Numenta Live Streams, 1 Apr 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: AI / Neuroscience Chat! Pooling in Neural Networks in silicon and in your brain - (artificial intell

Description

Spatial pooling, maxpool, CNN, HTM, temporal memory, bursting, boosting, etc. -- Watch live at https://www.twitch.tv/rhyolight_

A

Alright, let's get started thanks for watching I am Matt Taylor from the Mentos. This is I. Think the there's only the second day, I chat that we've done. I did one last week yeah. So this is only the second AI neuroscience chat that I have done and and I even have a little chat command for it. It AI chat and I just kind of described what it is today we're going to be talking about pooling.

B

A

Before I start talking about pooling, does anybody have.

B

A

Pooling is at all I there's there's a way to describe it. I guess in deep learning, I only know one way of describing it in deep learning, and let me I'll show you what I'm talking about here and on the screen. So this is the deep learning idea of pooling that I understand and that is you have an input space I like to call these input spaces.

A

They I mean these really represent neurons in if you're trying to relate deep learning to neuroscience these, these bits really represent you, the neurons or synapses, to a another neuron. You know, so you know how you know the pooling sucks. Why does pooling suck I guess I? Don't understand that I've never heard anyone say pooling, sucks, it's sort of pretty cool when you understand what it's doing, it's like it's, a it's, a cool trick that your brain is doing to take information in one space and translate it into another space.

A

Sort of so that's I'm, going to be talking about. The type of pooling that that this is showing is is is doing the same thing, but it's losing a lot of information. Yeah! That's, oh yeah, that's I! Guess that's your point right! It's losing!

A

It loses a ton of information so when you can, when you pool especially using max pool okay, so in this example each one of these we're basically taking the input space and then breaking it up into into separate spaces that will then be represented by another unit right, so you're, sort of adding another layer of units after the convolution to further decrease the features.

A

Female squelch down the feature space and, in this case I think this is probably showing an example of maximum pooling, which is you basically look in that space and you take the the max value the unit with the maximum value right and that's the one that's represented. But if you think about this, you could do this in lots of different ways. You mean you could apply any function over that space and transform a whole bunch of units into one unit right. That's pooling whether you're doing you could do max pooling min pooling mean pooling.

A

You could play lots of different mathematical functions in as a part of this pooling.

A

So it's it's an interesting concept and it's similar I think I mean this sort of of all this sort of came from biology, but in biology it works even even more differently than this I would say.

A

So we're going to talk about first, let's think, let's think about what a pool is okay, so that we call it pooling for a reason, if you're pooling, let's, let's draw a pool right, let's draw a stream of water into a pool right, so we've got a stream.

A

It's important that you think of this as a stream I think I should use a blue for this, but I don't think it really matters, but this stream is.

A

The stream implies like, if you sample this, if you take all all the water molecules right here and sample them out, you'd have a state. You know in that in that stream, in any way that what the pool is doing is it gives you an opportunity to take the data in the stream and organize it somehow.

A

You know map it in a different way, so here's the way I would draw it, and and- and you can have lots of pool you could there could be a pool up here and then you know another stream leading into that. There could be a pooling here, cooling here to be another stream, another pool and each one of these pools might mean something or represent something different, depending on how it is has been mapped to onto the stream sort of like the space that's being pulled into it.

A

You might see this space I like to call this. If you're talking about a layer of neurons, you could you could think about this in a deep learning manner as well, you could say a deep learning layer or a cortical layer. Let's just say a layer is a bunch of neurons that are all working together, a population of neurons. These neurons map I mean they have input right.

A

These neurons will have input and typically the input can be expressed because sensory input is typically topological, so the input can be expressed as some type of topological space.

A

So this is some topological space and it's you know really similar to the convolution convoluted feature. That's also a topological space right. So in this topological space, what we're doing differently is we're not going to say this group of this this one neuron, let's say, let's just say: there's one neuron here: I'm not going to say you're connected to one quarter of it and or this little portion here connected to that low portion and we're gonna. Do it differently? That's that's a start. That's a start!

A

So what we're gonna do- and this is what biology tends to do- is each one of these say: we've got each neuron, I'm gonna, just I'm for simplicity, sake, I'm, just gonna use one example of a neuron here. Each neuron has a potential pool of inputs that it might be connected to and and deep learning networks. Typically, you have dense connections between a unit and one layer and the units in layer previous to it.

A

Hiya Caleb 51 your first time here, I'm talking about neuroscience and a I, specifically pooling and I'm sort of relating the max pool idea in and convolutional neural networks and I to how it really works, sort of in in the brain and so I'm talking about a population of neurons in the cortex, getting sensory input or input, proximal feed-forward input, the important thing this is. This is going to be input. Every cell in here is gonna, get input into this from this space, and then this is feed-forward.

A

Feed-Forward typically means in the direction of the senses. So, like information flows from your senses up through your brain and when you say feed-forward, that means that's the direction it's coming from the senses.

A

So so the difference here between something like like this convolution pooling that we're doing here that there still is information being lost. Just like any time you pool you lose information, so the trick is the trick is to lose.

B

The right information or lose.

A

The information that's not necessary when.

B

You're learning something right.

A

So what neurons typically do is and neurons grow right. They could they grow from one point to another point and that's like so there's not like they're not connected to all of them. There's some potential amount of connections that these this cells or these cells dendrites are close enough to the other cells dendrites to make synapses right or our axons, and it's actually, these these axons are growing up into I. Think the dendritic Bush branches of these neurons, so we're talking about axons from this layer and dendrites from this layer I believe.

A

But the point is that you can't possibly connect to all of them. So the dinner.

B

The idea of dense connections.

A

Hey thanks for the follow okay love, 51 I appreciate it I'll be doing this stuff every week. So this is like my new job, explaining how we believe the brain works and and that when I say we, my my company Numenta pays me to do this, so we're.

B

Trying to understand.

A

How the brain works in the neocortex and and the thalamus we're trying to understand the common cortical circuit in the neocortex, which involves the thalamus and anyway I'm talking about pooling, because pooling is like a crucial concept that we have to understand: spatial pooling so I'm going to get to spatial pool. This is what I'm, basically describing a spatial pooling here and each and in spatial pooling this would be a mini column, which is a column of a small column.

A

This is not a macro column, not a cortical column, but a small column of pyramidal neurons that all share the same receptive field, the same basically potential pool of connections that they could always have. You.

B

Joined our community.

A

Did you join the HTM form? That's the best place the best place to to join right there that that's, a very active, got a very active community of people there. So we were well. We were welcome to come and ask questions. Stupid questions are fine and if you have a question here, I'm happy to take it. That's why I do these AI, chats I just want to engage the community? Give you guys a chance to any questions, live and I'll keep doing it on Mondays at 1:00 pyramidal.

A

Let me let me spell that out for you. It is like a pyramid. Pyramidal like this, so a pyramidal neuron is is what is typically pyramidal. Pyramidal cells, pyramidal cells, pyramidal neurons. These are these are typically what we think does most of the hard core computation, I guess of the brain. Here's here's sort of one one looks like so: there's, there's inhibitory, neurons and there's pyramidal, neurons and pyramidal. Neurons I believe are always on our excuse.

A

Me are never inhibitory they're, always excitatory, neurons, yeah, okay, I see you explained advanced knowledge, I, guess, I, don't know. What's I, don't know where I'm trying to I'm gonna try and meet you guys, River your ass, so I'm happy to step back whoops, step back and explain some things but I'm I'm talking about our model of pyramidal neurons, however, I mean I'm talking also about mini column, and we believe these mini column structures are most likely and forced by inhibitory.

A

Neurons that activate in the same layers as a flammable neuron, so I mean there's a lot more than just pyramidal neurons going on in your brain. If you took out all everything but pyramidal neurons, your brain wouldn't work, but we're focusing on modeling. The pyramidal cells like, like you see here with.

B

The effects of.

A

All the inhibitory cells, too so I mean we need to model the whole brain, but for the most part, these are the ones that seem to be holding the computations or like holding holding memory, while the inhibitory neurons perhaps enforce a structure so that they can that they can compute within okay, so I'm trying to explain pooling so the I, so that the tricky thing here is it's. It's super easy when I think I think it's it's a lot easier when you, when you think about the convolution, because it's it's really simple to think.

A

Oh I've got this unit and it just Maps on to this blob block of of the input in in this model and me trying to be racist. A little.

B

A

Still will say, this is an input space, okay,.

B

A

One of these mini columns that I have written drawn in blue here one of these might project- and this is just my my speculations- it projects topologically onto the space, so it's not going to project across the whole space. There's there's usually a localized inhibition happening here, so these this mini column is sort of competing with its neighbors in its in its in its network right up here about hot, which one is going to represent the activations here. So you have to also remember that in this space, there's neurons firing too.

A

So at every time step. This isn't a very good example. I mean graphic here, but every time step, let me go back to the whiteboard. Sorry, every time step we're seeing an activation pattern in in this input. Space just similar to you, would see an activation pattern over here in the in the convolved feature.

A

You'll see an activation pattern in this input space, but each of the mini columns are going to be have localized connections and we call this local inhibition, where you might have a mini column over here that that projects into this space, a mini column over here that's receptive field- is in this space, but we're breaking it up in a way that we don't have to like. If you look at the the convolve feature, we don't have to have less space we can make. We can make this structure as big as we want.

A

So the idea, one of the ideas of spatial pooling is, is to change the representation here. Transform this representation using these mini columns, receptive fields so that the mini columns will start to represent things in here, and these and like I, said this is not a dense connection. It's not, but it could it's not going to connect to every single. It will never connect to all these neurons in here. It only connects to a percentage of them. That's the way it works in biology. I mean this.

A

This mini column will never have the opportunity to connect to all these neurons because they're too far away I mean it does. It just doesn't happen like that in your brain, so all of these mini columns are essentially pooling features pooling things that they see in here and representing them when they see that that thing occurring.

A

So if you, if you look at this, how this we saw, we can sort of randomly set up these initial connections, rights from from one mini column to an input space and if we just randomly set them up so there's a potential pool and we just randomly set the weights of the connections around some Gaussian distribution so that they're almost connected then, as soon as we start streaming information remember, remember, I, said stream. That's super important stream!

A

That's another difference, I'm going to point out here as soon as we start streaming information through this, it's going to be spatial information by nature, but the fact that that one point comes after the other is super important. That's the temporal nature of the world. You have to think of this as a sensory array, not necessarily just an image that we're looking at as a part of a million other images that we're trying to classify. That's not how your brain works. That's how deep learning networks work. Yes, that's not how your brain works.

A

Your brain is expecting each image that comes or each whatever it is. Each sensory array to logically follow the next there's a reason this activation comes next, there's a reason the next activation comes next, it may be because you're moving the sensor through space, it may be because an object is moving against the sensor in space.

A

You know what I mean, so what you end up having here is: is you get a set of activations in your input space and you will get a set of mini columns that activate here, let's say: there's there's another one, and and because of these activations that occurred, we have another set of activations in this other layer in this population of cells. That now represents this data, so we have transformed the data out of this space and into another space and in this space that one of the important things is we can control.

A

We can just control how dense this representation is. How many many columns win this competition, because there's a there's sort of a game happening up here and it's all localized there's this localized game.

A

So all of the many columns that sort of are looking at this space are competing to decide which one has the most overlap with with that, the semantics within that meaning and and that's the one that's going to win and fire and represent that pattern from here on out- and we can learn that over time this- the this, if we all of these weights, because these are all weights here- all of these weights will be updated as different columns start tuning to to the spatial input that they see.

A

So these weights will learn by changing these weights. This population of neurons learns to tune to the spatial patterns that continuously stream across this input. Space I lost anybody. Does that all make sense. I'll take questions. I'm happy to take questions.

A

This is sort of like the basics of how a spatial Pooler in in neurons- and we have simulations of all this. This is like tried-and-true software tested. This is we've got this all in software. It works, it works, it works totally so, and what this gives you is a way to to represent semantics. This is sort of like how information gets encoded from your senses, into your brain and in a way like you've, got.

B

A

Encoding coming out of your senses, out of your retina out of the nerve in our optic nerve, auditory nerves, coming into your brain into your thalamus, and once it gets up into cortex, we have to translate. It has to transform to be represented in cortex, because the cortex, the neocortex, is where you're modeling objects based on the sensory input, and so this is sort of the first step of modeling those objects. It's taking the sensory input and transforming it into sort of a a dimension that the brain can control. If that makes sense,.

A

Let's see so this is spatial pooling and there's more to it than this there's boosting which I'm not gonna talk about what well yeah we'll talk about that. So one of the problems that we face by applying spatial pooling in this way- and this is and it's a problem- your brain- has found a solution for is I, wouldn't call it it's it's a little similar I guess you could say it's similar in concept to overfitting.

A

If you understand that concept, but what what quickly happens, if you set this up without something to enforce entropy in this space that certain mini columns are certain cell activations up here, just sort of lock in and they start representing too many patterns down here so you'll get it you'll get a small percentage of cells up here, representing a large percentage of the different patterns that come across this space.

A

So one way or the way, the way the brain seems to fix this problem is called.

A

Homeostatic homeostatic equal up, but basically it's homeostasis in neurons. How do you say it? It's some type, there's a freeze, that's home! You know, static, plasticity, I, think this! Is it the capacity of neurons to regulate their own excitability yeah right relative to network activity? So so this idea of spreading out the meaning? Okay, is something that we had that the brain seems to do and there's papers about this and mechanisms of doing it.

A

So our mechanism, we call boost- and it's like an implementation and software based on this spatial, pooling implementation and software of how to prevent that from happening right, prevent a few know. A few amount, a few mini columns from representing way too much data or patterns down here.

A

What sort of activation function and this okay, so we're talking about activation functions. We call this or in, if you're going to look it up, you would call it look for like k, winners or winner-take-all k, winners or something that's that's. The equivalent activation function that you would use in a deep learning network, because essentially, you might have a thousand two thousand mini columns up here and in response to this.

A

Here's, the here's, what you do I'm gonna there is a there's, a great I, say great, but that's a little there's a there's, a great HTM school about spatial pooling. If I could spell it right, so there's go.

B

A

The one Episode seven talks about this, this mapping to the input space, an episode, eight talks about learning, so an episode, seven at the end of it. Well it'll talk about I, don't even talk about boosting on that do, I, there's a whole other episode about boosting I think that's episode, nine in an episode, nine! That's what it is! Hey.

B

A

This is I, don't understand any, but thanks for the follow I appreciate it talking about brains, if you like brains, you're in the right place, I was just talking about our implementation of homeostatic plasticity in our and our HTM thing. Okay, so let.

B

Me just tell you how.

A

Boosting works, instead of so usually what happens is in response to some stimulus. Some activation pattern in the input field. You've got you know all these, thousands of mini columns and and and we all stack rank them based on their receptive field. They've got a certain amount of connections open over here right and there overlap with the bits in the space that they're covering right, so they're, trying to pattern match in the space saying and looking for a lot of activations in their receptive field, and so you might get one that's got.

A

Oh I've got I've got this column has 40 overlap. This one has 38 this one has 38 this one has 37 and will stack rank them. All of these columns will rank them ranked by overlap, and that is it's the overlap of its potential connections. It's receptive field of synapses that have already been established right connection, connected synapses to the input coming across those synapses and what will do is will will stack, rank them and then will draw a line.

A

Will pick will say: I only want K winners to represent this input and that's what we call K winners, and that is that is inhibition this this in HTML all this inhibition inhibition I'm expressing this as a local function, so there's local inhibition. So there's groups, you know each each mini column has a neighborhood of columns.

A

That is it's competing with and getting stack ranked up against right because that that's what you need if you have topological input, but a lot of times, you don't have topological input if you're, just looking at scalar data, for example, which is a big use case, with streams, temporal streams of scalar data. If you just look at a server and its metrics over time, that's all.

B

Temporal streams.

A

Of scalar data beat Alice thanks for following I appreciate it awesome I'm glad you guys are enjoying this content, so I'm talking about local inhibition, but there's also global inhibition, and just imagine that in global inhibition the potential pools of each mini column spread across the entire input space. So, basically there's one neighborhood: all of the mini columns are competing with each other. Every time that's global inhibition.

B

And that's much more efficient.

A

I mean, if you think about it, the more neighborhoods and the more the more competitions you have to compute your brains doing this in parallel constantly, but in the in the CPU there's a more localized the inhibition occurs. The more computation occurs like exponentially. More computation occurs.

A

Okay and, like I, said, if you, if you want to see that, go to that HTM school link and there's a there's, a video about boosting there. What will happen if the input is empty? Nothing will fire like these. These cells up in this layer that we're simulating will only fire in response to activations in the input space. If there's nothing there, they don't do anything.

A

There will be it's like. There's not nothing happens. There's compute stops, I mean there's nothing to do and in in the real world that doesn't exist. There's always something if there is, if you're awake, you're receiving data, even if nothing's going on around you, you still have sensation right. You can still feel like close your eyes. I can feel where my hands are like I still have sensors that are giving the input constantly. So you have to think about this as a stream over time.

A

This is not just a spatial classification task that we're talking about when we're talking about HTM Tec, hierarchical tempura memory, we're talking about a stream of temporal data flowing across this input space, and this responding and learning to it in real time online learning completely online learning. This is something that deep learning networks cannot do today.

A

So so this is the idea of spatial pooling, so we're so we're taking features essentially and we're going to start representing them differently up here and once we can represent them differently and once we've applied, oh I forgot to explain, boosting I'm, still trying to explain, boosting, ok, so so I explained the K winners and the the local inhibition the competition that occurs to decide which, which winner called many columns, are going to represent this data down here and I also said that if we let it go willy-nilly and we don't control and we don't enforce entropy at all, then a low.

A

A small percentage of these many columns will start representing a large percentage of the spatial patterns that come across so and that's why I was talking about homeostatic plasticity, because that's your brain solved. This solves this in a certain way and we're trying to do it in software similar way. So, oh so what we do- and this is a little compute heavy right, but you have to keep an a duty cycle for each one of these mini columns in here.

A

So essentially for every mini column is we have we typically 2048 will say: 2048 mini columns as a typical size of an HTM Network that will create and do and build experiments on, B tess's doesn't drop the drop block, performs spatial, pooling, drop, block, drop, block, I, don't know what drop buckets, sorry, so, I'm, not sure and and on by the way spatial pooling may be a bit of an overloaded term. You know these are both general terms. Spatial and pooling, so drop block might be spatial drop out.

A

Oh I, don't know what spatial drop out is either as far as I know this. This concept of spatial pooling is a unique discovery. That's a part of the HTM Canon, so I, don't I've never seen anyone else doing something like this. A spatial drop out that might be a technique. That's used in deep learning networks to try and spar safai the network, because deep learning folks have recognized that sparsity helps it helps against, especially against noise.

A

So there may be a tactic called spatial drop powders, but I'm not familiar with it so anyway to apply boosting you have to keep an active duty cycle for each one of these columns, and an active duty cycle is just imagine a time window say: I'm, I'm, gonna, look at this one column, this one mini column over time, so you you'll just keep a little counter and you'll say for the last hundred time. Steps. How active has this column been? That's a that's! Basically, it you're just going to check.

A

Is it active now and then add it and- and you know, tally it up and sit and give an indication. We call this ADP or ad ad see any active duty cycles. So we're gonna keep an active duty cycle for each one of these, and the boosting step will then do a computation based on all of the active duty cycles and come up with boost. We call it a boost factor, boost factors, one for each column and just multiply that columns I'm, not I'm, not exactly sure how it does this in a code.

A

But it's, but it's simpler. It's pretty much. What I'm explaining it's you.

B

A

Look across the active duty cycles because you want to prevent some of the columns from being too active, so so you'll prevent the columns that are have been very active from representing new input and you'll pass off that semantic, meaning to another column that has not been active so so you're, basically just keeping because, like I said otherwise, you'll get because there's you can represent a lot of information with these many columns. It can represent a lot of spatial input.

A

A lot of spatial patterns and you'll get the situation where you've got just a few of them representing a lot of spatial patterns, and you don't want that because you need you, you want. You want to enforce entropy in this layer so that you can spread the meaning out through all of those mini columns. Marc Brown is in the king good to see. Marc I was just talking about boosting homeostatic plasticity, so we were talking about pooling. I talked about local inhibition, I talked about many columns and how they map to input space.

A

We talked about K winners. We talked about active duty cycles, boost factors. So if, if there's, if there's any other topics, I'm going to talk about I'm gonna go on to a the idea of temporal pooling or object pooling next, but one of the key concepts, then you really have this to understand this, that's different from deep learning. If you're coming from the machine learning a deep learning world I keep hitting on this again and again this it's that this is a a stream of data. It's a stream of data! There's!

A

If you look at this there's a constant stream, that's constantly changing and your your cortical layers are observing that stream of data over time. It doesn't make sense without time.

A

Waiting yeah I haven't talked about bursting yet because that's sort of a part of the temporal memory system, we might we'll talk about bursting at some point I'm trying yeah, but the topic is pooling, I, don't mind talking about bursting, but it might get a little bit too far ahead too fast, because I haven't explained too far a memory at all bursting.

B

Sorted the key part, but.

A

It's it. The interesting interesting thing is I like explaining bursting so I'll give you a little teaser if, if a pattern comes across here so every step there are cells in these mini columns that are predicting what they're going to see next. Okay, that's that's the idea of temporal memory, which I haven't touched on yet, but there's going to be cells that are active up here like right here and, and that means not they're not trying to represent.

A

What's here, they're representing based on everything they've seen in this space, what do they think is gonna happen? Next? What pattern is coming next right? So if it's wrong, if this cell is wrong and you get the next input step and there's a totally different input that comes across here, a really interesting thing happens: every cell in the column then turns on- and you see this in the brain. This is called I, think they call I, don't know if they call it bursting on the brain or not.

A

But you see this in the brain, it's sort of like surprise OMA. You know it's like something happened. Yeah, surprise mark mark says it's, so it's so cool because you can see you know more neurons, activating in your brain when something shocks you, your brain, is kind of like pop, like what what.

B

The hell was that, right, what.

A

Was that and and it sort of it makes sense in a way because you don't understand what's going on and you activate all of these cells to tell it any number of things could happen, any a number of things could be happening right now. The possibilities are many.

B

A

And and as that pattern replays itself it locks in to one of these neurons that will decide to represent that pattern over time, as you see it over and over and then the next time you see it, that's the cell. That will predict right. That's how you learn. That's how you create memories. Well, one of the ways in which you learn it gets a key. It's a really crucial part of how you learn.

A

Honestly, it's I, don't know how you could learn without a mechanism like this in some way that takes like sensory information over time that you're constantly that's constantly flowing into your mind into your brain into your specifically through your thalamus into your neocortex and just trying to understand. What's happening all the time and if something unusual happens like a bunny pops up in front of me, all my mini columns are gonna burst because that's never happened before that. Wasn't what I was predicting. You know now.

A

I have to come up with some new explanation for this reality in front of me. That seemed impossible a moment ago. Right, that's bursting so say the mecca, the bursting mechanism. Let's talk about bursting in another AI chat, that's a good topic for another day, I shot, yeah I. Think.

B

A

I think it will have something so the tonic, you're gonna love it I got to get this podcast edited Marc, because we about tonic and burst modes with subbu, ty and now I can't remember all the details about it, but but it's important for sure, okay, I'm gonna we're going to go on to a different type of pooling. So assuming that we understand this idea of.

A

Spatial pooling, then we can assume.

B

A

Assume again, we have a a layer of cells, a layer of neurons and again there, let's assume, there's an input. There's sensory input coming up to this proximal.

A

Proximal feed-forward sensory input right.

A

So, in response to that, you'll have some mini columns that get activated in here. This is typically draw them and you know there will be cells and we haven't done temporal memory yet, but there will be cells in here that are active and those represent sort of the context of that spatial input. We'll talk about that when we talk about temporal memory, how is info? How is the information returned back? So that's a tricky question.

A

How is it returned back returned back to what you're talking about feed for you talking about feed back to, because because this sort of flows, information flows from sensory through through your brain up, you know essentially through the hierarchy and motions motion. Information emits out of this, so I haven't really talked about this, but every single cortical column in your brain outputs, a motion signal of a movement signal which is which flummoxed scientists for a long time, because why would you need that in every single part of your neocortex?

A

But it does, it seems like in every part of your keep neocortex, including you know your ears, your eyes. Your all of that has is creating some motion signals, so you're sort of constantly putting together a Moute emotion pattern for your whole body, all the time right. So there's always this constant, like output, and it's not coming from this particular layer. It's coming from a different layer that up the but let's ignore.

B

That, for a moment,.

A

I'm not talking about backpropagation I, hope, you're, not defer Oakhill. There is no back propagation as far as we know in the brain and the problem with one of the problems with back propagation is it's sort of this?

A

It's like you have to have access to so much information in a global way to to run run that algorithm and do that error correction. We don't see that there are there. Is the connectivity I think to do that? There may be it's probably done through apical credit assignment and my that's what I think, but anyway, I'm getting off-topic if you just saw the bunny for the first time you pool it, but you can draw later. If you want. Oh you mean if you've never seen a bunny before.

A

So I think you would probably so here's the thing you haven't seen a bunny before right, but you've seen rodents before you've seen furry things before. Maybe you've you've seen mammals.

B

A

I mean even if you saw a bunny you've, never seen one before you'd immediately know some things that were similar to it right. That's crucial! That's because that bunny, that you saw with your sensory input, has semantic overlap with other objects that you've already learned in your and your neocortex. So even without learning it before you can already relate to it. You can already relate to it and you can compare it to things that you already have representations for.

A

If you bring in your brain, so yes, there would be bursting there'd, be bursting because it's out of context, you've never seen a rodent appear in front of you on your desk or whatever before, and also that one looks a little odd and you've never seen a rodent. That looks like that with long years and a little bitty button tail right, but that's that's something you would have to learn and you'd have to inspect it.

A

You know you'd have to look at it and I don't and if you just had like a flash of it and you weren't able to inspect it like explore it and because all of our intelligence, I believe is based on the exploration of space and using your sensory input and coordination. With your movements to explore space, you have to know how you move through space and how the objects respond to your movements in space anyway. I think I'm getting off-topic, but that's.

B

Okay, it's AI chat.

A

So we can get off-topic if you want to, but I want to talk about another type of pooling. So this is a little bit of a trickier type of pooling, because I'm jumping ahead I'm going to talk about this idea of temporal pooling and we haven't really talked about temporal memory, yet so I think the easier way to think about this is to talk about object, pooling and- and this is the one of the key ideas here is- is this circuitry that we have in our in our cortical column?

A

So not only do you have this layer, you have another layer, you have another layer. There are something like 10 12, different layers, depending on who you ask that I think there's there's six like easily differentiable layers in your neocortex, but then there's multiple different cell types and many of the layers, and so each one of those could be said to be doing something different, and maybe you can treat that itself, each different cell type in a layer as a different layer itself, since they have different compute I mean there's even gonna, be layers down.

B

A

And as proximal into it feeds into this this thing that's doing spatial pooling at the same time, it's doing another process that we that we call temporal memory.

A

I'm, not gonna, explain that right now, but essentially it it activates different cells in each mini column, depending on the context over time that it's receiving the input. So so, essentially, it's storing sequences of spatial patterns. So you can using this mechanism that I'll explain in another a ichat, probably you can identify or you can make predictions. That's the key thing you can make predictions about what you think is coming next and the predictions will be in the context of all the other sequences that you've ever Sene.

A

So but the hard thing here is that you don't know what the sequences are. That's the tricky thing you you, you have the predictions, but in order to dive into this list layer and identify all the different sequences that this particular activation might right now be played be playing out based on your experience is particularly hard. Let's say mr. chat here is explainable: AI possible HTM s also. How would it go about explaining decisions being made? Okay, so I'll. Take that question explainable, so definitely I think it's possible, definitely and in fact, I.

A

Don't think that there that's like our mission. Our mission is to understand how intelligence works in in the new cortex in the brain, essentially so how much more explainable can get that's our mission is to explain a I explained intelligence so that we can make intelligence, that's crucial. So absolutely it's explainable. That being said, there's this idea of reality.

B

A

We haven't talked about this I. Don't think I'll of bring this up because it really hit me when I read it so I read a book by Max tegmark, actually, Jeff Hawkins suggested it so I read it and it was called this met our mathematical universe, something like that, and he has this idea of three different realities, and this is totally hits home for me this when you're talking about AI. This is this is a great way to think about it. You've got three different realities. First of all, there's like true reality.

A

There's the reality external reality. He calls it there's the reality. That's out there and it's everybody observes it. Everybody has a window to that reality, but but nobody has a complete view of that reality right now, then there is internal reality. So I have my own internal realities represented in my brain. You have your internal reality. It's represented in your brain. There is no way for me to completely share my internal reality with you. The way we do. This is through a third type of reality that max tegmark calls consensus reality.

A

That's the reality that we're both existing in right. Now, it's the reality in which I'm speaking words to you and you're understanding. Those words are trying to understand those words and making an internal representation of them as you're. Trying to understand these concepts so I have this internal reality that I'm trying to project through the consensus reality through education and through this twitch channel, for example, so that all of you can update your internal realities to have a better understanding of my internal reality right. Does that make sense so think about AI?

A

Now, because there's no.

B

A

My internal reality could be interpreted by you there's no way and unless you took my head off and put it on your neck, and in that case who are you anymore right, there's no way I could freeze all the neurons in my brain I could map them all out. I could put them all in a computer there's no way you could make sense of that.

A

There's no way anything could make sense of that except me period, and you know why, because all those representations are based on my body they're, based on all of my experiences I've ever had as I did this interview with neuroscientist I've done too many of these Michael I, followed on Twitter I'm, sorry he's a motor motor guy. He does motor inference stuff and I asked him. What is it? Did somebody drag my memory?

A

If you can it's a it's on my YouTube channel, it's on the YouTube channel under an interview with a neuroscientist and I, asked him what happens in your brain when you go reach for a cup right. What happens in your brain? We reach for a cup water's activating and he's like. Well, it's like the some experience of every time you've ever done ever reached for a cup or ever done.

A

Anything with your hand comes into play when you do that, like there is no separating the way I reach for a cup from me from my identity that internal reality cannot be separated. From my experience right I, you will never! You can never understand. If my interpretation of read is also your intake interpretation of read, that's that's the idea, so with AI you're going to have the same problem.

A

If we're reading intelligence, that's based off of how our brain works, even if we can explain all the minutiae of the cells and the layers and and what's happening here, what's happening here, you're, never going to be able to take an internal representation inside of a common critical circuit. That's down here! That's related specifically to the sensory input of this agent and it's experience with reality. You will never be able to compare that with anything else and make sense of it. Does that make sense now?

A

That being said, we can create some really elaborate consensus realities so that we can communicate with these AIS once we get to that point, but we're never. You can never expect to look into these cell activations that I'm drawing up here and expect to know exactly what experience that agent that entity that intelligent thing is experiencing. Jonathan Michaels. Thank you Falco. That was him Jonathan Michaels.

A

If you're interested I thought this, was it I really liked this? Oh, let me look it up real quick, Jonathan Michaels he's on Twitter to interview neuroscience there we go! Oh you.

B

A

Thank you so much. Okay,.

B

A

So, if you're interested in that idea of motion and grasping okay, I'm gonna try and catch up on chat here for a minute, because I've been just talking and talking so just many columns sequences in one map of mini-com sequences in another map, yeah is explainable, AI possible! That's for you! You might let night, like the next wall of text sent to predictor. Okay, that's on the HTM forum. Mark's talking about.

A

Explain it may be explainable in its principle: stockless s, but it unexplainable due to complexity. It's not complexity. It's unexplainable by nature! That's the key that you got to remember because it's an intelligent agents. Experience is unique to its movement through space. So, as you could you could take, you could train. Let's say: let's assume you had an AI robot and it was smart. However, we made it smart right and we we turned it on and we let it learn how to move around in a little room. Okay, so it's learn the room.

A

Now you could clone that thing and send it off into a bunch of different rooms and and the rooms would be different right and then turn them back on and say, they're all learning again and they all keep moving through those rooms. Each one of them has now diverged from its original internal representation. You cannot compare them to each other anymore. There may be some ways that you'll compare them for a small amount of time. While the overlaps are similar.

A

You know things they've learned, but as they learn new things, they're going to diverge and you're not going to be able to reliably take internal representation and and turn it into anything meaningful over time.

A

Okay, reality that is collective yeah, so that's the consensus reality where have I heard everything you learn is in terms of what you have learned before. I don't know, but that's absolutely true. Every everything you bring your full experience of the world to every new situation. There's no getting around it. You can't really like make it down to just here's the input it's like everything you've ever learned is a part of your experience.

A

That test says this reminds me of the Sibyl system and the anime psychopaths I I'm gonna. Leave you guys up to interview to investigate that. Have you tried to build reinforcement, learning, HTML, common reinforcement environment? So that is definitely a huge deal. So that's that's the future. Absolutely reinforcement, learning with HTM is most certainly the future, especially on the in the direction towards robotics and gaming applications for HTM.

A

That being said, and I've talked about this a little bit before in other chats to the sec, the the mechanism for enabling that is not quite.

B

Nailed down so.

A

If you want to look up the oops, it's the it's the it's the framework paper this, this sort of defines sort of one of the this one right here defines sort of the the idea of displacement cells, okay, displacement cells- if you can no.

B

A

Always get too deep because these tats get deep fast, but you have to understand how grid cells work. You can represent different locations in space yeah. What he's saying you have to have the the the.

B

A

Has to be learning, the HTM is a whole different animal, it's true, but they both have to learn together. They both have to learn together, I think they have something to tell it they have to have feedback they have to. There has to be back and forth. There has to be some type of integration. The key thing, I think, is that the the HTM system plays a part in creating movement right like I, had.

B

A little arrow over here, the.

A

Htm system has to play a part in the causation of movement. In that way, it has to understand how its movements affect the sensory input, that it gets back and I think the way that it has to do. That is by having some representation of displacements between locations somehow in one of these layers, so that it knows the repercussions of its movement because it knows where its sensor is it.

A

If it's a part of a movement, it has to predict where it's going to be next and then what its gonna feel when it gets there, and- and that is a part of the feedback that comes back a sensory input and we're not quite sure how these displacements layers play with the other layers that we've identified so, but once we figure that out, I think the link to reinforcement learning is like right there.

A

That's that's the point where you take where you take the action in the environment, and you have some type of goal: reward set up there and the in the lizard brain is Mark Brown likes to call it right. I hope your green with me mark here so that we can attach our our neocortical model of the world to that control center, that control system.

A

Somehow so that's sort of the idea and the reinforcement learning system would be more of a model of the older part of the brain, that's dealing with with reward and Punishment and stuff, but at the same time, it's constantly keeping a link to the models that it's attending to and exploring, etc and and and make no doubt about it. I feel like we're talking about thalamus in the neocortex here, I think, there's that's the huge link and that's that's the direction. We're headed, that's, definitely the direction we're headed.

A

We haven't gotten to a point where we're ready to to create any type of reinforcement, learning system yet, but we've got some toes in the water for sure also the way that the amygdala assaults, the memories that are forming yeah, that's a good, that's a good way to put it seasoning the memories.

A

The emotions are a part of this for sure and I think emotions gets get represented all across your neocortex I think they're, a part of every experience that you have your emotional state when you've learned something can affect the meaning of that absolutely and I think that Coke sort of that spreads through your whole brain. You know yeah, it's definitely so it's a way that the the lizard brain projects to the cortex, your fear response, your fight-or-flight response.

A

If you see an object, that invokes fear. That's that's. That's evidence right there that the emotion is stored right there with the object, representation right that makes total sense to me.

A

I haven't even gotten to object, pooling yet I'm thinking. Maybe since we've already reached like an hour I'm wondering if we should talk about temporal memory instead of object, pooling because it's I'm thinking, I'm gonna have a hard time explaining object, pooling without explaining temporal memory.

A

B

You guys still into this I could keep I could keep.

A

Doing this, but I'm gonna have to take a break. So why don't you give me five minutes and I will come back and we will have a little break. I mean while talking about we'll talk about some floral memory. I haven't prepared anything on this, so it's just gonna be sort of off the cuff, but so I'm tired you're too kind. Thank you so much. That's awesome! I'm glad you guys are enjoying this. I am I. Think this twitch thing is just amazing. I'm I'm very excited about it.

A

It feels like I'm, the only person consistently creating any like AI neuroscience, related content right now on Twitch and that's a great place to be because I feel, like this platform, is really blossoming right. Now, there's a lot of things going on in the Science and Technology game in twitch. If you know what I mean all right, guys, I'll put a little music on for you and I'm gonna go to be right back screen.

A

So if I have more.

B

Viewers now than I did before I.

A

A

A

Feel like I want twitch, so I can be just like whatever I can just do whatever I want so I'm gonna wear shades better we're gonna talk about temporal memory, I'll leave it on real low. If it bothers anybody. Let me know: I'm gonna draw something different than this to explain temporal memory. You guys.

B

All with means to Bristol er brain.

A

Hurts the music's out of key I didn't notice all.

B

My wife, all my.

A

Whiteboard things are gone.

A

Reset reset: okay, alright, that's! Okay! That's weird! Alright! You know it doesn't matter right. So let's talk about this input space again now, let's talk about a mini column. Here's here's.

B

A

Alright, this is in some layer of of cortex, so there's lots and lots of mini columns. So we're just going to talk about this one mini column and I think we would typically say we see this behavior in layer, 2, 3 and I always forget. I'll talk about cortical layers, another time and I'll have done my homework by then. So, let's just say, this isn't happening. Every single layer of cells in the in the neocortex but I think it's crucial at least two of them.

A

So each one of these mini columns represents a bunch of cells, pyramidal cells. So let's say a a mini column.

A

Has many I'm just going to call them cells and when I say cells, you can just think of that as a pyramidal neuron.

A

A

So, like I said before, I like to draw the input space in green.

A

We have some input and it's an.

B

A

Just a big grid of cells as well, which is that there it's also we can sort of see these as synapses or puts it well. The tip cells that are in either a lower layer.

B

Another layer, this.

A

Could be another layer could be another.

B

A

Or it could be sensory input like direct from the thalamus or that's relayed through the thalamus or perhaps.

A

There's even olfactory goes straight into your cortex, so maybe direct, maybe direct sensory, maybe through the thalamus. The point is this layer doesn't care, it doesn't know it doesn't care, it doesn't know it's. It doesn't know, that's one of the things that makes it common a common cortical circuit, we're looking something we're looking for something. That's can be generalized that generalizes everything right. That's why we're so smart, because we can generalize so much stuff. So we've got this.

A

This mini column forces all of the pyramidal cells within it to have the same receptive field to the infant space.

A

Okay, so assuming that it has some receptive field to the input space and it activates in response to cells, firing within its potential pool or source up the field, it will either activate or not and and decides whether it's going to be active based on how many of these overlap it's connection its connection space, that's where we were at before I started and and by the way is anybody in here that doesn't know what they're looking at, because I do have some feel free to use these chat commands if you want to learn the facts or links or whatever to other things, I added Zee.

A

B

Doesn't do much yet, but.

A

One of these days it will- and just in case anybody cares to know I've got three types of shows, a iChat, which is what we're doing right now. Every Monday at 1:00 I've got work sessions that are technical engineering sessions, Tuesdays and Thursdays at 9:00, and then fun Friday, which is whenever fun for me, whenever on Friday I'll, do something fun but check my twitch schedule or give me a follow.

A

If you want to get notified when I'm doing any of those things, it's all gonna be related to AI in some way, because that's what I do I work for the Mensa as an open-source community manager and I try and make educational materials explaining this type of stuff to people who might be writing it in code. So you want to write some of this stuff in code I'm here to help anyway, yeah someday I want to integrate HTM into the tooling either for the chat.

A

Server I've got a server rhyolites daemon the chat, bot, rhyolites team and that's not just a chat bot. That's a server so I'm gonna hook him up to time, series database and we're going to start sucking up information from twitch.

B

A

I already have an API to apply HTM anomaly detection overtop of it it's old, but I'm, not sure if it works, but I don't know what I'm gonna do yet mark I'm, still thinking it through I. Honestly, don't have time to do anything yet and I'm worried about new pick. I'm worried about Python 3 I need to get that sorted out before I. Do any really fun stuff. So you're gonna see me probably doing some boring technical Python work, trying to figure out how we're going to get a decent.

A

How we're gonna get a Python 3 environment for HTM, because our because our time is running out the clock is ticking. I should put a countdown up here, so it reminds me every day how many days I've got left before Python 2 is end-of-life anyway, if I had a yak shaving button, if I had a yak shaving button I'd push it right now, but I, don't so trying to explain tempura memory.

A

This is not an April Fool's joke, hey.

B

A

I recognize you, son son, 1, 1, 2, 3 I've been streaming on Twitch for like a month now so yeah like I, said mark I have no idea what I'm gonna do, but definitely a time index. Everything is gonna have to have time. I mean ideally yeah, we'll.

B

Talk about that later, we'll talk about that later. There's.

A

Lots of things we could do I think code thing to do keywords: code thing to do keywords; I, don't know what that means. Sorry, okay, so a mini column activates in response to some some activation in this input. Space thing to do keywords like my channel button, I yeah, I've,.

B

A

Of things to do like that, I've got a list of things to do to make the twitch experience better. It's just getting I came in over the weekend to do some of it. It's just busy right now, so I don't want to put too much of my time and effort into twitch, because I have a job and honestly, this was wasn't it before last month, so I still have to keep up with all my other responsibilities, and then this Python 3 thing is anyway.

A

I did not I'm not streaming to complain to you type, something in, and it does a time mark thing. Yeah.

B

Oh I I did do that time.

A

Mark right now, I just marked it Oh or you can clip it, which is in the web. Ui I, don't.

B

Know if this is.

A

What you're talking about, but with the keyword, yeah, there's a there's a command to to clip I think it's back slash clip, maybe no I, don't know, but.

B

Here's the clip button right here.

A

All right there, if you've ever wanted to clip something, but anyway I did figure out how to set stream markers. So I did figure that out.

B

Anyway, I'm getting off.

A

The topic, let me let me get back to what I want to talk about on the whiteboard, which was simple memory; okay, so this mini column activates in response to an activation in its input space and it, along with a whole bunch of.

B

Other mini columns.

A

You've got to remember: there are a bunch of mini columns there. This thing is full of them, full of them full of them full right, we're just looking at one and each one of these has got all these cells. You know.

B

Just like it they're all over.

A

Here, oh right tons.

B

A

Now, in response to like, like I, said in the previous, we were talking about this earlier when we're talking about spatial pooling. If we get this input and there's been nothing, nothing is predicted like we didn't. We didn't predict this input, we're gonna, we're gonna burst the whole column, okay, the, and that means that this is new.

B

This is one of the ways that this.

A

Is how I like to explain it there's two ways these cells fire one is one. Is that the the call the mini column bursts.

A

And the other is correct: prediction: okay, two ways that one of these cells fires, so let's say I hate that it got my I got my.

B

A

My orange you got here peanut butter in my anyway guys stop looking at my bald spot. I know it's sexy, but my eyes are right here. Okay, so let's say nothing is predicted, so we burst every column, I'm gonna, I'm gonna, get rid of this cuz I. Don't want to make my pin black, so we've got active cells, active cells, all the way down all the way down active cells.

A

Okay, so that's one reason: one of these could fire. The other reason is there's another state that HTM neurons can be in and we'll call that predictive I'm gonna make it red I. Guess we'll make it red hey in a beta okay. So so, let's say this style cell is in a predictive State. What this means is that it has recently got enough distal input on one of its distal dendrites that an MDMA spike has just occurred, which means which is a dendritic spike, so you can Google dendritic spike and there's a there's.

A

A video I made about what added riddick spike is. If you want to see that in fact I'll do it I'll do it for you cuz it's decent enough and I don't want to go over. Let me just show you then didn't. Do it tick, spike, Numenta, I think that's.

B

A

Here it is, what is a dendritic spike, well link.

B

A

Bits but I don't want to sorry I'm gonna pop I'm gonna pop okay. That's me, my point being is usually you the the soma spikes when it gets proximal input. That's what I'm trying to show there if the soma gets proximal input. These are proximal dendrites near the cell body, then the soma fires. It creates an activation, but there's all these other dendrites that are around the edges that are distal distal, far away right.

A

If those get stimulation, it doesn't cause enough stimulation anywhere for anything to happen it and for a long time we didn't know what they were even therefore, but then we found out that this thing called an indium, a spike, which means, if there's enough activity in one sort of localized area, of a set of an indirect segment, that it would send a signal back to the soma right and this, and it wouldn't be enough to cause the soma to fire.

A

It wouldn't breach the activation threshold, but it would get it so close to the activation threshold that it only needed a little bit of stimulus to fire. So that's sort of a different state of the cell and we call it a predictive state if it gets enough proxy excuse me distal dendritic activity, it can put it into a predictive state, there's not much more to this. So and that's what I'm talking about here.

A

If we have a cell, that's in a predictive state and I'll talk about where those that distal input comes from in a minute, but right now, I just want to tell you how it fires. So if it is in a predictive state inside this mini column, that means that this cell is predicting that this mini column is going to fire or it's going to be activated in response to some signal coming through the input. So if that happens, and the activation of the mini column does occur, and there is.

B

A cell in a predictive state that.

A

Means it's going to fire fast, it's going to fire fast, because it's sitting there waiting for proximal input. It's it just had some indication that it might be soon. It got real close to firing because of that in DMA spike. So if that occurs, then that means that that prediction was correct and it learns so the synapses are. The permanence is for its synapses are then reinforced that it's thought it was going to be predictive of the in DMA spike, put it into a predictive state.

A

It actually was correct because it was happened to be inside the mini column, that the input space activated when spatial cooling occurred, and it was correct, so it gets reinforced. The the distal connections that caused it to go into a prediction predictive state can be reinforced to reinforce that pattern because it was something's coming I, don't know. Okay, that's ominous. That means, but oh you're, talking to the audience so.

A

Where should I go? Where should I go next? The the next now I have to talk about. How does that cell get into a predictive state? That's the next logical thing, so that's sort of well I just talked about how a cell gets into it. It's looked at the distal connections, so all of these each one of these cells in the mini column, has distal connections. This one might be connected to that. One might be connected to this. One can be connected to any cell in any one of these other columns.

A

Okay, this is like this is the sequence memory this. This is where you this is going to blow your mind. I think this is what what bit King was saying is coming this. Is it once you get this? This will blow your mind.

A

This cell has all of these different potential distal connections that might put it into a a predictive state that represent one sequence. Now this cell could also have these for many different sequences, so it could go into a predictive state because of some some pattern that it sees evolving or a different pattern that it sees evolving. That's completely different. This cell can learn lots of different parts of patterns. That's just the way that memory sort of takes advantage of these sparse connections in your brain. Okay, let me take some questions.

A

Are the distal SP connections in each cell, connected to cells and other columns, not in other macro columns, not in other cortical columns? At this level there are distal connections and other layers between cortical columns? Yes, but not not in this mechanism that I'm teaching right now. This is all happening within one layer, one layer of a cortical column. We don't you, don't have to go outside that layer at all. You can. You can apply spatial pooling and temporal memory. It's like peanut butter and jelly.

A

They just you got to have one that, for the other. Temporal memory works right on top of spatial pooling and then we've got other mechanisms that can also work on top of temporal pooling. In the same way or spatial pooling in the same way, okay, but we'll get into those at some point- says the distal cell, oh I, guess so yeah. So this neuron inhibits others in a mini column. The mini column inhibits other mini columns from expressing themselves this.

A

This neuron I'm just talking about what's sort of happening inside the mini columns, so the the inhibition I about in the last hour. That was when we were talking about spatial, cooling and abyssion happens when the mini column, competition occurs. So we're sort of saying in the context of that competition. What are the cells within those many columns, doing and they're learning sequence memory, they're learning the spatial patterns that are flowing through this space and connecting them over time? Okay, that's!

A

What these connections are this this guy is going into a predictive state, because this cell was just on that cell was just on that cell was just on, and that represents something that it has seen before its last time. It saw that, let's say it fired next, that's why it's in a predictive state and then, if it, if it turns out it's wrong, then we decrement those synapses. So we don't. We have less of a chance of it being wrong in that way.

A

The next time, if it's right, we increment the synapses, so we nail in that pattern. Okay, is there septa field connected to the cells or the column, the mini column, I'm? Talking about the so there's two different receptive fields here in this sort of idea, we've got the mini columns receptive field over the input space. That's what this and this is. This is the other receptive field, every cell inside a mini column, every pyramidal neuron also has a receptive field distal receptive field. This is the proximal receptive field.

A

Okay, that's the mini column is enforcing, and this is a distal receptive field of the same cell. Does that make sense and what time frames is happening. Okay, forget about time frames this. There is no tithers there's there's a step. There is step after step. All you have to think about in the brain. I had this this.

B

Is a hard question this.

A

Is hard to understand, okay, but think about how time slows down when a traumatic event occurs or when your adrenaline kicks in how it feels like time slows down and other times where time goes very fast. There's that idea of time being relative that I think is very applicable here, because all your brain knows is input after input input, input, input, input, that's it this space, just changes. How fast is changing is as sort of irrelevant, that's a context that might be expressed in the space, but as far as this layer is concerned, it's irrelevant.

A

So it's it's a little hard concept to grasp, but, let's see a very short term, yeah very short term definite I mean net input after input really really short.

B

A

Let's see the peak of each alpha brainwaves sure sure very fast, continuous, absolutely continuous, yeah, more fair question be the difference between n steps and the depth of them. I'd.

A

Not sure what you mean by depth, every step would be like a full sensory input. You know, different regions of the brain are going to be processing portions of it and they share their their interpretations of space through.

A

At least through lateral connections across cortical columns and layer, two that's a little further down the road. Okay is inhibition, not only limited in time. Inhibition tau happens at every time, step every time step. There's this mini column, competition. Every input that lights up this board is a new competition that chooses the active columns in this layer right and then there's yet another computation. That is a temporal memory, computation that occurs within the space that the spatial pool.

B

A

Just just created it's the.

B

Spatial Pooler, the.

A

Spatial pooling operation here is translating the semantics of this meaning into a new language of mini column, activations and within that language of many columns. The temporal memory plays out this temporal memory only works within the context of this space. Okay, by getting everything right here, mark can Brecht me from wrong or in a beta, if you're still, there I think he left anyway. He is he's good. He talks about neuroscience stuff, too. He's got some good presentations.

A

B

A

World is not quantized quanta sized quanta sized are quantized. It does that with brain waves to unify into time steps brain waves, yeah.

A

The in well, you think of it: I don't like the neuroscience term, is inhibition, because this activation is essentially enforced by inhibit inhibitory cells, but I never thought of it. I, never understood inhibition until I understood the neuroscience. What it really is is a competition for, for which many columns are active. When I did the stack rank thing, that's in abyssion, that's it it's and it's inhibiting by saying I'm, drawing a line here and I'm, not letting any of the other.

A

The other mini columns express themselves, that's the inhibition part of it, but it's your if you're confused about inhibition you're, probably over complicating it. In your mind, it's really just the activation of mini columns based on stack, ranking there overlaps with each of their inputs and then choosing only the ones that most strongly overlap that input to represent data there to represent that input, that's the inhibition part and then the boosting sort of happens. In addition to that, the.

A

That well I, didn't I, didn't read a neuroscience paper: there's, not necessarily a in neuroscience paper. That explains this you'd have to read: read the neuron paper. That's that's the paper to read. Let me go back to the art papers section. Why neurons have thousands of synapses? If that that's definitely the paper you want to understand. If you want to understand this idea of inhibition and even in the in the Bamie, you did yeah I read it again. Honestly, there is no shame in reading these. Oh I had to read these several times.

A

All of them I feel like I'm, a pretty smart guy, but I had to read these papers several times, so don't feel bad about it. It's like I think it's normal I! Think that's what you do when you, when you do science when you're faced with hard challenging problems. Sometimes you have to read things several times before they make since and sometimes I'll reach something completely discounted, because I think, oh, that's not applicable at all and then a year later realize how insanely applicable it was anyway.

A

So alright I was trying to explain temporal memory, I hope, I'm still sort of on track, but I.

A

Might I probably need to wrap up soon, so I can get some things done before. I have to go pick up, my kids so.

A

So the tricky bit so I talked about bursting and we talked about the about the distal connections and why this one is active. The the the tricky thing that I think that took me a while to understand with this also has to do with it being a high order memory.

A

The inhibition timescale is a half-life effect in the basket and chandelier cell activations yeah I can see that I. Don't think we modelled that, but I can see that happening, I'm, not on that and I'm, not sure how I mean that's one of those things. That's one of those questions. How how strongly do we model these things? We could add like a whole procedure for for adding that sort of halflife inhibition effect over each time step instead of applying and I think every time step and not letting it move onward.

A

That might increase the accuracy of things that might I don't know. I, don't think that I know of we've ever tried that there's a lot of things like that that could be tried within the realms of HTM firry that I'm talking about here. There's a lot of there's a lot of knobs that could still be installed and twit and and maybe they'll have big effects, and maybe not. This is like frontiers of science type stuff here this is a it.

A

This is so mark says it makes a big difference and, like I said, I, don't think that we're we're modeling that at all, so anyway, it just shows you where we're there's a lot of greenfield here. There's a lot of greenfield here for for research, for, for whatever.

A

So, okay, so one last thing I want to try to explain is the tricky bit is: is high order memory versus single order memory. So if you understand like Markoff a Markov chain, this idea, where I don't think I'll explain it, but basically you make your decision about what state you're going to be in as a unit. It depended on the last the last time step so since the state of a system at the last time step. So you essentially are only looking at the last time step.

A

So that's the only thing you have to make your decision. What this is, what this is is something that's moving even more for than that. It's a higher order. Memory system it can, it can remember not only was this input a part of another input, but it also.

A

There was one previous that I learned that it was a part of too, and it can that can be expressed in this layer of cells and I, really think the best place to understand that and I'm gonna go back to this again is there's just there's this one HTM school video I worked a long time on these on these visualizations. For this one and I think it's still probably the best place to see how it works and I'm just going to mute it and go.

A

There's a point in here where you see you see the bursting happen and, and you can see how the sequence just split so I'm, just I'm explaining this note these notes and so I dialed in these notes. In this little thing and and then had the sequence, memory learn the system. No, no I want to do that. Don't do that! That's the press, the wrong button and I. Had the sequence.

A

Memory learn this pattern of notes and after a while it learned it it stopped bursting because that's what it does at bursts, embarrassing because everything's new and then eventually it stops bursting and it's nails it down. Every predicted cell is properly predicted, everything's reinforced. All the weights are stable and everything's locked in and then, if you go change, one of the notes and watch what happens? That's that's the interesting thing and so I'm not I'm. Gonna leave this for you guys to do. If you want to do it.

A

I'll share it at a brown I, don't know I'm, just gonna I'm, just gonna link it. Oh geez, I hate. The time continue. I, don't.

B

Want to do I just want.

A

The I'm just gonna leave this for you guys to do because it this is sort of fascinating I thought it was fascinating, so you're gonna. So when that when I change, the note you'll see something happen, that's that really nailed it in for me. So so take a look at that. If you want to understand now, I am going to log off I got all you guys hooked in with me right now, but I've got some stuff to do. I really appreciate everybody hanging out and before I.

B

A

Does anybody have any questions that they want me to ask or any suggestions about how I could make this twitch experience for you guys better or more interactive? If, if you come to my work sessions, Tuesday and Thursday I get real interactive with people. Well, I've been working on code with some people looking at the same Trello boards and everything so yeah thanks everybody for watching I, appreciate you jumping in and following me, etc. I'm not doing this to make money or anything on Twitch. This is really just to spread.

A

The information you know, I want I, want this technology to be free and open, transparent, free free is and speech transparent for anybody to understand it, no matter what company they work for or what, where they are, and the technological class rank whatever.

A

Okay, all right, let's do show me some backprop doing unsupervised one-shot memory thanks thanks mark! That's that's true. All.

B

Right, let's go.

A

Find somebody to raid that's what we do at the end. It's always good to give away your sentence. Send off your viewers to someone else. Oh.

B

A

Gonna, send it to this guy because he was recently talking about him. So let's raid this guy, he he is apparently I think doing some a lot of algorithm stuff. So that's exciting: let's do some! Let's, if you want to learn about algorithms. I was just explaining a couple of algorithms to you just now: the spatial Pooler, a temporal memory, algorithms, etc. So we will go read him when you guys are ready, got 12 of us thanks thanks for sticking through 12 of us are ready to raid I got 17 people watching right now.

A

That's excellent I really appreciate it all right we're ready to go. Let's do this take care! Everybody have a wonderful day.