The Thousand Brains Theory of Intelligence | Jeff Hawkins | Numenta

the title of my talk is the Thousand brains theory of intelligence at my company Nea we study U two things one is we study how the brain works specifically the neuro

cortex and we want to understand that from a biological point of view but we really also want to take what we've learned about the brain and apply it to Ai and machine intelligence so that's

what our mission is that's what we are doing and uh today I'm going to tell you about the progress we' made and understanding how the brain works and how that will affect the future AI but mostly I'll be talking about uh the

brain and what it is the brain does and how it creates intelligence so let's just jump into it um I'm going to start with this uh this

article that was written back in 1979 by the famous scientist Francis Crick called thinking about the brain and he made the observation even back then that neuroscientists or

scientists had a huge amount Collective too much amount of data about the brain we knew all kinds of facts but we didn't really have any understanding about how it worked and his observation was that

perhaps we don't need more facts we just need a new way of thinking about it so that was in the subtitle of this article that we needed a new way of thinking about the information we already had somehow the brain was doing something we

weren't just didn't understand it correctly later AR in the article he wrote what is conspicuously lacking is a framework of ideas within which to inter interpret all the different approaches

meaning all the different uh pieces of knowledge we had and that's what I'm going to present today um I'm going to present a way of thinking about the brain and explaining how what it does

and how it does it that is consistent with um All the known facts about the brain so let's just start with the a

picture of the human brain um and that's the neocortex it's the big sheet of cells the wrinkly sheet of cells that's on wraps around the top of your brain it's about the size of a large dinner

napkin maybe 1500 cmers in in human it's about 2 and a half millimeters thick now this include the organ of intelligence in humans um animals have various levels

of intelligence and not all of them have a new cortex but in humans and animals this is the organ of intelligence um there's a couple things we can say about it um it's the primary

organ for responsible for all of our sensory perception so when we see something uh and we recognize what it is or we touch something or we hear

something um it's the neocortex that's understanding that impul it also is the primary um creator of our motor Behavior so when we move our limbs or we

manipulate tools uh such as picking up a a physical tool or maybe using your smartphone or when we create language it's the neocortex that does this my neocortex is creating my speech right

now and yours is understanding it and it's also the organ of all thought or abstract thought so we think about math or science or philosophy or

Neuroscience it's the near cortex that's doing that now it has some remarkable attributes um first of all it learns continuously it never stops learning uh

throughout our lives we're always learning new things we walk into a new room we quickly learn the the format of the room or if I get a new a new tool I'll learn how to use that or I pick up

a new object or if I get a new app on my smartphone I can learn it very rapidly and continuously I never stop learning um we just always picking up new pieces

of our environment all the time um it learns rapidly unlike most artificial noal networks today you you generally can be exposed to something once and

you'll remember it um and so it's not like we have this very long uh training period we just quickly pick up things we say oh this is there this is that this person said this and so on and we

incorporate this knowledge into our into our model of the world it's very efficient the the neocortex or the brain is as a whole consumes about 20 watts

which is very low power and this is despite the fact that it's made up of very slow elements neurons themselves can't process very quickly um the fastest they can do anything is about

five milliseconds so we have the system that's made of a very slow elements uses very low power yet is extremely powerful um and so that's a wonderful mystery and

perhaps most important thing is its flexibility its extreme flexibility each of us learns uh to do thousands of different types of tasks in our lives

whether we program or use computers or we learn how to drive a car even simple things like how to open Windows and clean them or how to plant um vegetables and flowers in your garden we can learn

almost anything and um we all learn we just tremendously flexible we're not limited in what we can do and we see the parallels between different activities so we generalize well between different

activities now it it should be mentioned that today's AI um is not nearly as capable as a human brain uh it doesn't do most of the things that we're able to

do and and it's not even close and it of course it these the systems we have today are they don't learn continuously they don't learn quickly they take huge

amounts of power um and and they don't generalize they're not flexible so the brain is is doing something complet very very different than today's

AI so here's the outline of my talk first I'm going to talk about what the neuroc critics does and the way to think about it and that is it learns a model of the world um then I'm G to talk about how

that model is constructed first that it's it's a highly distributed model and it learns through movement and then the third Point here is how it learns um a

model of the world using something called reference frames and I'll explain this and then how do we know these reference names exist because they're derivatives from other cells called goodd cells and

place cells so hopefully uh this will be um easy to follow it's a little bit of Neuroscience but that's very interesting so let's just talk about what the netic

does it's tempting to think of the neocortex as like a computer like it gets some input and processes it and gets some output or if you're using one of the new chat AI systems you type in a query and it gives you an answer but

that's not the right way to think about it the brain what its main purpose is is to learn a model of the world to create an internal representation of the world that's

outside and that's what we do when we learn subsequently we use that model for various things we can use that model to recognize things know where we are we can use that model for planning and so

on but the goal of the system is to learn an a representation of the world an internal model of the world that's its first and foremost goal then subsequently we use that model to do

now the number of things you know about the world is huge it's very large we know thousands of physical objects how they look and feel and sound uh we know that objects are composed of other

objects take something as simple as a bicycle a bicycle is composed of a frame and wheels and pedals and a chain and a seat and so on um and each of those are

composed of other objects so we have this sort of hierarchical model of objects of objects we also learn where objects are located relative to each other this is part of our model so uh if

I go into a room and I see chairs and a table I not just see them as a list of objects but I see where they are relative to each other same with a bicycle a bicycle is not just a set of components there are wheels are

positions relative to the frame positions relative to the seat and so on we have to learn those relative positions and objects in the world have behaviors um take your smartphone you

bring up an app we you touch it certain things you expect to happen certain things you know how you swipe what happens you go this way what happens you push a button or the bicycle what happens when you turn the handlebars how do the brakes work well how do you what

happens when you pedal and change gears and so on and part of our model of the world is how objects behave we have to understand how the brain can represent that information then finally the model of Our World includes things which are

not physical they're not they're more abstract such as what's you know math or what's the definition of democracy so we have have these this very complex model of the world and it's

all in our neor cortex and of course as I said earlier we use these models to create goal oriented behaviors once you have a model of how things work you can say well I

want to accomplish something given my model of the world what would be the best actions most likely actions that would lead to the result I want and that's the whole goal of the neur cortex is to have the basically do goal

oriented behavior but you need this internal model first and finally I would say the model is predictable uh this is important to understand both how it works and how we

discovered how it works um um when I say it's predictive it means that it's constantly predicting what it's going to experience next something as simple as uh as I've touched something my brain is

going to my new cortex is going to predict what it's going to touch or if I see something it's going to predict what I see when I walk into a familiar room if I look left I expect to see one thing if look right I expect to see the other

thing this prediction is occurring all the time um we're not generally aware of it but we know that it's occurring because if anything in the world changes um you'll notice it and that's

proof that the brain is making a prediction so even if you're not thinking about something if something feels different or looks different immediately your attenion is drawn to it that tells you if the brain is making a prediction about that and one of the

reasons the brain does this one of the advantages of it it uses prediction as a training signal um so when we have a model of the world something is wrong

the prediction is incorrect brain knows to attend to that thing and says this part of my model is incorrect let's learn that part of the world let's see what's different how did how did I get this wrong and so we don't train

ourselves just by throwing a million images in front of us or doing them things a million times we built this model up incrementally we have a model we make predictions the predictions are

incorrect and we fix the model and so on now how is this all implemented in the brain and and in the neocortex and so we're going to dive a little bit into the Neuroscience here um and we're going

to talk about something called the cortical columns this is how the Neo cortex is structured this upper image represents a sort of artistic drawing of

of a a small piece of the Neo cortex so this is uh and it's showing it's organized into these columns um and the neus as I mentioned earlier is about two

and a half millimeters thick and these columns are roughly about a millimeter in diameter um so now if you actually were to look under a microscope at the ne cartrix you wouldn't see these columns like this they're not visible

like this but we know they exist and I'll explain why in a second in a human if we assume that the columns are about a millimeter in diameter then we have about 150,000 of them in our NE cortex

so that's how your NE cortex is organized into the 150,000 of these columns um that as you'll see in a moment are remarkably the same they look

very similar everywhere so this is a repetitive element that's neortx is built with now how do we know that NE columns exist um there's a lot of evidence I'll just show an image here

from uh a paper that was published in 1997 by Vernon M capsule and just gives you the idea of it here in this image there are shown six columns in the

cortex and the experimentor put a probe through them horizontally at an angle down you can see the line coming down diagonally and it crosses through different parts you know horizontally

through the ne cortex and and then you can see that the inputs to these columns are coming from different patches of skin on the hand of the case of a

monkey and so when they put the probe in they would notice that all the cells seem to respond to one patch of skin and that's occurs then after about a

millimeter travel um they' all switch over the cells start responding to a different patch of skin for about a millimeter and then and then after a millimeter they switch to another patch

of skin so the columns are defined by the fact that all the cells in a particular column are getting input from the same sensory patch if you will and um and then the next column over from a

different c patch that's that that general idea applies through out the the new cortex in the visional system and the auditory system and even higher levels which is a little bit difficult

to imagine at first but this is the general organization of the neuroc cortex now when scientists first started looking internal to the brain looking at

what did the the neocortex look like what do these columns look like um they started making pictures like this this goes back a long time to

1899 from famous scientist gahal and here you're kind of looking at an image of of the cells in uh a small region of Cortex and again it's two and a half

millimeters thickness across the all the regions and so um on the left image you can see the different some of the cell bodies and you can see there are different densities and different sizes

and this LEDs you to the idea that there are different layers of cells in the corex so now scientists talk about there's different layers of cells that cut across the two and a half millimeters but if you look at the

connections between the cells that's in the rightmost LI you see that most of the connections go vertically so most the information flows up and down across the two and a half millim up and down within a column between the different

layers of C this is what most of the processing occurs it comes into a column gets process up and down and then eventually it will go someplace else but most of the connections are vertical and

scientists like ourselves we make diagrams that look like this now there are literally thousands of papers that have been published on the architecture of the New York portex it's very

detailed and complex um area of science but we can try to do give you a summary of what it looks like something like this where you have different cell types they um connect to each other some in

One Direction some bidirectional they have different attributes and so on um I'll make just a few very high level observations first of all columns are

complex in a in a square millimeter neocortex there's about 100,000 neurons about 500 million synapses there are

dozens of different cell types and and and and they they're connected in various prototypical circuits so there's a very complex circuits it's not just a

big mess of wires it's very or organized um and there's actually something called mini columns and there's hundreds of mini columns inside of each column so we can summarize that whatever columns do

is also complex now I mentioned this because uh some AI researchers treat the cortex is doing simple things like just feature extraction but this much more is

going on here this is a very complex circuitry um that's doing something complex another surprising thing is that all columns have a motor out you there's a layer of cells which

project someplace in the brain that makes something move so even the parts of the cortex that get input directly from the eyes um they have cells which

project pack to the part of the brain that moves the eyes and so um you the part that's processing the input from the eyes is also directing how the eyes

should move um and that's a common feature everywhere the the part that gets input from your ears uh directs how your head and it angle is changing so when you change move your head you

change what you hear just like when you move your eyes you change what you see so this is a as far as you know a universal property in your equipments and as I mentioned earlier uh what is

perhaps most surprising is that all the columns are remarkably similar The Columns that are processing Vision look like the columns that are processing touch and the columns that are processing language they're not

identical there are some small chain differences but by and large they have the same overall architecture which is surprising so Veron malast again the same scientist I mentioned earlier he

wrote a paper a famous paper in 1979 and he made the following speculations he said look the reason the columns look similar is because they perform the same intrinsic function they're all doing the

same thing somehow vision is the same as touch which is the same as hearing which is the same as language not obvious but that's what the Neuroscience is telling us then he went on say well what it does

is what it's connected to so if you take a bunch of col connect them to the eyes you get Vision if you connect them to the ears you get hearing if you connect them to the skin you get touch and if you take a bunch of columns and outputs

and connecting to other colums you get things like high level thought and language so he made the observation well if we could understand what a column

does um that would be really important as he said would have great generalizing significance so this is some sense uh a major quest of Neuroscience is to understand how columns what they do and

how they work and we think we figured out basically what's going on here and that's the essence of the theory that I'm about to tell you about so what does a column do

uh let me just start with a little thought experiment which is how we sort of broke through on this and and had a real uh discovery about this one day I

was holding um a ntic coffee cup this one and I was touching it with my hand and of course if I touch my just with one finger on the cup um and as I move move it around the cup my brain makes a

prediction what it's going to feel so as I move my f my finger um it will predict oh in this case oh it's gonna it's G to feel the rounded side of the cup and if I move my finger to the top of the cup

and touch the rim it'll make a different prediction most of the time you're not you wouldn't be familiar you wouldn't be aware of these predictions but you can think about it and as your finger is moving you can anticipate what it's going to feel like that was an

interesting observation it says well how does it how does my brain know what it's going to feel before it actually gets to the edge of the cup in this case and um you can do this without looking you don't have to be looking at the cup you

can do this in the dark so it's not like oh I see what I'm going to feel no it tells you that the brain has a model of the cup and it knows where my finger is on the cup and it knows where my finger

will be after it's finished moving it needs to know that to make a prediction so that was interesting because it means that bra is tracking the location of my finger relative this

object as I'm touching it now I can touch the object with multiple fingers at the same time so here I'm grabbing the cup with my thumb and my index finger and my other fingers in the back in each patch of skin that's touching

the cup has its own prediction about what it will feel and if it felt something different you'd know right away if there was a rough patch or a crack or something was not rounded and was more like an edge you would notice

this so the brain has to keep track of the location of each patch of skin that's touching the cup and checking against its model of the cup saying what am I expected to feel as my fingers move

around that cup so this led to the following conclusions U to predict what you'll feel the neocortex must must know the location of each patch of skin

relative to the cup which is a surprising thing if you think about it and to know a location requires a reference frame now a reference frame uh we're all familiar with them you could

think of a reference frame as like the XY and Z coordinates uh cartisian coordinates we learned in school um it's a way of keeping track of the location or specifying the location of something

so the brain has to have a reference frame that's keeping track of the location of every patch of my skin that's touching this um cup and that reference frame is relative to the cup

it's not relative to my body or to something else it's relative to the cup so this is a surprising thing so how did the brain do this how did neurons actually do this but this is not you can it's logically determined so it's not

really speculation it has to be doing this so this was a big insight for us um and so then we we we develop a theory based on this we published this theory

in a paper called a theory of how columns in newor cext enable learning of the structure of the world and this the can be summarized in a few slides here um here's a a cartoon

image of a finger touching the cup and the Green Arrow represents the sensation input that's going into a column in the newo cortex that's a column that's rep that's capturing the input from the finger so that's the sense feature

that's coming in but it's going to be paired with a location so there's going to be another layer of cells that's representing the location of the finger relative of the cup and so I can now if

my finger moves around the cup I have different locations and different Sensations and this column can learn a model of the cup it can learn the shape or morphology of a cup uh or any object

that it's touching by sensing and moving and sensing and moving by pairing a location with a sensation which is that blue arrow the blue arrow is basically learning how to make these U what what

the different features are different locations on the cup so uh this is a a part of what we propos that there'd be these reference frames in every cortical column but now what happens when you

have multiple fingers touching the cup um oh I should mention here there's um you can then Define an object an object is just essentially another layer of cells would represent the object which is essentially the collection of all

sense features and locations so imagine now I have three fingers t uh touching the cup and uh so we have multiple columns well at first you might make the observation that each finger is at a different location so

each finger is going to a different column and each column has its own location relative to the cup in fact each column has its own model of the cup there are now three models of this cup

coming into these different um sensory patches uh and and the brain has to keep track of the location of each finger because they can move somewhat independently but something else is

going on here when I when I reach in if I touch the cup with a full hand I may not have to move my fingers around to actually recognize the cup so what's going on is that columns can talk to

each other uh to what we call voting there are connections here these horizontal green arrows I've shown here there are these long range connections that go between columns uh these are

very well documented um and we propos what they're doing is they're allowing the columns to vote they're basically saying each column has a sort of partial knowledge about the world each one's saying well I'm touching something I'm not sure exactly what it is I might have

a guess it's a Cy cup but it might be something else and by collaborating they can all come to a common agreement at once they can say yeah we all agree this is a cup um and that's that is what you

perceive it's like okay yes there are many models here but I now everyone knows it's a cppy cup now this perhaps is somewhat obvious when you think about touch you can think about your fingers

moving independently but the same process is going on in vision and this is not so obvious the way to think about vision is your brain doesn't look at an image of the world

uh this is welln to neuroscientist it's not like that um in some sense what is actually happening is the cortex has different columns each column is getting

input from a different patch of the retina and so each patch of the retina is like a different finger in some sense each patch of the retina is looking at a different location in the world and each patch of the retina is trying to model

understand what it's looking at and they vote together so Vision Works the same way uh although it may not be obvious to you uh one way you can perhaps make it a little bit more obvious imagine looking

at the world through a straw and so you can only see a little bit of the world at once this is like touching the world with a single finger you can still learn what objects look like by moving the straw around and you can still recognize

objects by moving the straw around that would be like touching and recognizing something with a single finger but if you look at the world through your entire retina it's like grabbing an

object with all your fingers at once and and now you can recognize what it is so this SL two are very interesting question from a neuroscientist point of view is how is it that neurons these

cells can represent reference frames to know where the location of something is um if if you know anything about brains this is puzzling and it was puzzling to us but we we concluded this must be

happening the answer wasn't so hard actually um turns out there are um there's a part of the brain that's not in the neocortex um there these two parts

called the hippocampus and the rhinal cortex in the humans they're very in the center of the brain about the size of your small finger uh and there's these cells

in the anthero cortex called grid cells that um are known to form a type of reference frame for environments um and so when an animal is

typically a rat um moves around in an environment like a box or maze these grid cells in the antic sort of represent where the animal is um there's

other cells in the hippocampus called play cells which are uh also represent where the animal is but more based on sensory information so it's a sort of a

pairing of location and sensory input um and these two cell types have been very well studied and so what we speculated was that the neocortex would use a

similar mechanism it would have something equivalent to grid cells and it was something equivalent to play cells uh so uh that's what we proposed in a series of papers that there would

be grid cells and play cells equivalents in every quter column in the Neo cortex and these would create reference frames for objects here's a here's another image just to help you get understand

what I'm talking about here in the left picture there's a a cartoon of a rat moving around an environment and I've labeled three locations there D and F

and grid cells represent where the rat is so when the rat is at location D the the grid tells a certain firing pattern and um when they at to location e they

have a different C firing pattern and if the rat goes back to D no matter how it gets there what direction and so on it's going to the same sort of cells will be firing so these bit cells represent locations in the environment as a

reference R what we're proposing is that in the cortex something almost the same thing is happening but instead of tracking the location of your body in a room like the rat we're tracking the

location of a sensory patch like with your finger and so as your finger moves over the coffee cup it's the same as the rat moving around the room there are grid cells representing the location of the finger relative to the coffee cup

and and so like every time you get to location on the rim labeled X those same cells would be active so now we put this together we can we can basically make this this a

very simple version of it it's not it's more complicated than this but we can say that there's a layer of cells in each cortical column which we would call cortical grid cells this is the reference frame representing where a

sensory input is and then' be somewhat of equivalent to play cell which is the sense features at a location um and therefore the each cortical column is modeling something in the world just the

same way as grid cells and play cells now we made this prediction um and when we made this prediction it was not known that there are grid cells in NE cortex but you'll see in a moment there's lots

of evidence for that now back right here um there's growing evidence for this uh and and I'm not going to walk you through this experiment it's very complex experiment

but I want just to give you a flavor that people are are doing these things so there's now a lot of evidence are grid cells in human NE cortex and other animals NE cortex as well this is a an

experiment that was done using humans in fmar one of those machines that can detect where activity is in the brain and humans were asked to train to do think about birds in this case the birds

had were different differed by the lengths of their necks and their legs and so what the scientists were able to show um is that the uh that when we

thought about when a human thought about birds and thought of like oh what would a bird be like or how would I organize them and sort them and things like that that they were able to show that the

neocortex is using uh grid cells to organize knowledge about the bird exactly what I was just talking about um so grid cells formed the basis for knowledge in this case when people

learned about birds it was forming the basis for knowledge about birds and um it's very clever experiments there are now numerous experiments show that there are grid cell like um cells throughout

the neuro cortex so this is a a good confirmation at least of the major proposal of this Theory there's one more part about brain I want to mention um

before we we wrap up here and that the brain is generally viewed as a hierarchical system it's and and so with most AI systems so let's talk about that

in this case um on the left here is a very simplified view of how people often think about the brain you get input from say from the eye from the retina it goes to a region called V1 which extracts

simple features then to another region called V2 which extracts more complex features and so on until you get to a region that you can recognize an object and today's AI systems are largely the

Deep learning networks largely are based on this idea although in deep loaring network instead of having three or four layers like in a human brain they have maybe a hundred layers so it's it's it's

same idea but much more depth to the models um but when you look at an actual brain which is the set image in the middle of the the slide um it doesn't look very hierarchical at all now this

very complex image is actually famous for neuroscientist every neuroscientist knows this image and and what it represents is a monkey's brain the little boxes are the little rectangles are different regions of the monkey's

neur cortex and so there's dozens of them here and then um the lines represent how they're connected together so you can make a few observations first of all this does not look like a simple

hierarchy and it isn't in fact most of the connections that are observed in the brain are not hierarchical like in a flowchart they're they kind of go all over the place another way of putting that is that more than 40% of all the

possible connections between regions exist which is not something you'd see in the hierarchy another surprising result is some of the the some of the regions the ones that get direct input from the eyes or the ears or the skin

these are the largest regions in the neocortex which is surprising because if you most people think that you're just extracting some sort of feature or some sort of simple feature action why would those be the largest regions whereas the

other higher level regions are smaller and um and then also that we find even regions of the brain that get direct input from the eye such is V1

they also are they get influenced by what we hear and what we touch so they're multimodal which doesn't make sense in a hierarchy either but all these things do make sense in the Thousand brains

Theory U that I'm talking about uh here's a way to think about it um in the Thousand brains theory of intelligence there isn't a single model of something like a coffee cup there are

actually thousands of models of a coffee cup they're not identical they're complimentary they exist in different cortical columns I'm not saying that every cortical column has a model of coffee cup that's not true but there are

many models of the coffee cup and um there are models that represent what the coffee cup looks like and there are models that represent what the coffee cup feels like and most of the connections in the brain are for voting

between the columns so when I see the coffee cup I can imagine what it's going to feel and I feel the coffee cup I can imagine what it's going to see or if I have a very impoverished view of something and I touch it I'll be able to figure out what

it is and so on so those are represented by these blue arrows here and we also have models at different scales um so the way to think about the cortex is there thousands of models of everything

and they're voting all the time uh on what's going on now there is hierarchy there's still hierarchy in the brain we we know that's true some of the connections are

but they're not just passing features they're passing models so you know a bicycle as I mentioned earlier is composed of wheels and and chain and a

frame and so on so when we when we build structured models of the world we are learning models of models and that's what the hierarchy uh lets you do it's not hierarchy of features but hierarchy

of models so that's kind of wraps up the basic Theory here so let me just summarize uh the Thousand brain theory of intelligence I'm barely talking about biological intelligence so far and the

main key walkways are the brain learns a model of the world this is how you have to think about it this is don't think about the brain it's performing some function it's learning a model of the

world and it is a distributed model meaning in the brain it's distributed over thousands of different cortical columns and those models are learned by sensory motor integration you can't the

brain cannot learn without movement this is key to how we understand the world uh the models then use reference frames to represent knowledge so we don't just store a list of facts we don't just

store statistics we build a structured model of the world using reference frames and then the models vote to reach a consensus this is the core of what's going on in your head um and everyone

else's head um now I'm going to just State the following machine intelligence uh I believe I'm certain of this that true machine intelligence must work on

the same principles as the brain these are not optional components uh the AI we have today uh despite its its really

excellent performance and certain tasks and how you how well it works it doesn't really have the attributes we talked about earlier it doesn't really know what it's doing it's more sta these are more statistical models they don't have

structured knowledge of the world they don't know how to act upon the world they don't have a sense of how the world behaves and how we can generalize a cost models these all require the reference names and the distributive modeling

system I've talked about here so although we have a lot of great AI today I don't believe it's going to be the dominant technology going forward into the rest of this uh Century I believe that machine intelligence will be based

on most of Machine T will be based on the principles I've laid out here so it'll be the dominant AI technology of the 21st century uh I'm very confident of that um at my company the Mento we

are uh we are working to make this happen I'm happy to promote it which why I'm talking about it here um we're trying to get other people working on this too but um this is doable this is not something we have to wait we know

how to we know how to do this today