Numenta Live Streams, 10 Apr 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Timing in the cortical column SMI circuits, whiteboard chat, neuroscience, artificial intelligence

Description

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

A

A

All right, this yeah, okay, so I, was trying to work out like basically turn this cortical column into a software diagram. Yes, okay, so I was thinking about. You know two three question.

B

You're trying to turn the diagram we talked about the other day which, in today's research me I'm questioning parts of that right. So do.

A

You have that in mind, yeah, of course, keeping that in mind I.

B

C

With you the other day, I said this is all speculative, so we don't really know what it's going to look like, but it's something like yeah.

A

I think it's I think it's sort of useful to do. Okay,.

D

Maybe I should draw it in this awkward diagram because you already know yeah, so this is sort of how.

A

I'm, just thinking of it as data in and data out, first of all, yeah we're talking to software, so sensory input in coming into layer. Four and this they're all gonna be very simplified: okay, yeah, okay, we know layer, five is over here and there's motor output yeah. It's a super simple circuit right.

B

Well, it's not.

E

B

The way that's something goal of the circuit to have a motor, oh sure,.

A

Sure but I'm trying to trace the data flow. The flow of information through.

B

The goal the goal of the cortex is to build a bottle world right and not to getting input and do something right. Building.

E

B

Model world involves movement yeah, but it's not like I. Don't anyone to think that hey you got an input. You're gonna do good a lot of the world and the model is moving input around right. But it's not to do something necessarily yet yeah.

A

I can see that okay, but again I'm just trying to trace the data flow and so and also as I, was trying to separate because there's obviously this is a. This is a driver signal. There's that's.

B

The driver saying oh, there's no way we're gonna, throw that you shouldn't put two three in. They should just put three okay, nothing that doesn't drive, but we, if you're, going to draw that box, sweet box, five, it's it's lavery, they're, fine,.

A

C

A

D

Let's put I was gonna put six.

A

A right six day right, yeah.

B

Six, a down here.

C

Yeah and there.

A

Would be another driver signal.

C

Now that driver yes, the other way is there not a driver, and this, where is not.

A

So this would be yeah feedback, modulatory, so yeah, yeah, okay,.

C

A

Yeah- and there is also a there's.

B

A no I'm, not really actually I. Remember we did. We have that in the individual columns paper, but we knew that wasn't quite right. Okay and you don't need it anymore, if 6a representations in the in the grits of modules and are unique to the object. Okay,.

E

So there is actually physically in most parts of than your cortex there's no evidence for three to.

B

Four, there are some places where people see it, but not much. Obviously,.

C

That is that I of that.

E

We're going on a separate.

A

B

Going back to for yeah.

A

Yes, yeah yeah, so so, let's move on to here.

C

E

Gonna be a driver, transit.

C

Driver that's well.

A

Documented and then we need 6b, that's.

C

That's the current thinking is great. You know in the discussion we had the other day, essentially the same.

A

C

Anything about that right.

C

C

Sixty-Five I'm, not certain about that second signal: let's keep going that's.

A

Talking about, but this would be location and.

C

This particular version of the model, this very location and orientation. This would be place because.

E

Here's a here's, the thing you need Matt when I move as I'm, anticipating my movement.

C

E

Right I have to get that anticipated. New location asked to leave 60, so 60 represents also the new location as a who we're.

A

Going to make a prediction here right, but.

E

That's not I need to do more than a prediction: okay, because it has to get all the way back to four hey I am from.

A

B

So if I just predict.

D

B

Other five: it.

E

Doesn't generate any activity, it's just a prediction: I.

A

Mean this will.

D

Eventually, get back to four right and.

B

Well, no, no I'm thinking in the column in the column, so here's here's the general idea that we talked about the other day. This is just a hypothesis yeah, okay, that's okay! You got to and the idea is that five could be better than the two.

B

See this is where it starts falling apart. Let's.

C

B

B

Let's go I, don't think it's true, but let's go with it for now. Let's say you had this: ok, ok,.

A

B

Then good work, you would say: oh as I.

B

Activate this guy too,.

D

B

Asking us to film things that we did. You know we haven't done before so, but.

A

B

I have to get a prediction back to. Therefore that's the goal. You also get a prediction back here and there has to be a you. Can you can end up with a prediction input before, but.

E

Everything else.

B

Flowing back, which has to not be a prediction, has been activity right. That's all the ways you wouldn't get here. The prediction stops. It doesn't go any further until the next input comes in right.

A

B

Well, under this scenario here.

B

Three would be what's actually occurring in the world and two would be a sort of a prediction of what's going to occur, based on my movement, something like that right and it.

E

Still has to get back.

B

To four, so one way to do that you could- you could have this like here like like this, but.

D

If I'm gonna do that, I'm.

B

I'll just go through this I guess: I'm thinking about it Michael is it three to predict for the Jews? You can predict four and then I would be done. There's.

A

No evidence of that as.

B

I said it's kind of screwy there's. If you look at.

B

The car in the classic descriptions in the in the brain and the literature does not show in three and two feeding back to four: that's a bit misleading because you actually look at the cortex and you see layer, four and you see layer, three.

B

What you see is the cells or a classically layer, three cells or sort of there's a lot about this lower border here and they send the dendrites down into here or they send their axons and their dendrites come into the top away of four. So there's a at the border where what's costly at four is getting. You puts the bottom, the Blair three and and so there's a blurring here.

B

But if I look at the Bible and for the top away three dozen, it doesn't appear anything. So this is a unknown. It's all wonky.

A

B

Also possible and the peas totally possible, with ourselves in layer three dad sends a projection of their axons down the layer four, but we just haven't seen them. Yeah or people haven't looked for them.

E

B

This into the unknown category here yeah, well, we can say for certain, is somehow a prediction has to get back to like four, so that I can city and.

A

After it's made this this loop and gotten this information, but.

C

Location incorporated.

A

Into to get both of these sensory motor sort of loops, yeah.

B

Talk to me about timing, but yeah.

A

This, maybe you should try.

B

Putting like one and then a to.

B

Comes up in here that activates, a bunch of cells here, yeah.

A

That activates, a bunch of cells directly projection.

B

Yeah yeah, would you have to bunch of cells here this all time? Yes,.

A

C

I'm stuffs time long time to.

B

Rapidly five milliseconds type of things just.

A

This happening.

C

At the same time is three even or.

A

Or they do that well,.

B

That's just one but comes in and Bingbing Bingbing Bing yeah. Now.

D

B

This start there, okay, at this point in time, I'm assuming that 6a and 6b already know where they're going where they are. You already decided where you're moving okay.

D

B

The command to moved these things are settled on their new positions. This is automated prediction, so start.

A

From the last time,.

B

Yeah so we're strong we're developers, we we have we've anchored all in the 1660 or now the new locations and the new input comes in yeah, so it's been predicted great. They can write a play for that. Specific representation goes to layer. Three layer, three protects a layer; five, that's classic. Everyone neuroscience book says that yeah, this back projection from five to quote to three is also well documented. So in this case now in this particular hypothesis, five will project back the two, and so that would be time five and this wouldn't change.

B

6B 6b is already there so with.

A

B

I might be just.

A

Below this mate already, you have incorporated this.

B

A

B

A

B

This signal is more like trying to figure out where I have in the world yeah.

A

B

If I don't know where I am, if I give it a Miss and put on four I'll pick I'll try to drive the cells in layer, six, that's it! Oh, where might I be right if I see a buddy unique feature, I know where I am in the world, but it's not like part of the flow of things. It's contribute.

A

It will call this.

B

Anchoring this area of anchoring these things it's it's a way of like picking the grid, some modules that are active yeah. If I'm saying about that.

C

A

Yeah see if it's not it's not part of the floating.

B

By it's more like okay, I'm angry at myself in the world, I'm rich.

E

A

Cells should be.

B

Active now and once I'm angry that then just doesn't do anything that makes.

A

B

A

The flower go away.

B

I got some feature at some point in the world: I now create a stable representation of that feature. Independent of orientation I now assign that to my model of the object, which is location and things at that location. Now, if I, if I move to a new location, I'm going to.

C

A

Going to affect your model of movement within the space.

B

Refer to scene would be something like the following: I'd prefer to see that.

B

Well, you could do this, maybe two different cell populations. Well, that's for the moment say this. This then, if I move, what that's going to do is going to then update these cells to be the new place right. So this drives this. But then my movement command moves us to a new thing. Then that's going to tell them your place. My new place is going to be. This is going to be my predicted. This is going to be predicted version. This is going to be, they get the actual version. This is a predicted place.

B

This is my actual gameplay and, and now with my prediction, place, I can put I can protect them now, I can actually. This is these are active predictions and this is Asylum fiction. So now I'm depolarize the right cells here and anyway in the dargah. In the talk we had the other day, that would be the flow Bing Bing Bing Bing Bing movement occurs and then I go Bing. Being prediction I, don't think this is. This is on any any green thing. Is a silent prediction, meaning.

A

It's not a driver, meaning.

B

That you can't observe it like ie cancer right if it's, if it's internal to the cell and its depolarizing, the cell, like what we're doing here. Oh yeah,.

A

Yeah, the cells.

B

Can't tell anyone the depolarized.

A

Yeah right they're.

B

Sitting there in the predictive state, but you can't know that yeah.

A

B

Most of the predictions in the world you make a not aware of because it's just these cells are being the primary.

D

B

Very confident that, but there are I can I can say hey what is it I'm going to feel like I put my finger on this edge here? I could anticipate so if I can actually imagine the feeling that means there has to be some cells, farming.

A

Yeah right, that's.

B

It I'm imagining that thing right.

A

So shadow appears this would be.

B

An imagination this would be I can actually read this out. This is a cell population. I could say say that what is that? Oh, it's an edge because there's some cells that are firing here, yeah.

A

B

I, don't want to confuse that with what's reality right yeah, that's an anticipation, yeah and it's an active participation. It's an active prediction. This is the actual thing that an action really is in the world. We don't want to mix those two together, right and and so, but down here, I I can't I can't sense. This prediction here I mean this: is an active cells here, I'm, just a polarizing myself on so I. Can't I can't read that out. So the idea here is in our in our models.

B

Both of these guys are you know these are these are cells that you would have the ability to perceive these are cells that are projected off the cortex and so on active predicted next thing. An actual next thing, I think.

A

It's interesting that these are so close together, yeah.

D

B

I imagine a ssin of a thing and the other one is well. This is a reason why that might be the case. They seem intimately tied together and the reason reason for that would be.

B

Let me just draw it over here. If I can do this.

C

All right, this may be a part of the answer.

C

This is for three and two yeah there's your input.

B

One of the things that happens here is that there's these sharing many common response properties across all of these things, right: okay, so that may it can play a key role in the solution here in the following way.

B

Imagine layer, 4 gets an input and layer. 4 input drives, which many columns are active right.

A

B

It's going to drive, which many columns are active both of these things right, it's sort of saying you guys are going to learn something but you're all going to learn the same thing in some way and so in our diagram. Here, I come back to here from layer 5 and this guy goes. You know to layer, 5.

B

What we might, what we're doing, though, is when performing representation to and form representation to either they're going to be sort of derived from the same many columns. If you will so so, I want these two things that represent layer, 3 is going to use the same, becomes a representation like I want to first that, yes, I'm coupling these teams.

B

These two guys together and the advantage of that is, is that what I learned when I learned when I have to learn a projection from five to two I'm, going to pick the same many columns that I used going from three to five. Instead, it's going to get you mean that, if I think from a mini-com perspective that this many columns active on the way over over this was gonna be active on the way back, this one's active, this was gonna, be active and then way back.

B

It's sort of forcing these to represent representations would be similar. Yeah.

B

This could be how this moment to go about. How, for is we don't get these projections from layer? Remember, we talked about the three doesn't protect back to live for. Yes,.

A

B

They do share the many columns. Yes.

A

B

Many columns are defined by inhibitory cells, these bipolar inhibit to ourselves, which span layers okay. So this could be the solution.

A

We should never move this because we're talking about how.

B

It could exist. Okay, when I say doesn't exist. The pyramidal cells in layer, three do not project ready to the stones.

A

That's not a projection. It's a sink.

B

It's not these cells per mo cells excites our parameter cells, do not connect to the excitatory cells here right.

A

B

A

B

A it, but that we do know that these these many columns span across all the layers here, all right, meaning there's a time between these guys and that time is, is going to be. The best hypothesis is that, where things diseases, these bipolar cells, or sometimes they called them, not the chandelier ones, the other one.

C

B

No anyway, it's called bipolar cells because they're a form of bipolar cells. Okay, so these are like these inhibitory neurons. It's that span many columns and they define the mini column. So in that regard one could say the following: when I pick that back to here, I'm going to invoke these set of many columns and I will invoke these said that many calms down here- it's not going to make our cell the cell basis, but.

D

A

Say this particular population.

B

Cells, it can invoke this particular population ourselves, yeah.

A

What I can say is that this.

B

Set of mini columns will invoke the same set of many columns here. It does, which is not a it's it's. It's not a very, very specific.

B

Prediction but there's still a prediction so.

A

So, do you think that it's sort of obvious that as many columns activated for that could select cells within these two layers to be available or somehow select them on the way back? This could set a different, selective, different set of mini columns that would enforce some activity later. Are we.

B

What I'm saying is that we want the back flow to make a prediction here: yes,.

A

B

One that we only think I'm a prediction likes from the layer 684. We think it is a very specific cell population to a very specific cell population. Yes, what I'm saying is that the prediction coming back, maybe a specific mini pop population to a specific mini column, population that make sense.

D

B

Fits the biology, I'm not sure the implications of that are, but that fits the biology. That's what I would expect to happen. If you invoke mini column here, it's going to invoke the meaning common across all layers yeah. So if a 5 could print it back here can say I'm going to tell you here's my prediction: oh.

C

This is really interesting.

B

Okay, we it's just gonna suck much more time to figure this out ten regions, but the idea here, because even these many columns, its fans, they should the smell all the layers. Every mother, every taking them layer seems the same. Many columns have basically the same so invocation. If you will it's hard to make that work.

A

In my head with these.

B

The key to making it work yeah, the thought here is.

B

Some of these predictions, some of these backflow predictions can be done via many colors. I could predict the next many columns, such as protecting back to layer two and its.

B

If I think about, and then I can predict the next minute and then it would evoke that depolarize the many columns in this layer of the cells and all these many comps in this layer. What.

A

I'm going to think about this if.

B

You, if you put if you depolarize the cells, a mini compote you're predicting is, is the input. It's like. It's a it's like a predict, the melody, the next one, the melody yeah.

D

B

Not predicting.

D

B

You know what, as long as it's a third we're good yeah.

A

B

Not saying I'm predicting the animals. Third, that's the.

B

Output of the spaceship oeuvre yeah- let's predict exactly this prediction we were doing from here. Here, was more specific than that. It was saying that you have a specific.

A

Cell population- or we think that Domino people are here.

B

B

That then it's always bothered me. Why do we have these many columns across all the layers yeah.

A

B

Makes no sense, I, never really liked just I understand why we have them in our in our in our sequence, memory algorithm. We have a reason why they're there right it's a beautiful idea, I'm almost certain to correct people love it, though science people really catch on to that they all that's a really great idea, but.

E

A

Have the same mini house.

B

Everywhere, that's never been understood, I just didn't understand it.

A

Could be the next paper.

B

To solve the circuitry thing work, and these many cups are clearly a part of the answer. Yeah.

A

So we want to add some part of.

B

D

B

We need many columns to do our high order representations. This could be a way of propagating predictions across the layers you essentially anyone could any one of these layers could say: here's the next of many columns, guys and.

A

Provide a spatial context, at least for whatever's happening.

B

Yeah this prediction: this new set of cell.

D

Populations here has to go back to there.

B

For a short time,.

A

B

Through the mini column channel and the.

A

B

Column channel telling fine, you just pick those new cells and activate the under on many columns and the other one remaining columns present a prediction for everybody.

A

This really this would be amazing, so I'm just.

D

A

It when you're, when.

D

You don't know an object and.

A

You're and you're trying to learn an object. Perhaps there's a lot of mini column activity happening because you don't know it and once you you're learning one of your moving through it into a known space. Maybe they sink I.

B

Think I think the structure, the context means they're, always synced up they're.

A

Always it's not like they're scanning these.

E

Bipolar shelves they exist.

B

The physical cells- this is what they do. Okay, so I, don't think you're going to see this sorting going on so they line up bingo. It's it's more like there's two channels of communication, there's neuron the neuron communication, which is some cell population to sell Popular's yeah.

A

B

Specific, but there's this back channel and the back channel is that the level of mini columns and it sort of it allows it allows any letter cells to tell any other way ourselves. This is the it's, not the detailed representation you're going to get, but it's the it's, the the more cruder meaning common representation. You're going to get and that's a great idea. We have to stop so yeah, let's stop there, okay, but we need to think about that. One. Okay, great I, always think of something. Okay,.