Numenta Numenta Research Meetings, 22 Apr 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Numenta Research Meeting - Apr 22, 2019

Description

Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_

A

Yeah, what's the company yeah.

B

Marvy found this article that he was reading and he presented it last week we discussed it, I think Thursday and Friday yeah and there's like interesting ideas in it and so I think it's good to go and try to review it again, really quickly what it is and and then we continue from there. So you want to do that.

A

So yeah there's a paper from it's from like 2017 from Cape Jeffrey's.

B

Crew, probably a dual axis rotation.

B

During the movement in three dimensions, yeah.

B

A

Think there might be multiple papers under that now. I'm trying also to with Dennis, is research. Articles I mean.

B

It was some special edition.

A

Yeah, so this I saw it a while ago, but last week, I. Finally like understood, update rules and like how to update orientation under the theory, and it's actually really nice in it, and it it poses head Direction cells differently. Then then, we've often talked about them, so a head, Direction selves I mean as an animal if the animals on a like on a flat surface, unlike if they're, if they're, basically, if they're coplanar with gravity or coplanar, with like the perpendicular to gravity directions, cells of course behave in like this way.

A

We understand where a given move into movement. A given rotation to the writing, causes the bump of activity to move around the ring with somewhat fixed amount of that movement. So.

B

Yeah then you can you.

A

As the analyzer can take those neurons and sort them visually like this, so that it's as if a bump of activity is moving around them and so one way that this could handle 3d one way this could handle animals, for example, walking up hills, walking over hills around hills, or you could also just even talk about flying animals going in any or any robotic system, that's going in any 3d orientation, but one way that that could all work would be just a simple projection down.

A

Just do this last week, I'll just use that picture where one way it worked is like. If the animals facing like to decide what direction Dukes house to, is they just like projected down to the to the like the earth plane and represent that angle of a and as the animal rotates around here, this rotates around here, but one complication with is once once the animal is not coplanar with earth. I'll call it the diversity.

A

As a beautiful point, then, these updates are no longer linear, where, depending on what way the animals facing on this inclined plane, the the projected Direction is how much it updates with the given rotation varies like it'll, be much larger up here, it'll be much smaller here and so that those updated rules become complicated.

A

So this this paper proposed a different way that head Direction cells could behave in 3d and it has data supported. It supported Adam, Marquez.

B

Kemeny yeah I think he said it clear last week bit another way of saying we just said my name, I believe, is to say that we one way we thought about doing 3d orientation. Is you have a bunch of 1d orientation modules that are different rejections of 3d or adjacent space?

B

Just like we imagined I could get impaler's Mirko, and so that's one way of doing it, and this paper proposes a very, very different way of doing it, and that is that is the boy doing mu cevaz last time is that you just have 2d orientation cells, but there is another thing called a reference vector which, in this case of the wrap, would be private. So if you know the orientation of gravity to your body- and you know your 2d orientation- that sufficient Sarno your 3d orientation yeah.

A

C

There, you know, is the angle, you know clumsy.

B

On the angle, you can.

C

Compute, that's another.

A

A

So in this production version, yeah rotating to the right will not cause a linear change to the head Direction cells, except in a special case, we're where you're on the plane, where your co-fighter with the as a neutral point, then what they propose is it's going to involve, bringing in this notion of a reference vector or a gravity vector or something something like that.

A

But what they propose is no matter what plane you're on right now, relative to as a new little plane, if no matter what how you're ending up on a hill if you're, no matter that is wherever it is as it rotates on his current plane, the head, Direction cells will update linearly and the way they make this work is, is the rat or you are always aware of some reference vector and in a plausible one.

B

In many cases, this gravity wouldn't it well, we got into the case of that. The head, Direction orientation circle was that was the line with the you say, we don't they're still yeah like it's like if you're climbing up in a while I could join mine straight in the face. If the rat is like climbing up the states and rotating around like this it'll.

C

Always still be totally yeah.

C

Okay yeah, so it's literally representing your orientation on that.

C

A

Yes, yes, so so these head, Direction cells have to update rules. One is the one we're familiar with rotating clockwise, causing sense.

A

Rotates 90 degrees. The second.

C

Update rule, though, to make it where this all works, because that.

C

A

Yeah and if you any, if you only had that update rule, then you can get all these contradictions, we're different paths if there, if you're walking around the hill or on top of the globe like if you, if you go down here and then up versus, if you just go up you're, going to get these contradictions where different paths to the top of this globe.

B

Will cause a different head, Direction cells to be active for a given, but we want. We want half integration to work, yeah with orientation and so.

D

That anybody come back to.

B

The same point of the goal, the same position.

D

I need everything.

B

Orientation that would occur that right, yeah correct, so the second update rule here is city. So yes, as you just said, all we've talked about so far is rotations on this plane. The.

A

Other set of possible changes in orientation is to rotate this plane in various directions and the update rule is okay, as this plane the plane that the rad is on rotates relative to the gravity vector the reference vector if it rotates into the vector like if it rotates toward the head. Direction cells, don't update that has no effect on the head, Direction cells. So if the rat walks.

B

Like up here and then like one way to put, it is if thee, if they've, met the direction I'm facing. If, if the plane only moves like this and like this in the direction of placing then any other any other change, it would yeah.

B

So there's yes in my relief, the rat walks, if it climbs up.

A

The wall climbs up on top head Direction cells just stay constant. If this plane rotates around the rotates about the gravity vector than the head, Direction cells rotate by the same amount, they rotate linearly as this plane rotates about the godly vector.

A

So if the rat-like sidles around this corner, that's a 90 degree rotation that causes a head. Direction cells to rotate by 90 degrees, technically I think maybe that'll be counterclockwise, but it would rotate by 90 degrees- and this, like these, just very simple update rules of just update the head.

A

Direction cells linearly, both rotations on the plane and rotations of the plane about the gravity axis, give you this path: integration, property with orientation where, where any path to any point on this cube or any point on the sphere, okay, something I don't know yet is how the upside down is handled. I, don't I, don't have that intuition, yet I don't know how well handles there are some.

B

A

There are people on twitch right now.

A

So but yeah I mean I, could just state like the confusion about the upside down, is like I'm, imagine walking, okay, the rat tilting. What's this way keeps walking straight? What's this way keeps walking straight Direction cells don't update because it's tilting into the gravity axis it walks around. The bottom keeps walking straight. That would suggest that that that that these are the two direction north directions, so that I could make a case that okay on the bottom, this one should be active when you're playing this well or I could say.

B

A

Yeah or you could make the case that okay, if the rat is right here and it's what it's strafing to the side, it's that sidles over the corner. It's sign steps down this way. Well, it's plain as sill, just rotating relative to the gravity vector.

A

So if the rat is like walking like I'm like here, then I'm here and I'm here, if the rat swap is walking down this direction, it's silence down to the bottom, and now it's facing that way and now I've just made an argument that the cells should be active when there are facing this way at the bottom. So I commit I can make arguments for both directions and I. Don't know how to how the.

B

A

Phase they in the discussion, they said, we've left this part out and they address it sharply.

B

So these, in order to represent the proper orientation.

B

So I'm assuming and I, don't update, addresses in the big room. I'm assume we never have seven grid cells in the anchor independently.

A

Okay, I mean you made a leap there. Then you brought up a good so.

B

A

Think it's Ellsbury no you're talking about.

B

Sergeant structure, so just one module like that, it's efficient and so I mean what is the solution to that? There's. A solution to combine that, with with the property solution, did have multiple orientation modules that are angry differently in Vanessa Feist.

A

Talking about this, if I had I, don't.

B

Know yeah I'm just trying to hang on to an idea of accounting for if I had multiple orientation bottles and they anchor it independently, would they sufficiently encode 3d orientation?

B

B

Separating out the update rule, which requires a reference vector yeah from the I, wanted represented this for a prediction purposes.

B

You know: I need a cell block addition to represent minor in 3d orientation and do I do that by combining what a single with orientation module with reference vector or do I just take a bunch of orientation modules that are outdated using these new rules. So.

A

I guess I I can answer. I can address that question indirectly in a couple ways. The one one thing to keep in mind is that the rules that we're talking about for updating these requires this ring the spring of cells to be able to read out the gravity vector.

B

But when I actually need some other encoding in colorful column how's that being done so the initiative, the.

A

Logical just first thing to reach or first first thing that comes to mind and I had no idea how to draw a sphere on a whiteboard, but basically not a 2d, Rainey's idea. Thomas. Imagine a ring. That's like the surface of a sphere or like there's a sound here. That's out here, there's a year etc, and this represents like the direction of the of the gravity vector well. How's that ring you just seen is a spiced outer ring it's this sort of like a ring in that, like.

B

D

B

Representing yet but you're saying I mean so that's imposing a lot of stuff.

B

I'm not sure, that's like a in some sense, if, if I have that gravity vector and I have back in coding, you just talking about there, that is a 3d.

D

B

Of the relative search in the gravity vector, it is.

A

So I mean this is like a 2d located. This is a point on the sphere, so as 2d- and this is the third ya dimension, you could also have I mean like this is the this is the 2d version or you could do like the latitude and longitude? Essentially, those are. Those are a little weird moogly, because it's like it's like a ring and a half a latitude and a longitude, you don't yeah yeah.

A

This is like a ring and a half again I, don't know about I, don't know I'm going to think about these, like.

A

B

Need one the gravity vector is: is.

B

B

Like it's the two of those things that and then the ring who provides the last one, it's the offers like it's turning like this yep, then so that's one way of looking at, but then that begs the question: how the gravity vector coded and how do I take this population from you, I want to know is who the words or a population of cells.

B

This is probably because I need to be able to read out which allows me to make some predictions about my input and.

D

So if I have a ring.

B

And I have this to dereference belong to I mean.

A

I said posted a like and then otherwise like a third option, that I don't know if I'm the third option, that I'm pondering, if it's worth talking about something that comes up in robotics, a lot, a way that that that 3d orientations get represented in robotics as through quaternions, which are these four dimensional numbers that make everything work really nicely. I'll come back to you.

A

If that's worth our attention right now, it may not be that it may be the case that we don't need to think about those other than when we're when we're figuring out how these cells all need to be interconnected, but it may not be necessary to talk about those, but anyway, the point is there is another org and I'll just a quarter. Nians is another thing, but I don't know. If we'll actually need to talk about that, so there are a few options. This is what I'm thinking about right now is exactly like.

A

If we do implement this as a set of cells, how should they all be connected to.

B

Each others, along with motor commands, so that they okay, where does this money, come from well structure? Doing that yeah they just you know, I keep pointing to things like okay, I'm, trying to structure. We know about the court in cortical column and that may not be right, but I want to try it well.

D

B

D

B

Good 680, representing that to day you know, we think of grid cells, it's a 2d representation and we need to somehow get the 3d here's another 2d representation which, if we add orientation we get the 3d. So maybe nothing out. Is there a set of models now that represent the gravity vector.

B

And then you know then I'm laying on top of this orientation on top of them. I think we're talking about.

B

The same structure, and so maybe they're both like little modules or module of things where you get a unique representation, and so maybe one way cigarette, maybe though maybe one of them is one is like the 2d gravity vector and the other is 3e linear, vector of itself and then maybe orientation slices through all thing, thankee thankee I keep coming back. An idealist is trying to be crazy.

B

They've got this column, you got these mini columns that seem to have some sort of like you know, orientation, look to them and we don't know why they goes all the columns of all the layers. We don't know why they go through all the layers, there's some cells and all the layers that look like orientation every lady. Not all the cells is the subset of themselves, but in all the layers, and then we have these two things which look like you know like they'd, be good places, for you know like a grid cell motto.

B

Maybe the other one is this. You know from back of every gravity vector modulus on like that. You know.

D

I on somehow, because.

B

Then you have to solve all these things. A lot you gotta get off the orientation. You got resolved the gravity vector you got resolved with good, so hot. So.

D

Anyway, I said when.

B

I, look at this I go yeah he's got enough information here, but where the hell is this guy coming from what does it look like we're to the neurons I'm ya know if you're thinking at that level, I mean you described in this paper. They, instead of a very simplistic bare bones, so.

D

B

Questions come trying to fit this into this.

B

A

Source of ideas, at least, is going to be. Our vestibular system is almost is pretty much.

B

Definitely encoding something like this yeah, that's true for my head direction that it's not gonna, be choose my finger.

A

But just if we want inspiration for how neurons represent a gravity vector or a reference.

B

Flow these things so I would think, but I was thinking like okay I, don't know, I can, except in the end, so on the forecheck, so I said yeah I might be using.

B

The cortex that's not going to work at all.

B

My finger is moving around an object.

B

But I also have to have the same mechanism we're going to work for conceptual objects where it's you know, high-level flaws, so there's got to be more of a pure mechanism for doing this, and now I was thinking like well. You know mention I do I. Do I establish a vector or preferred vector for this cop, because you know I mean I pick the long dimension.

B

So is that something we do for objects we hold in our hands? We just use some sort of sensory input to decide what the preferred vector is. You know I.

B

Don't want to stop. You know, I.

C

Found it very difficult to believe.

B

C

Would have to be some? What you want to do is, as you made motions you wanna, build a sensory, the under light, and you know you take whatever sources of input and you have a generic system that learns transformations that learns how to take proprioceptive input.

B

And whatever input it's heading.

C

In and make predictions, and if you got that would give it the right input it like gravity and so on. It will learn this particular mechanism. If you give it, some other system would learn to use that to do the transformation at the end of it. You want a generic system that can learn transformations, and this is an example of a simple transfer me much more complex transformation for you to do for doing. Mathematics.

B

And equations name clearly yet.

B

My first assumption there's going to be the same transformation.

B

You know I just went through out of nothing which I agree, I.

C

Mean why would gravity important equations.

B

B

That make sense for that system totally planned changes, but not when I move my fingers and you can say whether they're appropriate something maybe that stays for so long but they're. Ideally, you would use a pure century employer so.

C

Mack my point was not dead. I think it may be a mistake to look at this particular circuit in the.

B

But Marcus's pointed you might get some inspiration by looking at these two verses. It.

C

Has to kind of maybe in a simple, is reduced to this, but it.

B

Would might be able to do a lot more than that I'll make oh I, see point you know one thing I want to point out: I've mentioned this many times the past and I. You know we think about the input to a cortical column coming into layer. Four right.

B

This is your input and and then we have this other one that beats, and so then there's this the here so say six be six day and there's five here, and we know that this there's this other input that comes in right around the border between five and six and it seems to be everywhere and and so in the past you might remember I wish I was trying to think of what. Why did you have? What purpose and I I speculate once before that this might be? Establishing orientation? I talked about that it might be.

B

Establishing orientation and I could I need to know my orientation but make any sense of it was. It was a tunable idea, like I, have essentially input but I need to know my orientation and that's that I always talked about maybe like there was bigger cubes I said this might have a broader, a broader receptive field so that you know larger area and then help you define your orientation. That was a very poorly defined idea, but if I am or specific version, that would be saying: oh no, it's not real imitation.

B

You might be just signing your reference right, which is not the same as orientation, but it is this key again. What would be the thing? I would need to establish a reference vector for a coffee table. If I were to do by pure sensory cues I can imagine it would be sort of its its general.

B

Its aspect ratio, something like that right markers, you think it's on them. I was distracted. Okay, because I had to look on your face, like you were thinking something good.

B

So I know I just don't mean this out again because again, I'm gonna stick trying to fit this world into this world. I'm thinking like well, if I can rely camera line, vestibular I can't I mean oh yeah I, don't what is going to be sensory, something sensory that I and I don't think it's gonna be this input, because this is the thing I'm trying to predict so there'll be another place.

B

I have to have sensory input that could some sense establish, perhaps the reference vector so anyway, that's just out there trying to figure out what that is.

B

Is intuition that this would be orientation, but maybe that was it was slightly under fine. Maybe it's disruption.

A

Again highly speculative, but something to think about I mean it makes sense in the sense of since we know like down here. Receptive fields are much longer. Yeah like like, like 6a those cells as like as an oriented egg gets longer and longer the 6a cells respond more and more so I.

B

Mean I remember them just having a wider area response.

A

They do respond more.

B

Nearly one just say: hey.

A

You're singing, actually, yes, big, let you go out well, okay, that would be the have.

A

The ambiguity here is they're averaging over a lot of cells. It may be that, but but it's certainly the case that, like that.

A

This, if they say so receptive field, is this long area, and it's responding to that particular particular orientation here to get horizontal like this will cause the cell to respond, and then it's response gets more and more rule at least reliable, as this gets longer.

B

Remember it was more like you know: it's almost like it was tapering off and they put some cutoff point actually for some white, not bigger yeah. That would be another thing. Then that's that's.

A

Exactly what you can walk or something.

B

Like this right right.

A

Well, it is true there.

B

Well, the differences.

C

Are small I'm not sure that.

B

Same property- that's just definitely smaller, so that resonated a.

A

Little bit just because that you can imagine reference vectors being calculated from something larger.

B

Yeah I'm trying to determine the orientation of this little line here, but my reference vector is gonna, be this guy, and so my reference vector is going to stay the same, even if I rotate my head. This way right, I'm, not sure how that all works out, because it's just weird.

D

B

Mean if I'm actually rotating my my my my head here this, what cells are active, but these longer cells would be changing so.

B

We have a comic operating on its own, or maybe not maybe you know, there's also the fact that the later five has a long range. There's long range connections in later five and remember, there's two sets of one reads connection and stay free and they are five and maybe one way to think about those long range connections would like by devoting on rafters.

B

Maybe to me this is a great because the market presented this I said: oh, it's real high over L just been stuck on this idea that we're going to represent 3d orientation with a bunch of orientation modules and that they're all being a projection on the 3d Orion space radiation studies- and this is just otherwise another thing that came up last week, which I want to mention. So you can take private sua time is.

B

We saw the issue of representing 3d linear space fits all model space and again, when sumption was ok, we're going to need multiple original module has then barkis's and Merkel's paper.

D

B

I was pointing to my keys last week: that's not what it looked like in the we don't see any place where he would see all these different modules within a column in invest. I can look at. This is my. This is my recreation of the tank drawing the tank paper where in amount of supplies with cortical column, he was showing in the courses into one the cortex and.

B

In about a size of a quarter column, you had one who said he had one basically.

B

And it's not like, where would I have multiple grid cell modules in a single quarter of a cop, and so the question then becomes well. Maybe there is, maybe you know, is the same sort of check. That's being done here. We have this reference vector. Could something could I have a reference vector to this would make this work? I haven't even thought about that. Just like that's another, so going on from the one extreme, we have multiple grid cell models, but there are different intersections of the different projections of the 3d space.

B

We have both of our inpatient modules, projections of the orientation space. So now.

D

We have like, oh no, we don't need multiple.

B

Orientation models: we don't issue modular reference vector and maybe the same thing can happen here. You know there is no module, have a coolant reference vectors on how it meets three eat. That's very, very high-level concept: I have no idea how that would work, but I just think we don't see multiple law of these guys.

B

Do you have any intuition Marcus with when I see those were first, you go like oh that's stupid or that could work or what was that is.

A

Something as a before is that orientation is a little bit different from.

A

B

A

I guess the other thing that comes to mind. Something you've pointed out that yeah, so the current state of this of all these cells is essentially a point on it. Promise like the manifold of the cell activity here, is like it's moving around around this and wrapping around. However, as you pointed out that there's also this like this notion of like a mist you'll, sometimes sometimes one of these will be higher firing, yeah, and so that can be thought of is like. Suddenly, this manifold of activity is becoming three and really that came.

A

Second, that's not necessarily just 3ds, because I'm enough variation, just like that pit and then a fold of activity could like going some and these other strange directions through allowing these to be a different magnitudes, so that there's some flexibility there, but I don't know what to do with it and I. Don't know how to give it like this path: integration, property, like how do you? How do you take these missed bumps, and these varying rates and encode something.

B

What is reference vector was like a key thing in all of us. What, if excuse me if I already said this I'm sorry I can't remember again going back to the many columns all right. Is it possible that the many columns are including reference vectors and that that's the thing that's passing through everything and then, in that regard, to my sister grams? This is orientation and this is grid. So then you have this with expensive all like what would I expect to see here right.

B

And so what would happen if.

B

Targeting me translate like quit, they would be the property I expect to see from a reference vector.

B

What I'm thinking of myself, if I change the speed at which you I'm asking the question? I just showed you a picture and there was just some line it's morning. You know some grading right. Well, it might just look like it's just the orientation, but if IX you're shooting a real picture of something or a real object, you're looking at an object is changing in different directions. Would these cells behave differently, given this model of the rough inspector.

B

Subset right weibo said: oh all, the many consoles have any kind of fire. Now we're representing a sub set up. Now, that's how we've talked about it. There's a lot of evidence that suggests it's more complicated than that.

B

So I wonder if actually there's even more going on here than we knew and that it's more like what you just said. Maybe they seem to really be changing in other ways too, that we we didn't understand and because you know, even though our model would we say yes, there's a subset and this cell. Now it's going to fire a specific location of the object or some specific location in video or something like that, they're, not reliable. In that regard, they seem to be there's still weird wonky. It's sort.

D

Of like that yeah we see that.

B

I remember seeing all the points I'm seeing that's quite right to either like doesn't always seem to do it at the same time. I, don't know why I'm just wondering where and then this is this appears to be just a one-dimensional encoding. Then you need a two dimensional encoding for the reference vector.

B

There's a few ingredients here and they're going to be macking on to this. Somehow you may there's no I just want to throw that as another possibility that perhaps the reference vector is the key thing for resolving the good cell location and the key thing with resolving their orientation cell.

B

You know the idea here that maybe what if I was in I had a grid cell map. It's also wrapping around the sphere.

B

It's almost exactly the same way, and so with my ingredients would look like, like they're, already corrected some plane, but as the plane moves around the array to the plane rotates, then they you know so so I would somehow not differently. You know just the same way like in cases all I see the point of it being an open space versus a closed space back.

B

This is anything about this new way of thinking about this market.

B

If I, if I, had multiple, if I had multiple rings like this and one vector but I, had multiple grid cell on orientation, cell modules and each orientation cell, what we wanted to do is the each orientation Salman do they do were path. Integration means when I'm in the same spot, on the same orientation on that at the same celebrate. So if I, even if I had a bunch of these and they were- and they were anchored differently, so they all had that same behavior, but they're anchored differently.

B

Then so they say there were n modules and each one had you know and positions. Then.

B

Then I would have entered, am possible, unique orientations right so that I'm just saying could I have in a representation. Can I still have a representation of orientation that you need to the object? Yes, it'll be the object. It won't be unique to the location. Yes, okay, that's that's! Okay, at the moment,.

B

I get that. Why did you say that so rapidly? What was your conclusion for that? Well, just the well just that movement, not in rotation of movement, isn't going to update these oh yeah.

B

Unless every orientation moduli this.

B

Yeah so you're saying non rotational movement, meaning like if I'm moving up this way. Nothing will change doctor saying.

A

You know, honestly, what you said was correct that wasn't what I was hanging up I, just my regular movement like if you walk around the room. These aren't changing, but you're also right that, as you walk as you walk toward whatever toward the gravity.

B

Vector it also won't change.

B

But but it would be specific to the room and then this is the difference situation or them I, wouldn't know where I am on that object, because.

B

And if you walk around yeah but but they would be changing.

B

I'm just doing this, but this is the preference vector and I'm just doing this.

B

Then then none of these modules would be changing. Is that what you said like if I, if I rotate up towards the whole unit, not a birthday correct, that's weird, because then for the guys? Oh, so yes, so we're all! Starting with this and I just get super super super speck.

B

The words you could have you can have a reference vector: okay, that's going to be stable for the object. That's going to be essentially anchored to the object.

B

You're gonna have patient modules which could be. Would you be unique to the object.

B

But not unique to location, I.

B

Multiple ways, then you can have you know a good seller location.

B

I'm wondering if I'm just if I took tanks I did here then I wouldn't even have unique locations on III couldn't Theory have the same. Encoding I would have the same encoding at different places. This wouldn't be unique. This is just one module and it's it doesn't it's not unique, so it would be it's. It's not unique.

B

Object and it's not even really unique to the location, to location on the object. It's for most cases. It would be sufficient, but not at all. Then we have. We have these three things and the question is that these phrase things together sufficient to do everything we need like, for example, I, have some location, but it's not unique to the object, but I do online orientation which is unique to the object. So it's combine those two its community object- and this is.

D

Not unique to a.

B

Location, this is not technically in each allocation, but it'll be more unique than this, but maybe together the chances of you know a set up would be. Maybe together would be sufficient.

B

So not everything has to not all these representations have to be unique to the operative vehicle. Occasionally it's the.

D

Combination, that's.

B

Important this one here is not you need anything, it's just its anchor to the object, but it wouldn't be unique to the object or you need your location. So it's uh it's just anchored.

B

Yeah so I'm thinking like multiple orientation modules and I got multiple good saw modules and I have reference vectors slicing down through them in the intersection these the intersection where I have to resolve the same thing has to resolve in here and here, using the same reference that there is is going to guarantee that these guys are in cahoots and everybody's everything. I need the purpose that is being represented.

B

And they could be taking advantage of this kind of trick.

B

You know again I'm imagining. Perhaps these are like mini columns coming down through here that you've got this you've got this, whether it's a grid, so module orientation module, maybe they're just the same like this and then you're slightly I'm.

E

Really bad when I have my money.

B

Columns that are saying, okay, this is this guy, and this guy are the ones we want to care about right now, that's coated, that's say: okay, I, don't know my location specifically, but now I do it, because it's just unique pattern that I'm encoding here, which is the intersection.

B

Element solutions, gonna, look something like this: it's some sort of flavor yeah.

E

My like my mic is way better I.

B

Should be for a while here, so if we talk you have anything you want to talk about other questions or observations.

C

I feel like we should be looking for a system that can learn these things rather than the hard coded system. That's my bias.

C

Network can learn a lot of different methods. Low these transformations can be learned by rights to the extent.

B

Of it, you can think.

C

Of this as a function with input that gives a certain output given the same input you own the same up, but you can wonder that internal.

B

One of the things I can push you back on and not completely, but there's someone thrunite.

B

B

That I think it.

B

B

But there isn't, we don't require all those two populations of cells and I need to do the coincidence between those it's it's like you know like this will be there's just one.

B

But if I'm trying to fruit apple have a dendrite, they said: okay, I'm going to respond when variable is to environment. B is a fix, and these are two separate cell populations in different different places and I make sure this mess. It's just not going to have the equally sample both of them. That is where I get nervous. I, don't like that I'm just putting out the best. That's one tool that okay.

B

So to me, one because a 4mm cells can do these things. That gets a little bit learning.

B

A

The past often when I have used this trick, I said I justify that saying that this could be genetically determined connections like.

D

A

Often in our circuits there are portions of the circuit that could be genetically determined in other parts that can be learned and I usually try to put these tricks on genetically.

C

Just need you need to sample.

B

C

Slightly larger cone, but.

B

How do you guarantee that you're going to get this equally.

C

B

Of them, okay,.

B

So I just want to point out that I keep coming back to the there appears with these mini columns. There appears to be a structural computational elements. It's the only way I can describe. Why I think of why we would all have this. Why you have these orientations.

B

It's like there's some undistributing. This is another bus, intersecting, these guys and then somehow seems to be doing computational parsing or something. What was like it's hard to explain why you do it otherwise I mean a while there's a sample a set of cell, like you just have the set of cells appear connections. No. This is like forcing a computational structure on the thing. That's why I keep coming back to it. It may be wrong, but.

B

So I think you know you can learn it, but I also think there's just there's a structural component. Maybe someone who's genetic, so I mean obviously have to learn it. There's something.

C

To that again, it's just a push-up. What fresh, yeah, yeah I think.

B

I'm, keeping thinking that they do the right answer is going to come out of somehow understanding why we have the miniconjou all these things. So.

C

I've always been looking.

B

For what would the many columns be representing, like everyone could use an.

E

Assist a big question limitation.

B

Okay, that makes sense because they look like and visually they look like orientation codes. You know, and but maybe isn't the new idea today or in this last week is that perhaps it kind.

D

B

Like a teaching, but really maybe it's something differently, like the reference maker, which kind of retrospective might look like orientation under many situations and then I'm thinking like what would they then? What would my resale response properties be up here? If I sell? No, it's not orientation thinking here and then maybe we think about that. More I can come up with some early because.

D

Then we can go back in the literature and say.

B

Oh, maybe it looks like under the default case of this. You know whatever that look like orientation but mentality these situations. It would be something different.

B

Yeah I write and then, of course, I'm jumping the conclusion that made me the same thing: that's going on with orientation being thinking about going one rotation model moving around the sphere with a baby. That's the same thing, the other way the way gives.

C

Reference spectra and.

B

Then maybe the grid said we only keep in one group called vector R. It's all modules, doing the same thing.

B

This may be the I, don't know, sometimes all of them together it's going to provide you where you need to know.

B

So we had a whole new thing idea.

B

We were walking through some of some of the structural requirements.

A

I mean by my next step is quite easy to just draw as like, okay orientation like motor going in you have a ring. We know that plus question mark and then some kind of recurrent connection well.

B

The question is how.

A

Do you represent yeah and does everything interconnected away with one with motor commands so that all this can update and and so.

C

This is would be a location vector now, but of this would be orientation.

A

A 3d orientation, okay, oh.

C

Yeah, sorry, that's the entire thing. Yeah.

A

So I mean, but this orientation good Paul I wasn't clear about that. So that I mean this. This cell population represents a 3d orientation and it is, it has a set of update rules as a procedure motor commands, it updates itself and just what's the. What is the logical way to break that and that's all population is.

C

A even assume, there's some reference vector coming in as well. Well suppose suppose this.

A

Question mark represents a reference vector so so.

D

B

Going to update the.

A

Reference vector as well.

E

D

A

Like guess, maybe an optional.

A

Yes, yes, suppose I give this its initial activity for free I'm, like I'm gonna, tell you what this is an Italian with the reference vectors and yes, of course, like I, could put a timeline here like vestibular.

A

That was my question like they can like. Sometimes this system will have something that tells it its reference lecture or it gives it hints about this reference vector I'm focused on the what we for lack of a better word, the path integration.

B

And then, what's going back to thing, you like this input here until they're fired, which is the sensory thing? Well, maybe maybe.

B

Maybe the reference vector is, it can be. I always want to be the same reference vector on one on at the same location, on the object, but it doesn't have to be the same reference vector at different locations.

B

So so my point is that I'm, a column I can only see certain part of the world. So I have this this larger perimeter and have a smaller prick. The lord of rimini, this input a smoke, and so so, like I, can't see everything. I can't know when the true extent of the object is maybe I can go, I don't know, but the question what would happen is only I. This was a somewhat a local computation.

B

That's my reference vector could be a local computation, volu tons of some broader area of which I'm touching or something but may not be. Global's adopted. We've done with that problem. I know in the wild cortex we think of grid cells, as always each other in the same orientation anywhere in the room. So in that case, maybe through entire room has the same thing but I'm a finger touching an object. That may not be the case.

B

I'm finger, touching an object, I, don't know, I know it's just another fun I'm, just starting off this off my head.

B

Yeah this markets left, we got the share of there. This.

D

Reminds me very very much what.

B

Happened when we did we started doubling in the grid cells was like. Okay, you know Marcus, dug into the living room, grid cells and came back all the stuff. We did.

B

That somehow in the same process a month or two so really.

B

That maybe the same thing will happen here. Let's make it a bit obvious.

B

All right answer keep.

E

Typing because I forget I can talk, I was excited.

B

Okay, so we want to learn it, we got it and not rely on the stupid letter. We have a good line, some sensory input. You might have voting about it. We might have voting when they find the vector I could be what's going on the live voting which we haven't done. You haven't really had a need for yet you know they said why we're voting ahead of time.

B

That would be another thing that all the top local comments, all over columns could vote on the orientation.

B

That's an interesting idea.

B

Okay, thank you. Thanks fun.

D

E

Gonna roll up here and get a better view of the.

D

Whiteboard, oh good Marcus.

E

Would you driving a chair.

D

E

It's a couple pictures and get a good view of.

E

I'm gonna exit the beating, okay.

E

All right, so you guys got a snapshot of that hopefully, and that is the end of the research meeting. Thanks for joining I'm gonna go ahead and cut my cut. My stream and I'll be back at one o'clock to talk about responsible, AI, license and check out my events calendar on my twitch stream. It's twitch.tv, slash, rhyolite, underscore slash events to see exactly what I'm gonna be working on. The rest of this week. Take care.