically to layer 6a in higher order, sensory cortex. So it seems to be this sorry, you were about to ask something: jeff, sorry,.

B

No jump in I don't mind. I was curious. I I'm just wondering I don't know that projection. Do you know that if it's that apical dendrites or is it, is it more to basal um dendrites um that projection.

C

They I know that they use genetic markers, so it wasn't just the layer in which it uh the axons were physically located. I know they showed that, specifically they synapsed onto these ntsr neurons, which are the layer 16 neurons. I don't remember if they.

B

Could be layer one, it could be direct yeah.

C

I think that's from the case.

B

Yeah I mean because if it's layer one then there's a general rule about any backward projection, quartal region, the region, projects heavily to layer one. So then that would be I've sort of a general purpose of methodology.

C

The ca tracing, though I know, specifically, showed that ca1 projects to mostly the layer six in the higher order, cortex, which was an interesting finding.

B

uh But that goes through the prefrontal cortex and not directly right. There.

C

Are there are some direct rejections.

B

C

Interesting yeah yeah: um it's only the higher order. Sentry cortex, though so you don't see it in b1, but you see it in like um so another interesting finding. Is you see uh theta oscillations only in working memory tasks? Do you see entrainment between these sort of three structures, so a non-working memory, tax sensory cortex seems to mostly oscillate on alpha, but when you start trying to remember things over delay periods, you see sensory cortex start oscillating more on theta.

C

It also seems to causally propagate from the hippocampus to the acc then to sensory cortex. So if you sort of do timing studies, you can see that the the oscillations seem to propagate in that direction.

C

So the idea is that these these episode codes enable you to reactivate representations in the absence of sensory input. So if it's the case that you project to these layer 6 a neurons, it has been shown that they provide driving inputs to these layer, 5a neurons, which have been shown to provide driving input back to layer, 2 3.. So there's some evidence. That's supportive of this again, obviously not definitive evidence or proof, but working memory seems to activate all layers except layer, four.

C

So during delay periods, you seem to see silence in layer four but activation other layers.

C

It seems to be the case, so some models of working memory propose that prefrontal cortex sort of mirrors a representation in sensory cortex, but it's not actually being re-represented in sensory cortex. A lot of recording studies are inconsistent with that. For example, if I record a neuron, that's responsive to someone's face- and I ask you to remember that person's face- we see to see you- we seem to see reactivation of the same neuron that was activated during the actual sensory experience, so it seems to reactivate the same neurons.

C

Also, we see evidence that firing starts in the deep layers and propagates upwards and it seems to be the case that replay is triggered by the hippocam. So there's some interesting consistencies with this idea.

C

um So let me walk through how this uh I propose this might work. So if it's the case that ca1 keeps replaying the same episode code, every 200 milliseconds, now I'm going to propose that just the replay of the same episode code doesn't have to be anything smart happening. Hippocampus. It just has to be the same code over and over and over again will immediately make it possible for you to connect a b and c across this.

C

This sort of pause, the one additional mechanism that you kind of need is uh timing within the macro column. So timing, you know not the timing you're describing but coordination between macro columns, so they have to sort of aggregate input at the same time, so I propose that there are sort of input, states and output states there's a lot of evidence that suggests uh that when layer 4 is activated, it actually inhibits layer, 5 and when the deep layers are activated, it kind of inhibits the superficial layers.

C

So there I propose that when macrocom's receiving input, it's not sending its output and then when it's passing its output, it's not aggregating actual sensory experience, and this is an important uh mechanism for for what I'm about to describe.

C

So now the idea is the entrainment. The consistency between the structures is the input output states oscillating such that the output states happen at the same time as the hippocampal replay.

C

So let's sort of follow the dynamic of what might happen. This is where you're going to see the stable sequence codes in layer 5b be represented. So let's say a random code happens. These layer, 6a neurons, randomly activate a subset of them, which randomly activates a subset of these layer, 5a neurons, which project to these layer 2 3 network. The idea is, when a is first experienced, you create a mapping between this random episode code and the actual most recently experienced sensory experience.

C

So what ends up happening is in the next output state. You see a pattern separated output from layer, 5b now a key feature. These layer, 5b neurons, seem to burst for a very long period of time, so they seem to burst between 50 and 200 milliseconds and they seem to be insensitive to any other output. So what that implies is any other representations in layer. 2 3 that occur during this output state will not change the see the layer, 5b output and this is going to be key to sort of learning a stable output.

C

So then there's a replay. While this is outputting the. If the uh hippocampus fires that same representation, then you'll just replay a again and now when you replay a then you get this sort of biasing, okay, so nothing too crazy, but every output state during the input of a you're just going to get a replay of a during the delay period. So now I'm going to the delay period, there's no sensory experience, but during the output state the idea is that you keep getting this replay as long as you're attending to it, meaning you're.

C

Maintaining the theta oscillation entrainment and the same episode code is playing you're just going to keep getting replay of a over and over again and a consistent sequence. Biasing from these layer, 5 b, 5 a neurons. So now the intuition, maybe is somewhat obvious. Now, when b is input, you're still going to get the sparse representation of b, because every output state consistently the hippocampus is driving a replay of the same sequence biasing.

C

So then, then, the next time there's a replay. You get a replay not just of a, but you get a b a and then b and then you, then you will get a sparse pattern from uh the the layer, 5 ars, neurons, specific to the sequence, a and b. So you can kind of play this forward. uh During this delay period, you're just going to keep replaying a then b a then b a then b and then, when you input c, then you get the same, sparse pattern for just c.

C

So the other key thing, though, is during these replay episodes notice. The r the uh the layer 5b um neurons, that are being bursted throughout the entire output period are specific to the pattern separated l23 neurons for the sequence abc, and the idea is that during this replay, because they burst for for 50 to 200 milliseconds you're, building plasticity between a b and c to the same sequence code.

C

So the idea is, if you keep triggering a replay of a which then goes to b, which then goes to c and you're enforcing the layer, 5b neurons, to burst in a pattern. That's representative of specifically the sequence abc then you're building plasticity. Now between a and b and c to the same output code, did I go too fast or did that make.

A

Any intuitive so you're, uh so this so this plasticity is uh this: can these connections are being reinforced during this delay period? Exactly right? Does that mean that the longer the delay, the faster you're going to learn, because because you're going to reinforce that multiple times during the delay, or I think.

C

A

C

It might matter no that's an interesting proposal. um I didn't. I didn't really consider the requirement for the delay um for the the length of the delay period. I'm not sure it's an interesting thing. Okay,.

A

But during the delay, basically, these connections, these random. Well, these connections are being reinforced through this connection. With this random episode code, yes,.

C

A

C

Replay every replay reinforces the mapping of the entire replay to the final output code.

A

Yeah and then suppose, you're now doing inference on the sequence. You know some time later. You know many days later or whatever uh you don't need to have this same episode code right now, you've built up these connections. Now you can uh do that. Okay,.

C

Exactly so, that's that's sort of a concept of how perhaps the hippocampus kind of trains, the neocortex and over time the neocortex doesn't require the hippocampus. It might provide ideas on how that might happen.

B

So I I have a question about this, but it's it's not about the mechanism, it's about something, that's not in the mechanism. So if you, if you want to finish the mechanism, then okay.

C

um I I will uh quickly finish: I have two more slides on the mechanism and we can go there. So the idea here conceptually is the next time a is input. In a macro column, the output will be unique sequence code for abc instead of the original pattern. Separated output for just a um and what's interesting, is it sort of provides, maybe intuition on how episodic memory works, which is the an episode code, can reactivate a bunch of macro columns simultaneously through layer, 6a and there's three pieces of evidence.

C

That's supportive of this people with hippocampal damage and rats with hippocampal damage are very impaired, with sequence learning trace conditioning across not just mammals, but also in amphibians and fish. If you remove their the analog of the uh the homologue of the hippocampus, they are unable to perform trace, conditioning meaning they cannot connect things with delay periods.

C

um It also seems to be because the hippocampus is necessary, so maybe provides intuition or concepts on how this might happen. Okay, jeff, take it away.

B

So um we're talking about the layer, five intrinsically bursting cells um in playing this role and you're you're, proposing a a mechanism or why they do this intrinsically bursting uh physiological thing, um but just a little aside by the way in some animals, that's in it's labeled layer, 5a and some it's layer, 5b, so really saying a and b doesn't really help. You have to say regular spiking got it um so, but you did on your paper. You always said rs yeah yeah anyway.

B

So now the question is: those are very interesting cells, those layer, five intrinsic reversing cells. They are. They are known to be the motor output of the cortex everywhere in every cortical root column. Everywhere they are the vet cells and motors projects, but so they have a motor wall they also- and so they just project subcortically uh to some motor center and they also split their axon and project to uh higher order relay cells in the thalamus, and so this dual role is very perplexing.

B

um How is the motor output is going supportively, or is that what we're projecting to the next region? Also a motor output or not? But I, as far as I can tell you're, not you're, not addressing those roles at all here: you're you're, giving them a different role um which is sort of this. uh Helping helping us do this integration of working memory and first, a question for you is: have you thought about how those intrinsically bursting cells could also be representing behavior?

B

How is it that they are behavior cells as well as doing this, um and have you thought about the other questions like well? How could they be projecting forward as well in cortex.

A

C

So so uh thoughts on that are, if it's the case, that progressively higher levels of cortex are trying to do some form of progressive unsupervised clustering. Then the idea is these layer. These intrinsically bursting neurons are clustering. A sequence of input, so third layer of cortex represents sequences of sequences of sequences, an object in the third layer of cortex, the subcortical motor output.

C

Perhaps here's a this is how I'm thinking about it is the reason that's useful is the subcortical motor areas are trying to learn the the the object that exists so, for example, the superior colliculus, the projection from layer 5b to the superior colliculus?

C

You can actually look around a room without a v1 you can you can throw something at someone's face and they'll blink. They can still see, but they don't proceed. They don't perceive sight the idea being that v1 v2 v4, what it's helping the superior colliculus do is recognize objects.

C

So the superior click list can then learn mappings between these stable codes and some meaningful. Okay.

B

That's interesting: okay, that's interesting! um I have a completely different take on. It doesn't mean it's right, um but it just just might be useful for you to think about it. This way, um my assumption is that the superior colliculus does its thing and what the cortex is trying to do is it's trying to say look. I I have a better model of the world than you do.

B

I know, what's going on up here, you're just pretty dumb and old, um I'm gonna, but I can't control muscles myself so I have to. I have to talk to you to do something. So the is the thing that controls eye movements, and so the cortex doesn't really want to control eye movements. It wants to control. It wants to suggest to the superior caliculus. Here's how I'd like you to move at this point in time, and so to my view, those connections are learned that the court checks um sends its trust.

B

It's basically saying I have a code that uh that that I want you to associate with behavior you're exhibiting right now and and therefore in the future. If I give you that code, you might do that. Behavior, for me, it's not teaching the superior calculus what's going on in the world, it's just the cortex. Taking advantage of it saying please do my bidding at this point in time. It's a very different view than what you've had yeah, but it doesn't mean it's right, I'm just pointing out it's different yeah.

C

It's interesting one, one additional thought there, which uh maybe is consistent with both views, is these layer, five b neurons, also project extensively to the striatum um across all of sensory cortex and and that striatum also then projects uh two superior colliculus and other motor areas. So you might you might imagine something where, when these layer, 5b neurons post a, I think, the object, we're observing is object, a or sequence a then for reinforcement. Learning in the striatum I can learn. Is this a good object, or is this a bad object?

C

What behavior do I want to do in reaction to this object and the stratum can learn for reinforcement, learning, the mapping between objects and behavior? Perhaps that's something: I'm I've been thinking a lot about.

B

Another another thing it's worth considering here is um you went over very briefly here of the idea that you're proposing that there's two states, alternating uh all the time in the cortex between inputs and outputs inputs and outputs- and you didn't say it here, but in your paper you said: that's assuming it's an alpha frequency, yep and- um and we kind of came to not that exactly that conclusion.

B

But the layer, five cells themselves, one of the possibilities how you can explain the layer, five intrinsically bursting cells- is that you know they're projecting forward in cortex and they're, also projecting down motor behavior, and um I find I find that difficult to understand.

B

And so one of the one of the things we came up with is an idea. We haven't really pursued this much, but it was in the 2019 paper um that that the layer five cells are alternating um um on a cycle between the meaning having a meaningful output, subcortically, which is a motor behavior and a meaningful output, um cortical cortically, which is not a motor behavior. um So this, but we only did it at the layer. Five sub.

B

We didn't propose that the entire cortical column is oscillating between inputs and outputs, which is an interesting idea of yours, um but that just the layer, five cells themselves could be oscillating between two different states. Two different meanings: we associated with these things. We call displacement cells, but I'm just I'm just trying to make these contrast. Yeah clarify things here and I'm also just making sure I understood your model correctly because um I went through it. I said: oh, these are things you did were kind of cool and clever, and these are things I think.

B

Well, you didn't address this issue and yeah.

C

B

Want to give you a chance to make sure I just didn't misunderstand that, but anyway, you're giving these intrinsically bursting cells in some sense, a different function um and not really addressing directly their motor behavior, correct, yeah, yeah, which I think we have to address, because it is the motor output, of course, okay.

B

So something else has has to be added.

C

Yep cool, so I um are we do I have. I have 10 minutes left. Is that right, I'll try and go fast.

A

Yeah we have a little.

C

Bit of time, okay, so another interesting thing, I think you folks have solved this problem differently, so so I'll describe sort of what I've taken from predictive coding and adaptive resonance theory. That might be maybe use useful to augment. I'm how you folks have thought about this, but the problem that I was trying to solve was uh sort of the catastrophic forgetting or the stability plasticity dilemma, which is, if you have continuous streaming input really from birth to death, to the neocortex for 80 years.

C

How do you ensure, as it learns new things, that it doesn't accidentally forget, old things um and uh htm or sorry, art and predictive coding suggest that this happens through gating learning using surprise? So the idea is you only learn when your predictions have failed. Otherwise we don't update representations.

C

Now, predictive coding has pretty biologically, in my opinion, implausible mechanisms for how this happens. So I didn't find it particularly convincing, but stephen grossberg came up with the very interesting idea on how match versus mismatch might be signaled in the thalamus that I think might be worth considering.

C

So the idea is these layer, six, a neurons project back to the same relay neurons, that project into a macro column and what we want to happen is somehow success, verse failure of these predictions, meaning I correctly predicted my upcoming input or I failed to be signaled differently.

C

So, if you look inside the thalamus, here's where subutai I referenced the uh relay iron burst only when getting sub-threshold modulatory input. Those are the things.

A

C

Thanks um and what you can show is these layer, 6a neurons projects, distal synapses, and do not drive output on these relay arms, but they do put them into an upstate, um in other words, a predictive state, and if you take one of these upstate relay neurons and give it driving input, it will burst.

C

So the idea now is when predictions are matched you will get bursting of these relay neurons when they are failed. You only get tonic firing and perhaps this somehow is used in macro columns to decide whether I'm going to potentiate synapses on these layer, 4 inputs or I'm going to do long term depression.

C

So that's one sort of interesting idea for match versus mismatch. Stephen grossberg also proposes some neuromodulatory mechanisms for what's the threshold for a successful match versus a mismatch, which is some interesting ideas in there also what's. He also proposes, which I find interesting, is this idea that there is an explicit surprise signaled by these sort of intralamin or thalmic.

C

So there's not proof of this, but the connectivity is consistent with this idea. That trn, which is this inhibitory layer around all of thalamus, has electrical coupling and it seems to be the case, these layer, six, a neurons, only synapse back on inhibitory, neurons on core relay neurons, not the uh the inhibitory neurons to these major. These major neurons there's actually a recent paper that further showed that so what this implies is due to electrical coupling. If you get bursting of core relay neurons, you will get weak, matrix, neuron firing.

C

If you get tonic, firing and relay neurons, you will get bursting and matrix firing and what's cool is, if you record these intra-lame orthotic neurons, you see exactly this dynamic, which is they only seem to fire in the presence of surprise, and if you follow the connectivity there's further interesting things.

C

These are the only phthalic neurons that project to ach or tonically active neurons in the striatum and they've shown that the the burst pause dynamic of these neurons in the stratum is required for plasticity in in the stratum, and if you remove the input from the insulin or thalamus you, you prevent uh plasticity in this rhino. So you it requires a pause and acetylcholine or I don't know how to pronounce that word. um It's that pause is required for uh plasticity, um because it inhibits plasticity specifically on direct medium spinning around.

A

C

Some really interesting things going on here. I think um you can also show these matrix neurons project uh to a key ner nucleus that innervates the entire cortac cortex with ach so anyway, some interesting things to think about. In my paper I propose some specific mechanisms. It's just one of many possible ways. This is used, but the idea is that perhaps learning new representations is gated by failed predictions.

C

A

Well, you know, I think you know that general idea. I think we would uh agree with the way we modeled uh failed. Predictions in our in our 2016 paper is that we get sort of bursting along these mini columns, uh and so you get you know an entire set of mini columns would become active and that triggers uh you know, learning of new transitions and new sequences, so it doesn't have to kind of leave the layer because, if you think about you know every layer, you know trying to learn sequences.

A

um You know, at least in that model. You know we can learn everything just completely within that layer, so we layer we can learn from mismatches and from surprise uh just within that, uh but uh there's an interesting idea here as well.

B

There's a lot of thoughts about this in my head at the moment um I realized I wrote some of this in my new book and I don't think we've written it anywhere else, um so uh no one you wouldn't be aware of it, but but we've discussed this in momentum, there's a we would. As somebody say, we would agree with the general idea that um that when you have a mismatch, that's a signal to learn right.

B

That's a signal that something's wrong with your model of the world, and um but what I think tied to that is also this. The very key is this idea of attention. So it's this is more obvious if you're doing sensory motive, learning where you're moving your fingers or moving your eyes um and and you're going along, also there's something in the in your sensory field. That's not right! It's not predictive what we tend to do in that case. What we do do not just tend to do.

B

What we do do is our attention is immediately drawn to that thing and we actually narrow down on input to some subset of the field we were looking at before we try and what would which essentially, we believe if you're familiar with our 2019 paper, we build a model of the world using reference frames. um Bits are like reference frames and that that reference frames are populated with other objects, reference frames other so there's reference names that are populated in locations with reference names and what you're doing is.

A

B

Saying, okay, I have a model of something it has these components of these relationships to each other. One of those is wrong, so I now focus on that say. The thing I thought was that location x is not right, and so I need to relearn that thing at that location x. So it's a much more complex involved mechanism. It's not inconsistent with what you're saying is this more it's more complex.

B

We have to attend to a subset of our input space and we think that thomas is internally involved in that uh it's not just a mismatch. It's like okay! I have a mismatch. We need to attend. We need to find the location where that thing is. You know to relearn that component of that object, yeah, um and so it's it's just a it's a much more complicated um uh process.

B

um That's going on.

A

Okay, yeah, so this this kind of circuitry um you know, could be tied to kind of this attentional mechanism of you know, focusing your attention to these surprise areas and and that that can trigger learning as well, so they're, not necessarily inconsistent with one another. I think no.

B

That's, I guess that's where I was going yeah and, of course the thomas has been implicated in all kinds of attention uh models from for 30 40 years. No one really understands it. I I think, did we did you ever publish that paper? We did with the woman from mit. What's her name.

A

uh Carmen carmen.

B

A

Still kind of get together every once in a while, you talk about it, we haven't, uh we did we published a co-sign.

B

A

Max poster.

B

We've been, we spent a fair amount of time studying the thalamic relay cells and they're complicated. They have these very unusual synapses. They have multiple ones, only some seem to be used. It almost looks like the thalamus is doing some sort of um um uh matrix.

A

Like routing or.

B

Routing dating mechanism- yeah yeah, that's there's another one anyway, a routing mechanism that you can. You can redirect input and change things, and- and so it's doing something really complicated yeah, but it does it with a fairly simple uh well-described, uh physiology and anatomy. So, but it's still puzzling to us yeah, um I think you're touching on the surface of something which is more complicated anyway, we we did a paper super.

B

They did a paper with carmen uh where we're looking at all these details and trying to speculate what kind of um functions they might play. um So I'm just pointing out that I think it's more complicated than this, but not that I disagree with anything. You said here just going on.

A

Jeff was the word you were reaching for a crossbar.

B

No, it wasn't a crossbar. What's the multiplexer, that's what I was thinking of.

A

Well, that's that's the ultimate form of a multiplexer. Okay, all.

B

Right, okay, but that wasn't the word I was looking for, it might be a good word. I was looking for the word.

C

um Do you have a hard stop? I can I we can skip five. I can just go through sensory motor prediction, or do you want.

A

um We can go for a little bit longer.

C

Yeah, that's fine, so I'll quickly do this. I think this might be, uh might have some interesting overlaps with uh your folks's grid cells. So I'll be interested to hear your thoughts on this. So obviously, as you said, somehow, the brain predicts the difference between me moving and the world around me moving if your eye sort of traces, this room you're getting a stream of different sensory input, but you are aware that it's because your eye is moving, not the world versus.

C

If I showed you this stream of input, your brain can tell that the image is moving, not your eye. So how does this happen?

C

So I attribute this to sort of the corticocortical neurons in layer six, and I think the proposal is is if, let's say in the following simple setup: in your left receptive field, there is a 45 degree bar and then in your center receptive field, there's nothing. Now. The layer 5b neurons seem to project sparsely to layer, six cortical cortical neurons, and the idea is also the motor input from the anterior. Cingulate also seems to project to layer.

C

Six, there's no proof that it's it projects the corticocortical neurons, but it seems to have axons that go throughout there. So if you put these in a predictive state, you might imagine that you end up getting a unique code here. That represents uh a 45 degree bar on this left receptive field and the motor commands look to the left and that output propagates throughout lateral macro columns.

C

What that means, then, is if the next thing you see is a 45 degree bar in your center perceptive field, you're effectively creating a mapping laterally between macro columns, you're effectively saying if I'm looking, if I see a 45 degree bar my left receptive field and I desire to move my eye to the left to look at it, then you end up seeing 45 degree bar in the center, so the next time you do this so the next time you get a motor command to look to the left.

C

You will get these layer 6a neurons to fire before there's any sensory input.

C

So the the idea is that layer, 6a is predicting its input before it receives it through these lateral connections and one one piece of evidence for this, which again is not definitive at all, but you do see deep layer neurons than sensory cortex with movement related response properties before movement even begins. So you see this movement in tension output, so.

B

Is that the end of this one yeah yeah go so I when I read this, I immediately said: oh no, I think you're wrong about this. So I'll tell you what okay and then and then the next two paragraphs later you said: here's why this might not work. Well, it's a combinatorial explosion, problem and there's been two uh and we've. There's been two hypotheses of how this this um this this remapping or prediction.

B

Before the you know, I basically the column prediction is going to see while the eye is moving right and and so it hasn't gotten the input. Yet one is the kind of mechanisms you propose, which I don't think are possible, and you point at this. I don't know if you thought of it. One of the viewers said it, but it's a common foil explosion problem and then you, then you came up with a bunch of ways. Well, maybe we can get around it. I don't believe it, okay, the other.

B

The other way is, you know, because oh it's done hierarchy and you know the information goes up to some other higher region and it comes down. Tells me what I'm supposed to be looking at there's a timing problem with that yeah there's not enough time for it. I don't believe it works either. We have proposed a totally different solution which I'm almost certain is correct. So I'm just going to boldly say that, um which is the whole idea is if every column is doing learning a complete sensory motor model.

B

You don't need to rely on projections from anywhere else. You just you've got a reference frame. You just need to know where you're moving to the next point, your reference frame and then you can predict what you're going to see. So it solves that problem elegantly and I think that's the solution to this problem. So I didn't, I didn't, buy this one I'll accept it. But at this point I didn't buy this one and, and you did say it's got a combinatorial explosion, problem yeah.

B

It does um that's the problem with this kind of mechanism, but.

A

I think it's more.

B

Elegant to say that you know look every column, it doesn't need to rely on someone else. It just says you tell me how I'm moving and I will know where I'll be, and I have a model for what the thing I'm observing and that model tells me what I should do you.

C

Think that mechanism solves because one problem I was trying to think through that. Maybe your mechanism solves more elegantly than this is the clear fact that the brain can do one-shot generalization to translations changes in size, and we still can't do that in basic cnns. So those.

B

That's bringing up a whole kind of words that you haven't dressed yet, but you know, even in the sequence memory, something something our sequence memory didn't do and yours doesn't do either, but I know we have to do it is um is, if you want to think about melodies, you think about pitch and variance right. You, um you can play me a sequence, a melody with a quick one, I'll, remember it in one shot.

B

You can play it back to me a minute later in a different key and I won't even notice yeah and I will predict the correct note in the new key yep. um So there is a there's. Clearly, uh an environment that's introduced there um and I actually believe that the sequences are learned in the invariant form, not in the original form. There's some evidence both ways, but I believe it's learned in the invariant form and so there's a whole other step. We have to introduce to this.

B

That's one type of invariance and that's uh can be addressed. With these this concept, we had called displacement cells, um which again it was a 2019 paper, um um but there's a whole bunch of environments, and so um and- and we don't really understand this completely but another one is scale in spatial scale like like vision or something like that and um and there'd be different mechanisms that might be involved in these different invariances.

B

uh So in the spatial scale, um one of things we're relying on the theta rhythm is very much intimately uh tied to uh the progression of grid cells.

B

Like you know, path, integration, movement um and um if you speed up and slow down the theta rhythm, basically you'll scale everything up and scale everything down, um and so you could change the map you have to move like, for example, if I'm looking at your face and you're close to me or you're, further away from me to scan to circad from eye to eye, it takes different different amounts of movement.

B

I don't I'm not aware of that. I just think I'm seeing your face, but somehow I've taken a model of your face and I've scaled. I can scale my movements to the appropriate moment in time, um which is like. Oh, I have to move my eyes left more left and I I'll make the same predictions.

B

So how does that you know if I have a single you know? How do I do that? How do I change from a egocentric motor behavior to an allocentric um movement behavior things like that yeah.

A

The big topics here.

B

But I think there's a whole series of topics related to invariants um and uh and and we didn't talk about those much the only place- we've ever addressed- that at all any of our papers was this very speculative small section on displacement cells. But we didn't talk about in the sequence memory um yeah and we didn't talk about. Maybe we mentioned it.

C

Supportive facts, for that is um birds that do birdsong, do not actually learn in a frequency in variant way.

B

C

Thing offset is slightly, they don't recognize it all. Yeah.

B

C

Nice seems to be neo4j yeah.

B

Yeah yeah yeah exactly they have this very specific mechanism for a learning song and that's that. But it's really hard for us to tell when the song is in different keys right.

A

B

There's a nice uh there's, this guy research. I can't remember his name. I can never remember anybody wrote the book. You know this is your brain on music and I talked to him about this and he um he says, there's evidence actually that we do remember the key. We first heard a song in, but we obviously we clearly remember it in an invariant form as well.

B

So if someone, if I would ask you just to play back one of the popular beatles songs, for example, it's very likely you'd sing it in the correct key yeah, even though you would say I don't know what cheat was in. But if I ask you to say start on this note, you will be able to do it. No problem, so you're not even aware that you know the right key, but it's there.

C

B

But anyway, so um I think I think when you, I think the issue of sensory motor prediction is far more complicated than this. It's really you know you have to like how do I know what I'm going to predict if you read you know, you know the cop example in my 2019 paper. How do I know I'm going to predict it's not just oh it was. It was being sent elsewhere and now I'm moving it. It's like no, I'm about to touch something or see something.

B

That's not even in any part of my receptive field right now and I'm going to make the correct prediction, uh I'm moving my fingers to the part of the cup and I'm not touching it, and I know it's going to feel that. So I think this is a much bigger problem and personally, I just hope this is the weakest part of your paper, but your paper's got a lot of great things in it. Okay, um I thought just like. I don't believe this one.

C

Yeah, I convinced you on a few other things and all.

B

Yeah, I know I think it's great don't get me wrong. I think there's a lot of really interesting stuff in the paper.

C

I'll skip the other part, but yes, there's lots left to do. um I haven't done any simulations, so I you know that stuff. There's a lot of neuron types I haven't incorporated.

C

um I want to incorporate the striatum I'm going to work on frontal cortex, so this is just a passion project I'm having fun with, but uh I appreciate you folks uh letting me speak here, and this has been fun.

A

Well, thank you for uh thank you for coming on. That's great, so are you gonna actually try to do some simulations? Do you think.

C

uh Well, I won't need to partner with someone that can do that better than me, so I've been reaching out to folks. If you folks are interested, I you know would love to do that, but I've been sort of networking with people that would be interested in participating. Where do you want.

B

To go with this research, I mean: are you a rising moment? It's very biological. Do you want? Are you interested in translating to machine learning, which is what we're doing right now? um What where's? What's your goal? You want to keep working on the neuroscience.

C

Yeah, I think both are really interesting to me, I'm a little as my sort of passion project, I'm kind of just seeing where it goes and following the momentum I think um modeling it in terms of machine learning is really interesting to me, um but also just the neuroscience is fascinating to me. um So so I'm kind of like I'm following both paths and wherever people seem to be most interested I'll kind of lean in.

B

How much time are you able to dedicate.

C

This is like why this is my weekend project when it's. It helps me the stresses of running a company, it's nice, to relax into a textbook.

C

This is the side project to stay sane.

A

And maybe sheltering in places helping you.

C

Yes, absolutely.

B

Well, I think that's an interesting in terms of how we might engage with you on some of this stuff really depends on. You know what you want to accomplish and what you want to do.

C

um I think the machine intel like finding a way to implement this and merge it with html or apply some of this in terms of html would be very fascinating to me. I would love to help in any way. I could.

B

You guys are, you were at all the work we're doing now in machine learning.

C

uh Probably not all of it but I've. I try to read all the papers. You folks release.

A

I don't know what we're doing yeah we've, uh so we've uh sort of taken this uh spatial cooler context and the uh you know those ways of introducing sparse representations and we've applied that to machine learning um and shown that you know, as we would predict from the neuroscience that, if you have really sparse representations and even in machine learning, you get much more robust representations, because really high dimensional sparse patterns are very unlikely to interfere with one another.

A

That same basic idea can be then applied to catastrophic, forgetting and continuous learning, and so we're currently in the middle of figure, how to add, active dendrites to deep learning systems um and to build in kind of the prediction. Mechanisms that that you discuss with the idea that you know active dendrites can recognize some sort of context.

A

um You know represented with sparse patterns. Then the the output of the network is going to become very context dependent, and uh you know, by learning different contexts in a way that don't where the individual contacts don't interfere with one another. You can do actually uh catastrophic. I mean you can do continuous learning without catastrophic, forgetting um just by exploiting this property of sparsity that you know, new patterns interfere with old ones. I think that's that's almost a requirement, I think, for doing continuous learning properly. Do you.

C

Folks ever plan to incorporate other structures other than the neocortex.

A

um So the thalamus is definitely very important to us. As we we've talked about um you know. We draw a lot of inspiration from the hippocampus and hippocampal formation, but you know it doesn't project everywhere in the neocortex, so I think as a first approximation, we don't want to rely on it, but I think we get a lot of inspiration from it. I think our tendency, if I had to at a high level kind of differentiate between some of the approaches, you've taken yeah.

C

A

Tendency is to try to just look at the cortical column. What can the cortical column do without resorting to external things, because uh you know we think the cortical column is much more powerful than you know. People give it credit.

B

And we would include the thousands in that. We think.

A

We will tell this to us: is.

B

Another layer or two of the cortex, it's just a couple of those.

A

But our tendency is to really just look within the the generalized definition of cortical column uh and all of the functionality rather than saying okay, it has to come from. You know this.

B

Other stuff, like you, brought up, you know when you brought up the whole idea of uh of um uh working memory like okay, that's something we wouldn't include, we would say no we're not doing you know we're dealing with like how do I just learn a basic predictive model of the world.

A

B

Model in the world, and I'm not saying okay, you know robot, you know, remember these seven things now, while we go on something else, right, yeah, um it's like how do I just understand how you know you mentioned watching dishes or something like that, but you know how do we just do these most basic modeling systems.

A

B

We're so impoverished in our current video.

A

B

um Yeah, you know one of the things you didn't mention, which I was in your paper, which I really liked, and it's all right. I just I throw an advertisement for right now. You did a nice analysis of the different um uh oscillation frequencies and made a nice proposal.

B

I think I don't know if I've ever seen that cohesively before, uh where you know here's a proposal for gamma. He has a proposal alpha. He was supposed to you, know vader data or whatever it was, and then how you know, a different state of of awareness and different states of attention that these different things played in. I thought that was very nice. uh That was in the hierarchy section if you skipped over um just plug it in, for in case I.

A

B

I'm going to think about that some more too um yeah.

A

It's something.

B

We've only recently sort of really delved into more.

A

I also like this idea of having this kind of random code. That's repeated as a way of kind of holding on to memories uh for a period of time.

A

You know I don't know that we have to resort necessarily to hippocampal code, but just the idea that there's a consistent random code that can be used to tie together uh things and extend that. I think that's a really nice idea. Yeah.

C

Yeah- and I think one thing that excites me about that or my intuition- or something that might be interesting, I'll- make it less strong a statement is we know that it seems to be the case that the hippocampus is somehow relevant for short-term memory and over time the neocortex takes over from the hippocampus. So we we have this fuzzy idea that something like that happens.

C

Perhaps it's the case that this this replay is what the hippocampus does to enable something to be stable for a short period of time, but as the neocortex learns through mechanisms, you folks have described that becomes less relevant and perhaps there's some mechanic going on.

B

Yeah, I think that's absolutely right. um You know it used to be thought. First, it's a very well documented thing that you know, obviously without a hippocampus.

B

And um it used to be thought that the hippocampus would transfer its memories to the neocortex and I never believed that didn't totally yeah.

A

B

Not many people would agree with that, but there.

A

B

Been studies, I think that one of us don't ask me because I remember never remember anything, but there was a very, very clear study that showed that that's not the case, the depth of that it's not a transfer, it's that the cortex learns and the hippocampus learns, but the the cortex will not finish its learning if the hippocampus is not there, which is consistent with what I think you just said, um which makes a lot more sense to me.

B

So then, again, even if we wanted to let's say model something in a in a machine, we don't necessarily have to model the entire hippocampal function. We just have to say: oh here's, a guiding signal to tell you to keep learning or something like that yeah. So I agree with that comment. um You know there's the way I kind of view it is.

B

um Is that the hippocampus is this sort of highly evolved system that was sort of grew organically over eons of evolution, where you know you're learning to navigate in an environment, then you're doing episodic memories and all this kind of stuff. And then what happened was the the basic processes that that are functioning in the hippocampal complex, which would be the anthronic projects, the sabiculum and hippocampus that those basic processes got put into a cleaner, smaller, unique form.

B

That became the corner of a column, and so it's like a it's like a it's the you know a complex overgrown bag of tricks. Hippocampus becomes just not another or a svelte, simple version of it that gets repeated. um So we were to see a lot of the same things going on. That's why we think it's a good cell replaceable equivalence in the in the cortex, but we don't have to copy all that messiness to understand the context.

A

Yeah well, thank you max. I know we take out a lot of your time and I should say you know you know you said you're inspired by us, but you're inspiring as well. Just the fact that you know you're able to do. You know you're able to run a company and then on the somehow on the side as a passion project do such a detailed neuroscience model uh is, is really inspiring.

B

It is inspiring, it makes me feel like lager, because when I was running a company I couldn't do that. I didn't I could my weekends was like I just like you know, decompress or something anyway, so you have to be able to do this while you're running a business, it's pretty damn impressive, yeah I'll um yeah! Okay! Are we down here we're going to stop.