back to indexJeff Hawkins: Thousand Brains Theory of Intelligence | Lex Fridman Podcast #25
link |
The following is a conversation with Jeff Hawkins.
link |
He's the founder of the Redwood Center
link |
for Theoretical Neuroscience in 2002, and NuMenta in 2005.
link |
In his 2004 book, titled On Intelligence,
link |
and in the research before and after,
link |
he and his team have worked to reverse engineer
link |
the neural cortex, and propose artificial intelligence
link |
architectures, approaches, and ideas
link |
that are inspired by the human brain.
link |
These ideas include Hierarchical Tupperware Memory,
link |
HTM, from 2004, and new work,
link |
the Thousand Brains Theory of Intelligence
link |
from 2017, 18, and 19.
link |
Jeff's ideas have been an inspiration
link |
to many who have looked for progress
link |
beyond the current machine learning approaches,
link |
but they have also received criticism
link |
for lacking a body of empirical evidence
link |
supporting the models.
link |
This is always a challenge when seeking more
link |
than small incremental steps forward in AI.
link |
Jeff is a brilliant mind, and many of the ideas
link |
he has developed and aggregated from neuroscience
link |
are worth understanding and thinking about.
link |
There are limits to deep learning,
link |
as it is currently defined.
link |
Forward progress in AI is shrouded in mystery.
link |
My hope is that conversations like this
link |
can help provide an inspiring spark for new ideas.
link |
This is the Artificial Intelligence Podcast.
link |
If you enjoy it, subscribe on YouTube, iTunes,
link |
or simply connect with me on Twitter
link |
at Lex Friedman, spelled F R I D.
link |
And now, here's my conversation with Jeff Hawkins.
link |
Are you more interested in understanding the human brain
link |
or in creating artificial systems
link |
that have many of the same qualities
link |
but don't necessarily require that you actually understand
link |
the underpinning workings of our mind?
link |
So there's a clear answer to that question.
link |
My primary interest is understanding the human brain.
link |
No question about it.
link |
But I also firmly believe that we will not be able
link |
to create fully intelligent machines
link |
until we understand how the human brain works.
link |
So I don't see those as separate problems.
link |
I think there's limits to what can be done
link |
with machine intelligence if you don't understand
link |
the principles by which the brain works.
link |
And so I actually believe that studying the brain
link |
is actually the fastest way to get to machine intelligence.
link |
And within that, let me ask the impossible question,
link |
how do you, not define, but at least think
link |
about what it means to be intelligent?
link |
So I didn't try to answer that question first.
link |
We said, let's just talk about how the brain works
link |
and let's figure out how certain parts of the brain,
link |
mostly the neocortex, but some other parts too.
link |
The parts of the brain most associated with intelligence.
link |
And let's discover the principles by how they work.
link |
Because intelligence isn't just like some mechanism
link |
and it's not just some capabilities.
link |
It's like, okay, we don't even know
link |
where to begin on this stuff.
link |
And so now that we've made a lot of progress on this,
link |
after we've made a lot of progress
link |
on how the neocortex works, and we can talk about that,
link |
I now have a very good idea what's gonna be required
link |
to make intelligent machines.
link |
I can tell you today, some of the things
link |
are gonna be necessary, I believe,
link |
to create intelligent machines.
link |
Well, so we'll get there.
link |
We'll get to the neocortex and some of the theories
link |
of how the whole thing works.
link |
And you're saying, as we understand more and more
link |
about the neocortex, about our own human mind,
link |
we'll be able to start to more specifically define
link |
what it means to be intelligent.
link |
It's not useful to really talk about that until.
link |
I don't know if it's not useful.
link |
Look, there's a long history of AI, as you know.
link |
And there's been different approaches taken to it.
link |
And who knows, maybe they're all useful.
link |
So the good old fashioned AI, the expert systems,
link |
the current convolutional neural networks,
link |
they all have their utility.
link |
They all have a value in the world.
link |
But I would think almost everyone agree
link |
that none of them are really intelligent
link |
in a sort of a deep way that humans are.
link |
And so it's just the question of how do you get
link |
from where those systems were or are today
link |
to where a lot of people think we're gonna go.
link |
And there's a big, big gap there, a huge gap.
link |
And I think the quickest way of bridging that gap
link |
is to figure out how the brain does that.
link |
And then we can sit back and look and say,
link |
oh, which of these principles that the brain works on
link |
are necessary and which ones are not?
link |
Clearly, we don't have to build this in,
link |
and intelligent machines aren't gonna be built
link |
out of organic living cells.
link |
But there's a lot of stuff that goes on the brain
link |
that's gonna be necessary.
link |
So let me ask maybe, before we get into the fun details,
link |
let me ask maybe a depressing or a difficult question.
link |
Do you think it's possible that we will never
link |
be able to understand how our brain works,
link |
that maybe there's aspects to the human mind,
link |
like we ourselves cannot introspectively get to the core,
link |
that there's a wall you eventually hit?
link |
Yeah, I don't believe that's the case.
link |
I have never believed that's the case.
link |
There's not been a single thing humans have ever put
link |
their minds to that we've said, oh, we reached the wall,
link |
we can't go any further.
link |
It's just, people keep saying that.
link |
People used to believe that about life.
link |
Alain Vital, right, there's like,
link |
what's the difference between living matter
link |
and nonliving matter, something special
link |
that we never understand.
link |
We no longer think that.
link |
So there's no historical evidence that suggests this
link |
is the case, and I just never even consider
link |
that's a possibility.
link |
I would also say, today, we understand so much
link |
about the neocortex.
link |
We've made tremendous progress in the last few years
link |
that I no longer think of it as an open question.
link |
The answers are very clear to me.
link |
The pieces we don't know are clear to me,
link |
but the framework is all there, and it's like,
link |
oh, okay, we're gonna be able to do this.
link |
This is not a problem anymore, just takes time and effort,
link |
but there's no mystery, a big mystery anymore.
link |
So then let's get into it for people like myself
link |
who are not very well versed in the human brain,
link |
Can you describe to me, at the highest level,
link |
what are the different parts of the human brain,
link |
and then zooming in on the neocortex,
link |
the parts of the neocortex, and so on,
link |
The human brain, we can divide it roughly into two parts.
link |
There's the old parts, lots of pieces,
link |
and then there's the new part.
link |
The new part is the neocortex.
link |
It's new because it didn't exist before mammals.
link |
The only mammals have a neocortex,
link |
and in humans, in primates, it's very large.
link |
In the human brain, the neocortex occupies
link |
about 70 to 75% of the volume of the brain.
link |
And the old parts of the brain are,
link |
there's lots of pieces there.
link |
There's the spinal cord, and there's the brain stem,
link |
and the cerebellum, and the different parts
link |
of the basal ganglia, and so on.
link |
In the old parts of the brain,
link |
you have the autonomic regulation,
link |
like breathing and heart rate.
link |
You have basic behaviors, so like walking and running
link |
are controlled by the old parts of the brain.
link |
All the emotional centers of the brain
link |
are in the old part of the brain,
link |
so when you feel anger or hungry, lust,
link |
or things like that, those are all
link |
in the old parts of the brain.
link |
And we associate with the neocortex
link |
all the things we think about as sort of
link |
high level perception and cognitive functions,
link |
anything from seeing and hearing and touching things
link |
to language to mathematics and engineering
link |
and science and so on.
link |
Those are all associated with the neocortex,
link |
and they're certainly correlated.
link |
Our abilities in those regards are correlated
link |
with the relative size of our neocortex
link |
compared to other mammals.
link |
So that's like the rough division,
link |
and you obviously can't understand
link |
the neocortex completely isolated,
link |
but you can understand a lot of it
link |
with just a few interfaces to the old parts of the brain,
link |
and so it gives you a system to study.
link |
The other remarkable thing about the neocortex,
link |
compared to the old parts of the brain,
link |
is the neocortex is extremely uniform.
link |
It's not visibly or anatomically,
link |
it's very, I always like to say
link |
it's like the size of a dinner napkin,
link |
about two and a half millimeters thick,
link |
and it looks remarkably the same everywhere.
link |
Everywhere you look in that two and a half millimeters
link |
is this detailed architecture,
link |
and it looks remarkably the same everywhere,
link |
and that's across species.
link |
A mouse versus a cat and a dog and a human.
link |
Where if you look at the old parts of the brain,
link |
there's lots of little pieces do specific things.
link |
So it's like the old parts of our brain evolved,
link |
like this is the part that controls heart rate,
link |
and this is the part that controls this,
link |
and this is this kind of thing,
link |
and that's this kind of thing,
link |
and these evolved for eons a long, long time,
link |
and they have their specific functions,
link |
and all of a sudden mammals come along,
link |
and they got this thing called the neocortex,
link |
and it got large by just replicating the same thing
link |
over and over and over again.
link |
This is like, wow, this is incredible.
link |
So all the evidence we have,
link |
and this is an idea that was first articulated
link |
in a very cogent and beautiful argument
link |
by a guy named Vernon Malcastle in 1978, I think it was,
link |
that the neocortex all works on the same principle.
link |
So language, hearing, touch, vision, engineering,
link |
all these things are basically underlying,
link |
are all built on the same computational substrate.
link |
They're really all the same problem.
link |
So the low level of the building blocks all look similar.
link |
Yeah, and they're not even that low level.
link |
We're not talking about like neurons.
link |
We're talking about this very complex circuit
link |
that exists throughout the neocortex.
link |
It's remarkably similar.
link |
It's like, yes, you see variations of it here and there,
link |
more of the cell, less and less, and so on.
link |
But what Malcastle argued was, he says,
link |
you know, if you take a section of neocortex,
link |
why is one a visual area and one is a auditory area?
link |
Or why is, and his answer was,
link |
it's because one is connected to eyes
link |
and one is connected to ears.
link |
Literally, you mean just it's most closest
link |
in terms of number of connections
link |
to the sensor. Literally, literally,
link |
if you took the optic nerve and attached it
link |
to a different part of the neocortex,
link |
that part would become a visual region.
link |
This actually, this experiment was actually done
link |
by Merkankasur in developing, I think it was lemurs,
link |
I can't remember what it was, some animal.
link |
And there's a lot of evidence to this.
link |
You know, if you take a blind person,
link |
a person who's born blind at birth,
link |
they're born with a visual neocortex.
link |
It doesn't, may not get any input from the eyes
link |
because of some congenital defect or something.
link |
And that region becomes, does something else.
link |
It picks up another task.
link |
So, and it's, so it's this very complex thing.
link |
It's not like, oh, they're all built on neurons.
link |
No, they're all built in this very complex circuit
link |
and somehow that circuit underlies everything.
link |
And so this is the, it's called
link |
the common cortical algorithm, if you will.
link |
Some scientists just find it hard to believe
link |
and they just, I can't believe that's true,
link |
but the evidence is overwhelming in this case.
link |
And so a large part of what it means
link |
to figure out how the brain creates intelligence
link |
and what is intelligence in the brain
link |
is to understand what that circuit does.
link |
If you can figure out what that circuit does,
link |
as amazing as it is, then you can,
link |
then you understand what all these
link |
other cognitive functions are.
link |
So if you were to sort of put neocortex
link |
outside of your book on intelligence,
link |
you look, if you wrote a giant tome, a textbook
link |
on the neocortex, and you look maybe
link |
a couple of centuries from now,
link |
how much of what we know now would still be accurate
link |
two centuries from now?
link |
So how close are we in terms of understanding?
link |
I have to speak from my own particular experience here.
link |
So I run a small research lab here.
link |
It's like any other research lab.
link |
I'm sort of the principal investigator.
link |
There's actually two of us
link |
and there's a bunch of other people.
link |
And this is what we do.
link |
We study the neocortex and we publish our results
link |
So about three years ago,
link |
we had a real breakthrough in this field.
link |
Just tremendous breakthrough.
link |
We've now published, I think, three papers on it.
link |
And so I have a pretty good understanding
link |
of all the pieces and what we're missing.
link |
I would say that almost all the empirical data
link |
we've collected about the brain, which is enormous.
link |
If you don't know the neuroscience literature,
link |
it's just incredibly big.
link |
And it's, for the most part, all correct.
link |
It's facts and experimental results and measurements
link |
and all kinds of stuff.
link |
But none of that has been really assimilated
link |
into a theoretical framework.
link |
It's data without, in the language of Thomas Kuhn,
link |
the historian, would be a sort of a pre paradigm science.
link |
Lots of data, but no way to fit it together.
link |
I think almost all of that's correct.
link |
There's just gonna be some mistakes in there.
link |
And for the most part,
link |
there aren't really good cogent theories about it,
link |
how to put it together.
link |
It's not like we have two or three competing good theories,
link |
which ones are right and which ones are wrong.
link |
It's like, nah, people are just scratching their heads.
link |
Some people have given up
link |
on trying to figure out what the whole thing does.
link |
In fact, there's very, very few labs that we do
link |
that focus really on theory
link |
and all this unassimilated data and trying to explain it.
link |
So it's not like we've got it wrong.
link |
It's just that we haven't got it at all.
link |
So it's really, I would say, pretty early days
link |
in terms of understanding the fundamental theory's forces
link |
of the way our mind works.
link |
I would have said that's true five years ago.
link |
we had some really big breakthroughs on this recently
link |
and we started publishing papers on this.
link |
So we'll get to that.
link |
But so I don't think it's,
link |
I'm an optimist and from where I sit today,
link |
most people would disagree with this,
link |
but from where I sit today, from what I know,
link |
it's not super early days anymore.
link |
We are, the way these things go
link |
is it's not a linear path, right?
link |
You don't just start accumulating
link |
and get better and better and better.
link |
No, all this stuff you've collected,
link |
none of it makes sense.
link |
All these different things are just sort of around.
link |
And then you're gonna have some breaking points
link |
where all of a sudden, oh my God, now we got it right.
link |
That's how it goes in science.
link |
And I personally feel like we passed that little thing
link |
about a couple of years ago,
link |
all that big thing a couple of years ago.
link |
So we can talk about that.
link |
Time will tell if I'm right,
link |
but I feel very confident about it.
link |
That's why I'm willing to say it on tape like this.
link |
At least very optimistic.
link |
So let's, before those few years ago,
link |
let's take a step back to HTM,
link |
the hierarchical temporal memory theory,
link |
which you first proposed on intelligence
link |
and went through a few different generations.
link |
Can you describe what it is,
link |
how it evolved through the three generations
link |
since you first put it on paper?
link |
Yeah, so one of the things that neuroscientists
link |
just sort of missed for many, many years,
link |
and especially people who were thinking about theory,
link |
was the nature of time in the brain.
link |
Brains process information through time.
link |
The information coming into the brain
link |
is constantly changing.
link |
The patterns from my speech right now,
link |
if you were listening to it at normal speed,
link |
would be changing on your ears
link |
about every 10 milliseconds or so, you'd have a change.
link |
This constant flow, when you look at the world,
link |
your eyes are moving constantly,
link |
three to five times a second,
link |
and the input's completely changing.
link |
If I were to touch something like a coffee cup,
link |
as I move my fingers, the input changes.
link |
So this idea that the brain works on time changing patterns
link |
is almost completely, or was almost completely missing
link |
from a lot of the basic theories,
link |
like fears of vision and so on.
link |
It's like, oh no, we're gonna put this image
link |
in front of you and flash it and say, what is it?
link |
Convolutional neural networks work that way today, right?
link |
Classify this picture.
link |
But that's not what vision is like.
link |
Vision is this sort of crazy time based pattern
link |
that's going all over the place,
link |
and so is touch and so is hearing.
link |
So the first part of hierarchical temporal memory
link |
was the temporal part.
link |
It's to say, you won't understand the brain,
link |
nor will you understand intelligent machines
link |
unless you're dealing with time based patterns.
link |
The second thing was, the memory component of it was,
link |
is to say that we aren't just processing input,
link |
we learn a model of the world.
link |
And the memory stands for that model.
link |
The point of the brain, the part of the neocortex,
link |
it learns a model of the world.
link |
We have to store things, our experiences,
link |
in a form that leads to a model of the world.
link |
So we can move around the world,
link |
we can pick things up and do things and navigate
link |
and know how it's going on.
link |
So that's what the memory referred to.
link |
And many people just, they were thinking about
link |
like certain processes without memory at all.
link |
They're just like processing things.
link |
And then finally, the hierarchical component
link |
was a reflection to that the neocortex,
link |
although it's this uniform sheet of cells,
link |
different parts of it project to other parts,
link |
which project to other parts.
link |
And there is a sort of rough hierarchy in terms of that.
link |
So the hierarchical temporal memory is just saying,
link |
look, we should be thinking about the brain
link |
as time based, model memory based,
link |
and hierarchical processing.
link |
And that was a placeholder for a bunch of components
link |
that we would then plug into that.
link |
We still believe all those things I just said,
link |
but we now know so much more that I'm stopping to use
link |
the word hierarchical temporal memory yet
link |
because it's insufficient to capture the stuff we know.
link |
So again, it's not incorrect, but it's,
link |
I now know more and I would rather describe it
link |
Yeah, so you're basically, we could think of HTM
link |
as emphasizing that there's three aspects of intelligence
link |
that are important to think about
link |
whatever the eventual theory it converges to.
link |
So in terms of time, how do you think of nature of time
link |
across different time scales?
link |
So you mentioned things changing,
link |
sensory inputs changing every 10, 20 minutes.
link |
What about every few minutes, every few months and years?
link |
Well, if you think about a neuroscience problem,
link |
the brain problem, neurons themselves can stay active
link |
for certain periods of time, parts of the brain
link |
where they stay active for minutes.
link |
You could hold a certain perception or an activity
link |
for a certain period of time,
link |
but most of them don't last that long.
link |
And so if you think about your thoughts
link |
are the activity of neurons,
link |
if you're gonna wanna involve something
link |
that happened a long time ago,
link |
even just this morning, for example,
link |
the neurons haven't been active throughout that time.
link |
So you have to store that.
link |
So if I ask you, what did you have for breakfast today?
link |
That is memory, that is you've built into your model
link |
the world now, you remember that.
link |
And that memory is in the synapses,
link |
is basically in the formation of synapses.
link |
And so you're sliding into what,
link |
you know, it's the different timescales.
link |
There's timescales of which we are like understanding
link |
my language and moving about and seeing things rapidly
link |
and over time, that's the timescales
link |
of activities of neurons.
link |
But if you wanna get in longer timescales,
link |
then it's more memory.
link |
And we have to invoke those memories to say,
link |
oh yes, well now I can remember what I had for breakfast
link |
because I stored that someplace.
link |
I may forget it tomorrow, but I'd store it for now.
link |
So does memory also need to have,
link |
so the hierarchical aspect of reality
link |
is not just about concepts, it's also about time?
link |
Do you think of it that way?
link |
Yeah, time is infused in everything.
link |
It's like you really can't separate it out.
link |
If I ask you, what is your, you know,
link |
how's the brain learn a model of this coffee cup here?
link |
I have a coffee cup and I'm at the coffee cup.
link |
I say, well, time is not an inherent property
link |
of the model I have of this cup,
link |
whether it's a visual model or a tactile model.
link |
I can sense it through time,
link |
but the model itself doesn't really have much time.
link |
If I asked you, if I said,
link |
well, what is the model of my cell phone?
link |
My brain has learned a model of the cell phone.
link |
So if you have a smartphone like this,
link |
and I said, well, this has time aspects to it.
link |
I have expectations when I turn it on,
link |
what's gonna happen, what or how long it's gonna take
link |
to do certain things, if I bring up an app,
link |
what sequences, and so I have,
link |
and it's like melodies in the world, you know?
link |
Melody has a sense of time.
link |
So many things in the world move and act,
link |
and there's a sense of time related to them.
link |
Some don't, but most things do actually.
link |
So it's sort of infused throughout the models of the world.
link |
You build a model of the world,
link |
you're learning the structure of the objects in the world,
link |
and you're also learning how those things change
link |
Okay, so it really is just a fourth dimension
link |
that's infused deeply, and you have to make sure
link |
that your models of intelligence incorporate it.
link |
So, like you mentioned, the state of neuroscience
link |
is deeply empirical, a lot of data collection.
link |
It's, you know, that's where it is.
link |
You mentioned Thomas Kuhn, right?
link |
And then you're proposing a theory of intelligence,
link |
and which is really the next step,
link |
the really important step to take,
link |
but why is HTM, or what we'll talk about soon,
link |
So is it more in the, is it backed by intuition?
link |
Is it backed by evidence?
link |
Is it backed by a mixture of both?
link |
Is it kind of closer to where string theory is in physics,
link |
where there's mathematical components
link |
which show that, you know what,
link |
it seems that this, it fits together too well
link |
for it not to be true, which is where string theory is.
link |
Is that where you're kind of seeing?
link |
It's a mixture of all those things,
link |
although definitely where we are right now
link |
is definitely much more on the empirical side
link |
than, let's say, string theory.
link |
The way this goes about, we're theorists, right?
link |
So we look at all this data, and we're trying to come up
link |
with some sort of model that explains it, basically,
link |
and there's, unlike string theory,
link |
there's vast more amounts of empirical data here
link |
that I think than most physicists deal with.
link |
And so our challenge is to sort through that
link |
and figure out what kind of constructs would explain this.
link |
And when we have an idea,
link |
you come up with a theory of some sort,
link |
you have lots of ways of testing it.
link |
First of all, there are 100 years of assimilated,
link |
assimilated, unassimilated empirical data from neuroscience.
link |
So we go back and read papers,
link |
and we say, oh, did someone find this already?
link |
We can predict X, Y, and Z,
link |
and maybe no one's even talked about it
link |
since 1972 or something, but we go back and find that,
link |
and we say, oh, either it can support the theory
link |
or it can invalidate the theory.
link |
And we say, okay, we have to start over again.
link |
Oh, no, it's supportive, let's keep going with that one.
link |
So the way I kind of view it, when we do our work,
link |
we look at all this empirical data,
link |
and what I call it is a set of constraints.
link |
We're not interested in something
link |
that's biologically inspired.
link |
We're trying to figure out how the actual brain works.
link |
So every piece of empirical data
link |
is a constraint on a theory.
link |
In theory, if you have the correct theory,
link |
it needs to explain every pin, right?
link |
So we have this huge number of constraints on the problem,
link |
which initially makes it very, very difficult.
link |
If you don't have many constraints,
link |
you can make up stuff all the day.
link |
You can say, oh, here's an answer on how you can do this,
link |
you can do that, you can do this.
link |
But if you consider all biology as a set of constraints,
link |
all neuroscience as a set of constraints,
link |
and even if you're working in one little part
link |
of the neocortex, for example,
link |
there are hundreds and hundreds of constraints.
link |
These are empirical constraints
link |
that it's very, very difficult initially
link |
to come up with a theoretical framework for that.
link |
But when you do, and it solves all those constraints
link |
at once, you have a high confidence
link |
that you got something close to correct.
link |
It's just mathematically almost impossible not to be.
link |
So that's the curse and the advantage of what we have.
link |
The curse is we have to solve,
link |
we have to meet all these constraints, which is really hard.
link |
But when you do meet them,
link |
then you have a great confidence
link |
that you've discovered something.
link |
In addition, then we work with scientific labs.
link |
So we'll say, oh, there's something we can't find,
link |
we can predict something,
link |
but we can't find it anywhere in the literature.
link |
So we will then, we have people we've collaborated with,
link |
we'll say, sometimes they'll say, you know what?
link |
I have some collected data, which I didn't publish,
link |
but we can go back and look at it
link |
and see if we can find that,
link |
which is much easier than designing a new experiment.
link |
You know, neuroscience experiments take a long time, years.
link |
So, although some people are doing that now too.
link |
So, but between all of these things,
link |
I think it's a reasonable,
link |
actually a very, very good approach.
link |
We are blessed with the fact that we can test our theories
link |
out the yin yang here because there's so much
link |
unassimilar data and we can also falsify our theories
link |
very easily, which we do often.
link |
So it's kind of reminiscent to whenever that was
link |
with Copernicus, you know, when you figure out
link |
that the sun's at the center of the solar system
link |
as opposed to earth, the pieces just fall into place.
link |
Yeah, I think that's the general nature of aha moments
link |
is, and it's Copernicus, it could be,
link |
you could say the same thing about Darwin,
link |
you could say the same thing about, you know,
link |
about the double helix,
link |
that people have been working on a problem for so long
link |
and have all this data and they can't make sense of it,
link |
they can't make sense of it.
link |
But when the answer comes to you
link |
and everything falls into place,
link |
it's like, oh my gosh, that's it.
link |
That's got to be right.
link |
I asked both Jim Watson and Francis Crick about this.
link |
I asked them, you know, when you were working on
link |
trying to discover the structure of the double helix,
link |
and when you came up with the sort of the structure
link |
that ended up being correct, but it was sort of a guess,
link |
you know, it wasn't really verified yet.
link |
I said, did you know that it was right?
link |
And they both said, absolutely.
link |
So we absolutely knew it was right.
link |
And it doesn't matter if other people didn't believe it
link |
or not, we knew it was right.
link |
They'd get around to thinking it
link |
and agree with it eventually anyway.
link |
And that's the kind of thing you hear a lot with scientists
link |
who really are studying a difficult problem.
link |
And I feel that way too about our work.
link |
Have you talked to Crick or Watson about the problem
link |
you're trying to solve, the, of finding the DNA of the brain?
link |
Yeah, in fact, Francis Crick was very interested in this
link |
in the latter part of his life.
link |
And in fact, I got interested in brains
link |
by reading an essay he wrote in 1979
link |
called Thinking About the Brain.
link |
And that was when I decided I'm gonna leave my profession
link |
of computers and engineering and become a neuroscientist.
link |
Just reading that one essay from Francis Crick.
link |
I got to meet him later in life.
link |
I spoke at the Salk Institute and he was in the audience.
link |
And then I had a tea with him afterwards.
link |
He was interested in a different problem.
link |
He was focused on consciousness.
link |
The easy problem, right?
link |
Well, I think it's the red herring.
link |
And so we weren't really overlapping a lot there.
link |
Jim Watson, who's still alive,
link |
is also interested in this problem.
link |
And he was, when he was director
link |
of the Cold Spring Harbor Laboratories,
link |
he was really sort of behind moving in the direction
link |
of neuroscience there.
link |
And so he had a personal interest in this field.
link |
And I have met with him numerous times.
link |
And in fact, the last time was a little bit over a year ago,
link |
I gave a talk at Cold Spring Harbor Labs
link |
about the progress we were making in our work.
link |
And it was a lot of fun because he said,
link |
well, you wouldn't be coming here
link |
unless you had something important to say.
link |
So I'm gonna go attend your talk.
link |
So he sat in the very front row.
link |
Next to him was the director of the lab, Bruce Stillman.
link |
So these guys are in the front row of this auditorium.
link |
Nobody else in the auditorium wants to sit in the front row
link |
because there's Jim Watson and there's the director.
link |
And I gave a talk and then I had dinner with him afterwards.
link |
But there's a great picture of my colleague Subitai Amantak
link |
where I'm up there sort of like screaming the basics
link |
of this new framework we have.
link |
And Jim Watson's on the edge of his chair.
link |
He's literally on the edge of his chair,
link |
like intently staring up at the screen.
link |
And when he discovered the structure of DNA,
link |
the first public talk he gave
link |
was at Cold Spring Harbor Labs.
link |
And there's a picture, there's a famous picture
link |
of Jim Watson standing at the whiteboard
link |
with an overrated thing pointing at something,
link |
pointing at the double helix with his pointer.
link |
And it actually looks a lot like the picture of me.
link |
So there was a sort of funny,
link |
there's Arian talking about the brain
link |
and there's Jim Watson staring intently at it.
link |
And of course there with, whatever, 60 years earlier,
link |
he was standing pointing at the double helix.
link |
That's one of the great discoveries in all of,
link |
whatever, biology, science, all science and DNA.
link |
So it's funny that there's echoes of that in your presentation.
link |
Do you think, in terms of evolutionary timeline and history,
link |
the development of the neocortex was a big leap?
link |
Or is it just a small step?
link |
So like, if we ran the whole thing over again,
link |
from the birth of life on Earth,
link |
how likely would we develop the mechanism of the neocortex?
link |
Okay, well those are two separate questions.
link |
One is, was it a big leap?
link |
And one was how likely it is, okay?
link |
They're not necessarily related.
link |
Maybe correlated, maybe not.
link |
And we don't really have enough data
link |
to make a judgment about that.
link |
I would say definitely it was a big leap.
link |
And I can tell you why.
link |
I don't think it was just another incremental step.
link |
I don't get that at the moment.
link |
I don't really have any idea how likely it is.
link |
If we look at evolution,
link |
we have one data point, which is Earth, right?
link |
Life formed on Earth billions of years ago,
link |
whether it was introduced here or it created it here,
link |
or someone introduced it, we don't really know,
link |
but it was here early.
link |
It took a long, long time to get to multicellular life.
link |
And then for multicellular life,
link |
it took a long, long time to get the neocortex.
link |
And we've only had the neocortex for a few 100,000 years.
link |
So that's like nothing, okay?
link |
Well, it certainly isn't something
link |
that happened right away on Earth.
link |
And there were multiple steps to get there.
link |
So I would say it's probably not gonna be something
link |
that would happen instantaneously
link |
on other planets that might have life.
link |
It might take several billion years on average.
link |
I don't know, but you'd have to survive
link |
for several billion years to find out.
link |
Yeah, I think it is a qualitative difference
link |
in all other evolutionary steps.
link |
I can try to describe that if you'd like.
link |
Sure, in which way?
link |
Yeah, I can tell you how.
link |
Pretty much, let's start with a little preface.
link |
Many of the things that humans are able to do
link |
do not have obvious survival advantages precedent.
link |
We could create music, is that,
link |
is there a really survival advantage to that?
link |
What about mathematics?
link |
Is there a real survival advantage to mathematics?
link |
Well, you could stretch it.
link |
You can try to figure these things out, right?
link |
But most of evolutionary history,
link |
everything had immediate survival advantages to it.
link |
So, I'll tell you a story, which I like,
link |
may or may not be true, but the story goes as follows.
link |
Organisms have been evolving for,
link |
since the beginning of life here on Earth,
link |
and adding this sort of complexity onto that,
link |
and this sort of complexity onto that,
link |
and the brain itself is evolved this way.
link |
In fact, there's old parts, and older parts,
link |
and older, older parts of the brain
link |
that kind of just keeps calming on new things,
link |
and we keep adding capabilities.
link |
When we got to the neocortex,
link |
initially it had a very clear survival advantage
link |
in that it produced better vision,
link |
and better hearing, and better touch,
link |
and maybe, and so on.
link |
But what I think happens is that evolution discovered,
link |
it took a mechanism, and this is in our recent theories,
link |
but it took a mechanism evolved a long time ago
link |
for navigating in the world, for knowing where you are.
link |
These are the so called grid cells and place cells
link |
of an old part of the brain.
link |
And it took that mechanism for building maps of the world,
link |
and knowing where you are on those maps,
link |
and how to navigate those maps,
link |
and turns it into a sort of a slimmed down,
link |
idealized version of it.
link |
And that idealized version could now apply
link |
to building maps of other things.
link |
Maps of coffee cups, and maps of phones,
link |
maps of mathematics.
link |
Concepts, yes, and not just almost, exactly.
link |
And so, and it just started replicating this stuff, right?
link |
You just think more, and more, and more.
link |
So we went from being sort of dedicated purpose
link |
neural hardware to solve certain problems
link |
that are important to survival,
link |
to a general purpose neural hardware
link |
that could be applied to all problems.
link |
And now it's escaped the orbit of survival.
link |
We are now able to apply it to things
link |
which we find enjoyment,
link |
but aren't really clearly survival characteristics.
link |
And that it seems to only have happened in humans,
link |
to the large extent.
link |
And so that's what's going on,
link |
where we sort of have,
link |
we've sort of escaped the gravity of evolutionary pressure,
link |
in some sense, in the neocortex.
link |
And it now does things which are not,
link |
that are really interesting,
link |
discovering models of the universe,
link |
which may not really help us.
link |
How does it help us surviving,
link |
knowing that there might be multiverses,
link |
or that there might be the age of the universe,
link |
or how do various stellar things occur?
link |
It doesn't really help us survive at all.
link |
But we enjoy it, and that's what happened.
link |
Or at least not in the obvious way, perhaps.
link |
if you look at the entire universe in an evolutionary way,
link |
it's required for us to do interplanetary travel,
link |
and therefore survive past our own sun.
link |
But you know, let's not get too.
link |
Yeah, but evolution works at one time frame,
link |
it's survival, if you think of survival of the phenotype,
link |
survival of the individual.
link |
What you're talking about there is spans well beyond that.
link |
So there's no genetic,
link |
I'm not transferring any genetic traits to my children
link |
that are gonna help them survive better on Mars.
link |
Totally different mechanism, that's right.
link |
So let's get into the new, as you've mentioned,
link |
this idea of the, I don't know if you have a nice name,
link |
We call it the thousand brain theory of intelligence.
link |
Can you talk about this idea of a spatial view of concepts
link |
Yeah, so can I just describe sort of the,
link |
there's an underlying core discovery,
link |
which then everything comes from that.
link |
That's a very simple, this is really what happened.
link |
We were deep into problems about understanding
link |
how we build models of stuff in the world
link |
and how we make predictions about things.
link |
And I was holding a coffee cup just like this in my hand.
link |
And my finger was touching the side, my index finger.
link |
And then I moved it to the top
link |
and I was gonna feel the rim at the top of the cup.
link |
And I asked myself a very simple question.
link |
I said, well, first of all, I say,
link |
I know that my brain predicts what it's gonna feel
link |
before it touches it.
link |
You can just think about it and imagine it.
link |
And so we know that the brain's making predictions
link |
So the question is, what does it take to predict that?
link |
And there's a very interesting answer.
link |
First of all, it says the brain has to know
link |
it's touching a coffee cup.
link |
It has to have a model of a coffee cup.
link |
It needs to know where the finger currently is
link |
on the cup relative to the cup.
link |
Because when I make a movement,
link |
it needs to know where it's going to be on the cup
link |
after the movement is completed relative to the cup.
link |
And then it can make a prediction
link |
about what it's gonna sense.
link |
So this told me that the neocortex,
link |
which is making this prediction,
link |
needs to know that it's sensing it's touching a cup.
link |
And it needs to know the location of my finger
link |
relative to that cup in a reference frame of the cup.
link |
It doesn't matter where the cup is relative to my body.
link |
It doesn't matter its orientation.
link |
None of that matters.
link |
It's where my finger is relative to the cup,
link |
which tells me then that the neocortex
link |
has a reference frame that's anchored to the cup.
link |
Because otherwise I wouldn't be able to say the location
link |
and I wouldn't be able to predict my new location.
link |
And then we quickly, very instantly can say,
link |
well, every part of my skin could touch this cup.
link |
And therefore every part of my skin is making predictions
link |
and every part of my skin must have a reference frame
link |
that it's using to make predictions.
link |
So the big idea is that throughout the neocortex,
link |
there are, everything is being stored
link |
and referenced in reference frames.
link |
You can think of them like XYZ reference frames,
link |
but they're not like that.
link |
We know a lot about the neural mechanisms for this,
link |
but the brain thinks in reference frames.
link |
And as an engineer, if you're an engineer,
link |
this is not surprising.
link |
You'd say, if I wanted to build a CAD model
link |
of the coffee cup, well, I would bring it up
link |
and some CAD software, and I would assign
link |
some reference frame and say this features
link |
at this locations and so on.
link |
But the fact that this, the idea that this is occurring
link |
throughout the neocortex everywhere, it was a novel idea.
link |
And then a zillion things fell into place after that,
link |
So now we think about the neocortex
link |
as processing information quite differently
link |
than we used to do it.
link |
We used to think about the neocortex
link |
as processing sensory data and extracting features
link |
from that sensory data and then extracting features
link |
from the features, very much like a deep learning network
link |
But that's not how the brain works at all.
link |
The brain works by assigning everything,
link |
every input, everything to reference frames.
link |
And there are thousands, hundreds of thousands
link |
of them active at once in your neocortex.
link |
It's a surprising thing to think about,
link |
but once you sort of internalize this,
link |
you understand that it explains almost every,
link |
almost all the mysteries we've had about this structure.
link |
So one of the consequences of that
link |
is that every small part of the neocortex,
link |
say a millimeter square, and there's 150,000 of those.
link |
So it's about 150,000 square millimeters.
link |
If you take every little square millimeter of the cortex,
link |
it's got some input coming into it
link |
and it's gonna have reference frames
link |
where it's assigned that input to.
link |
And each square millimeter can learn
link |
complete models of objects.
link |
So what do I mean by that?
link |
If I'm touching the coffee cup,
link |
well, if I just touch it in one place,
link |
I can't learn what this coffee cup is
link |
because I'm just feeling one part.
link |
But if I move it around the cup
link |
and touch it at different areas,
link |
I can build up a complete model of the cup
link |
because I'm now filling in that three dimensional map,
link |
which is the coffee cup.
link |
I can say, oh, what am I feeling
link |
at all these different locations?
link |
That's the basic idea, it's more complicated than that.
link |
But so through time, and we talked about time earlier,
link |
through time, even a single column,
link |
which is only looking at, or a single part of the cortex,
link |
which is only looking at a small part of the world,
link |
can build up a complete model of an object.
link |
And so if you think about the part of the brain,
link |
which is getting input from all my fingers,
link |
so they're spread across the top of your head here.
link |
This is the somatosensory cortex.
link |
There's columns associated
link |
with all the different areas of my skin.
link |
And what we believe is happening
link |
is that all of them are building models of this cup,
link |
every one of them, or things.
link |
They're not all building,
link |
not every column or every part of the cortex
link |
builds models of everything,
link |
but they're all building models of something.
link |
And so you have, so when I touch this cup with my hand,
link |
there are multiple models of the cup being invoked.
link |
If I look at it with my eyes,
link |
there are, again, many models of the cup being invoked,
link |
because each part of the visual system,
link |
the brain doesn't process an image.
link |
That's a misleading idea.
link |
It's just like your fingers touching the cup,
link |
so different parts of my retina
link |
are looking at different parts of the cup.
link |
And thousands and thousands of models of the cup
link |
are being invoked at once.
link |
And they're all voting with each other,
link |
trying to figure out what's going on.
link |
So that's why we call it the thousand brains theory
link |
of intelligence, because there isn't one model of a cup.
link |
There are thousands of models of this cup.
link |
There are thousands of models of your cellphone
link |
and about cameras and microphones and so on.
link |
It's a distributed modeling system,
link |
which is very different
link |
than the way people have thought about it.
link |
And so that's a really compelling and interesting idea.
link |
I have two first questions.
link |
So one, on the ensemble part of everything coming together,
link |
you have these thousand brains.
link |
How do you know which one has done the best job
link |
Let me try to explain it.
link |
There's a problem that's known in neuroscience
link |
called the sensor fusion problem.
link |
And so the idea is there's something like,
link |
oh, the image comes from the eye.
link |
There's a picture on the retina
link |
and then it gets projected to the neocortex.
link |
Oh, by now it's all spread out all over the place
link |
and it's kind of squirrely and distorted
link |
and pieces are all over the...
link |
It doesn't look like a picture anymore.
link |
When does it all come back together again?
link |
Or you might say, well, yes,
link |
but I also have sounds or touches associated with the cup.
link |
So I'm seeing the cup and touching the cup.
link |
How do they get combined together again?
link |
So it's called the sensor fusion problem.
link |
As if all these disparate parts
link |
have to be brought together into one model someplace.
link |
That's the wrong idea.
link |
The right idea is that you've got all these guys voting.
link |
There's auditory models of the cup.
link |
There's visual models of the cup.
link |
There's tactile models of the cup.
link |
In the vision system,
link |
there might be ones that are more focused on black and white
link |
and ones focusing on color.
link |
It doesn't really matter.
link |
There's just thousands and thousands of models of this cup.
link |
They don't actually come together in one spot.
link |
Just literally think of it this way.
link |
Imagine you have these columns
link |
that are like about the size of a little piece of spaghetti.
link |
Like a two and a half millimeters tall
link |
and about a millimeter in wide.
link |
They're not physical, but you could think of them that way.
link |
And each one's trying to guess what this thing is
link |
Now, they can do a pretty good job
link |
if they're allowed to move over time.
link |
So I can reach my hand into a black box
link |
and move my finger around an object.
link |
And if I touch enough spaces, I go, okay,
link |
now I know what it is.
link |
But often we don't do that.
link |
Often I can just reach and grab something with my hand
link |
all at once and I get it.
link |
Or if I had to look through the world through a straw,
link |
so I'm only invoking one little column,
link |
I can only see part of something
link |
because I have to move the straw around.
link |
But if I open my eyes, I see the whole thing at once.
link |
So what we think is going on
link |
is all these little pieces of spaghetti,
link |
if you will, all these little columns in the cortex,
link |
are all trying to guess what it is that they're sensing.
link |
They'll do a better guess if they have time
link |
and can move over time.
link |
So if I move my eyes, I move my fingers.
link |
But if they don't, they have a poor guess.
link |
It's a probabilistic guess of what they might be touching.
link |
Now, imagine they can post their probability
link |
at the top of a little piece of spaghetti.
link |
Each one of them says,
link |
I think, and it's not really a probability distribution.
link |
It's more like a set of possibilities.
link |
In the brain, it doesn't work as a probability distribution.
link |
It works as more like what we call a union.
link |
So you could say, and one column says,
link |
I think it could be a coffee cup,
link |
a soda can, or a water bottle.
link |
And another column says,
link |
I think it could be a coffee cup
link |
or a telephone or a camera or whatever, right?
link |
And all these guys are saying what they think it might be.
link |
And there's these long range connections
link |
in certain layers in the cortex.
link |
So there's in some layers in some cells types
link |
in each column, send the projections across the brain.
link |
And that's the voting occurs.
link |
And so there's a simple associative memory mechanism.
link |
We've described this in a recent paper
link |
and we've modeled this that says,
link |
they can all quickly settle on the only
link |
or the one best answer for all of them.
link |
If there is a single best answer,
link |
they all vote and say, yep, it's gotta be the coffee cup.
link |
And at that point, they all know it's a coffee cup.
link |
And at that point, everyone acts as if it's a coffee cup.
link |
They're like, yep, we know it's a coffee,
link |
even though I've only seen one little piece of this world,
link |
I know it's a coffee cup I'm touching
link |
or I'm seeing or whatever.
link |
And so you can think of all these columns
link |
are looking at different parts in different places,
link |
different sensory input, different locations,
link |
they're all different.
link |
But this layer that's doing the voting, it solidifies.
link |
It's just like it crystallizes and says,
link |
oh, we all know what we're doing.
link |
And so you don't bring these models together in one model,
link |
you just vote and there's a crystallization of the vote.
link |
Great, that's at least a compelling way
link |
to think about the way you form a model of the world.
link |
Now, you talk about a coffee cup.
link |
Do you see this, as far as I understand,
link |
you are proposing this as well,
link |
that this extends to much more than coffee cups?
link |
Or at least the physical world,
link |
it expands to the world of concepts.
link |
And well, first, the primary thing is evidence for that
link |
is that the regions of the neocortex
link |
that are associated with language
link |
or high level thought or mathematics
link |
or things like that,
link |
they look like the regions of the neocortex
link |
that process vision, hearing, and touch.
link |
They don't look any different.
link |
Or they look only marginally different.
link |
And so one would say, well, if Vernon Mountcastle,
link |
who proposed that all the parts of the neocortex
link |
do the same thing, if he's right,
link |
then the parts that are doing language
link |
or mathematics or physics
link |
are working on the same principle.
link |
They must be working on the principle of reference frames.
link |
So that's a little odd thought.
link |
But of course, we had no prior idea
link |
how these things happen.
link |
So let's go with that.
link |
And we, in our recent paper,
link |
we talked a little bit about that.
link |
I've been working on it more since.
link |
I have better ideas about it now.
link |
I'm sitting here very confident
link |
that that's what's happening.
link |
And I can give you some examples
link |
that help you think about that.
link |
It's not we understand it completely,
link |
but I understand it better than I've described it
link |
in any paper so far.
link |
So, but we did put that idea out there.
link |
It says, okay, this is,
link |
it's a good place to start, you know?
link |
And the evidence would suggest it's how it's happening.
link |
And then we can start tackling that problem
link |
one piece at a time.
link |
Like, what does it mean to do high level thought?
link |
What does it mean to do language?
link |
How would that fit into a reference frame framework?
link |
Yeah, so there's a,
link |
I don't know if you could tell me if there's a connection,
link |
but there's an app called Anki
link |
that helps you remember different concepts.
link |
And they talk about like a memory palace
link |
that helps you remember completely random concepts
link |
by trying to put them in a physical space in your mind
link |
and putting them next to each other.
link |
It's called the method of loci.
link |
For some reason, that seems to work really well.
link |
Now, that's a very narrow kind of application
link |
of just remembering some facts.
link |
But that's a very, very telling one.
link |
So this seems like you're describing a mechanism
link |
why this seems to work.
link |
So basically the way what we think is going on
link |
is all things you know, all concepts, all ideas,
link |
words, everything you know are stored in reference frames.
link |
And so if you want to remember something,
link |
you have to basically navigate through a reference frame
link |
the same way a rat navigates through a maze
link |
and the same way my finger navigates to this coffee cup.
link |
You are moving through some space.
link |
And so if you have a random list of things
link |
you were asked to remember,
link |
by assigning them to a reference frame,
link |
you've already know very well to see your house, right?
link |
And the idea of the method of loci is you can say,
link |
okay, in my lobby, I'm going to put this thing.
link |
And then the bedroom, I put this one.
link |
I go down the hall, I put this thing.
link |
And then you want to recall those facts
link |
or recall those things.
link |
You just walk mentally, you walk through your house.
link |
You're mentally moving through a reference frame
link |
that you already had.
link |
And that tells you,
link |
there's two things that are really important about that.
link |
It tells us the brain prefers to store things
link |
in reference frames.
link |
And that the method of recalling things
link |
or thinking, if you will,
link |
is to move mentally through those reference frames.
link |
You could move physically through some reference frames,
link |
like I could physically move through the reference frame
link |
of this coffee cup.
link |
I can also mentally move through the reference frame
link |
of the coffee cup, imagining me touching it.
link |
But I can also mentally move my house.
link |
And so now we can ask yourself,
link |
or are all concepts stored this way?
link |
There was some recent research using human subjects
link |
in fMRI, and I'm going to apologize for not knowing
link |
the name of the scientists who did this.
link |
But what they did is they put humans in this fMRI machine,
link |
which is one of these imaging machines.
link |
And they gave the humans tasks to think about birds.
link |
So they had different types of birds,
link |
and birds that look big and small,
link |
and long necks and long legs, things like that.
link |
And what they could tell from the fMRI
link |
was a very clever experiment.
link |
You get to tell when humans were thinking about the birds,
link |
that the birds, the knowledge of birds
link |
was arranged in a reference frame,
link |
similar to the ones that are used
link |
when you navigate in a room.
link |
That these are called grid cells,
link |
and there are grid cell like patterns of activity
link |
in the neocortex when they do this.
link |
So it's a very clever experiment.
link |
And what it basically says,
link |
that even when you're thinking about something abstract,
link |
and you're not really thinking about it as a reference frame,
link |
it tells us the brain is actually using a reference frame.
link |
And it's using the same neural mechanisms.
link |
These grid cells are the basic same neural mechanism
link |
that we propose that grid cells,
link |
which exist in the old part of the brain,
link |
the entorhinal cortex, that that mechanism
link |
is now similar mechanism is used throughout the neocortex.
link |
It's the same nature to preserve this interesting way
link |
of creating reference frames.
link |
And so now they have empirical evidence
link |
that when you think about concepts like birds,
link |
that you're using reference frames
link |
that are built on grid cells.
link |
So that's similar to the method of loci,
link |
but in this case, the birds are related.
link |
So they create their own reference frame,
link |
which is consistent with bird space.
link |
And when you think about something, you go through that.
link |
You can make the same example,
link |
let's take mathematics.
link |
Let's say you wanna prove a conjecture.
link |
What is a conjecture?
link |
A conjecture is a statement you believe to be true,
link |
but you haven't proven it.
link |
And so it might be an equation.
link |
I wanna show that this is equal to that.
link |
And you have some places you start with.
link |
You say, well, I know this is true,
link |
and I know this is true.
link |
And I think that maybe to get to the final proof,
link |
I need to go through some intermediate results.
link |
What I believe is happening is literally these equations
link |
or these points are assigned to a reference frame,
link |
a mathematical reference frame.
link |
And when you do mathematical operations,
link |
a simple one might be multiply or divide,
link |
but you might be a little plus transform or something else.
link |
That is like a movement in the reference frame of the math.
link |
And so you're literally trying to discover a path
link |
from one location to another location
link |
in a space of mathematics.
link |
And if you can get to these intermediate results,
link |
then you know your map is pretty good,
link |
and you know you're using the right operations.
link |
Much of what we think about is solving hard problems
link |
is designing the correct reference frame for that problem,
link |
figuring out how to organize the information
link |
and what behaviors I wanna use in that space
link |
Yeah, so if you dig in an idea of this reference frame,
link |
whether it's the math, you start a set of axioms
link |
to try to get to proving the conjecture.
link |
Can you try to describe, maybe take a step back,
link |
how you think of the reference frame in that context?
link |
Is it the reference frame that the axioms are happy in?
link |
Is it the reference frame that might contain everything?
link |
Is it a changing thing as you?
link |
You have many, many reference frames.
link |
I mean, in fact, the way the theory,
link |
the thousand brain theory of intelligence says
link |
that every single thing in the world
link |
has its own reference frame.
link |
So every word has its own reference frames.
link |
And we can talk about this.
link |
The mathematics work out,
link |
this is no problem for neurons to do this.
link |
But how many reference frames does a coffee cup have?
link |
Well, it's on a table.
link |
Let's say you ask how many reference frames
link |
could a column in my finger
link |
that's touching the coffee cup have?
link |
Because there are many, many copy,
link |
there are many, many models of the coffee cup.
link |
So the coffee, there is no one model of a coffee cup.
link |
There are many models of a coffee cup.
link |
And you could say, well,
link |
how many different things can my finger learn?
link |
Is this the question you want to ask?
link |
Imagine I say every concept, every idea,
link |
everything you've ever know about that you can say,
link |
I know that thing has a reference frame
link |
associated with it.
link |
And what we do when we build composite objects,
link |
we assign reference frames
link |
to point another reference frame.
link |
So my coffee cup has multiple components to it.
link |
It's got a limb, it's got a cylinder, it's got a handle.
link |
And those things have their own reference frames
link |
and they're assigned to a master reference frame,
link |
which is called this cup.
link |
And now I have this Numenta logo on it.
link |
Well, that's something that exists elsewhere in the world.
link |
It's its own thing.
link |
So it has its own reference frame.
link |
So we now have to say,
link |
how can I assign the Numenta logo reference frame
link |
onto the cylinder or onto the coffee cup?
link |
So it's all, we talked about this in the paper
link |
that came out in December of this last year.
link |
The idea of how you can assign reference frames
link |
to reference frames, how neurons could do this.
link |
So, well, my question is,
link |
even though you mentioned reference frames a lot,
link |
I almost feel it's really useful to dig into
link |
how you think of what a reference frame is.
link |
I mean, it was already helpful for me to understand
link |
that you think of reference frames
link |
as something there is a lot of.
link |
Okay, so let's just say that we're gonna have
link |
some neurons in the brain, not many, actually,
link |
10,000, 20,000 are gonna create
link |
a whole bunch of reference frames.
link |
What does it mean?
link |
What is a reference frame?
link |
First of all, these reference frames are different
link |
than the ones you might be used to.
link |
We know lots of reference frames.
link |
For example, we know the Cartesian coordinates, X, Y, Z,
link |
that's a type of reference frame.
link |
We know longitude and latitude,
link |
that's a different type of reference frame.
link |
If I look at a printed map,
link |
you might have columns A through M,
link |
and rows one through 20,
link |
that's a different type of reference frame.
link |
It's kind of a Cartesian coordinate reference frame.
link |
The interesting thing about the reference frames
link |
in the brain, and we know this because these
link |
have been established through neuroscience
link |
studying the entorhinal cortex.
link |
So I'm not speculating here, okay?
link |
This is known neuroscience in an old part of the brain.
link |
The way these cells create reference frames,
link |
they have no origin.
link |
So what it's more like, you have a point,
link |
a point in some space, and you,
link |
given a particular movement,
link |
you can then tell what the next point should be.
link |
And you can then tell what the next point would be,
link |
You can use this to calculate
link |
how to get from one point to another.
link |
So how do I get from my house to my home,
link |
or how do I get my finger from the side of my cup
link |
to the top of the cup?
link |
How do I get from the axioms to the conjecture?
link |
So it's a different type of reference frame,
link |
and I can, if you want, I can describe in more detail,
link |
I can paint a picture of how you might want
link |
to think about that.
link |
It's really helpful to think it's something
link |
you can move through, but is there,
link |
is it helpful to think of it as spatial in some sense,
link |
or is there something that's more?
link |
No, it's definitely spatial.
link |
It's spatial in a mathematical sense.
link |
How many dimensions?
link |
Can it be a crazy number of dimensions?
link |
Well, that's an interesting question.
link |
In the old part of the brain, the entorhinal cortex,
link |
they studied rats, and initially it looks like,
link |
oh, this is just two dimensional.
link |
It's like the rat is in some box in the maze or whatever,
link |
and they know where the rat is using
link |
these two dimensional reference frames
link |
to know where it is in the maze.
link |
We said, well, okay, but what about bats?
link |
That's a mammal, and they fly in three dimensional space.
link |
How do they do that?
link |
They seem to know where they are, right?
link |
So this is a current area of active research,
link |
and it seems like somehow the neurons
link |
in the entorhinal cortex can learn three dimensional space.
link |
We just, two members of our team,
link |
along with Elif Fett from MIT,
link |
just released a paper this literally last week.
link |
It's on bioRxiv, where they show that you can,
link |
if you, the way these things work,
link |
and I won't get, unless you want to,
link |
I won't get into the detail,
link |
but grid cells can represent any n dimensional space.
link |
It's not inherently limited.
link |
You can think of it this way.
link |
If you had two dimensional, the way it works
link |
is you had a bunch of two dimensional slices.
link |
That's the way these things work.
link |
There's a whole bunch of two dimensional models,
link |
and you can just, you can slice up
link |
any n dimensional space with two dimensional projections.
link |
So, and you could have one dimensional models.
link |
So there's nothing inherent about the mathematics
link |
about the way the neurons do this,
link |
which constrain the dimensionality of the space,
link |
which I think was important.
link |
So obviously I have a three dimensional map of this cup.
link |
Maybe it's even more than that, I don't know.
link |
But it's clearly a three dimensional map of the cup.
link |
I don't just have a projection of the cup.
link |
But when I think about birds,
link |
or when I think about mathematics,
link |
perhaps it's more than three dimensions.
link |
So in terms of each individual column
link |
building up more and more information over time,
link |
do you think that mechanism is well understood?
link |
In your mind, you've proposed a lot of architectures there.
link |
Is that a key piece, or is it,
link |
is the big piece, the thousand brain theory of intelligence,
link |
the ensemble of it all?
link |
Well, I think they're both big.
link |
I mean, clearly the concept, as a theorist,
link |
the concept is most exciting, right?
link |
The high level concept.
link |
The high level concept.
link |
This is a totally new way of thinking
link |
about how the neocortex works.
link |
So that is appealing.
link |
It has all these ramifications.
link |
And with that, as a framework for how the brain works,
link |
you can make all kinds of predictions
link |
and solve all kinds of problems.
link |
Now we're trying to work through
link |
many of these details right now.
link |
Okay, how do the neurons actually do this?
link |
Well, it turns out, if you think about grid cells
link |
and place cells in the old parts of the brain,
link |
there's a lot that's known about them,
link |
but there's still some mysteries.
link |
There's a lot of debate about exactly the details,
link |
how these work and what are the signs.
link |
And we have that still, that same level of detail,
link |
that same level of concern.
link |
What we spend here most of our time doing
link |
is trying to make a very good list
link |
of the things we don't understand yet.
link |
That's the key part here.
link |
What are the constraints?
link |
It's not like, oh, this thing seems to work, we're done.
link |
No, it's like, okay, it kind of works,
link |
but these are other things we know it has to do
link |
and it's not doing those yet.
link |
I would say we're well on the way here.
link |
We're not done yet.
link |
There's a lot of trickiness to this system,
link |
but the basic principles about how different layers
link |
in the neocortex are doing much of this, we understand.
link |
But there's some fundamental parts
link |
that we don't understand as well.
link |
So what would you say is one of the harder open problems
link |
or one of the ones that have been bothering you,
link |
keeping you up at night the most?
link |
Oh, well, right now, this is a detailed thing
link |
that wouldn't apply to most people, okay?
link |
But you want me to answer that question?
link |
We've talked about as if, oh,
link |
to predict what you're going to sense on this coffee cup,
link |
I need to know where my finger is gonna be
link |
on the coffee cup.
link |
That is true, but it's insufficient.
link |
Think about my finger touches the edge of the coffee cup.
link |
My finger can touch it at different orientations.
link |
I can rotate my finger around here and that doesn't change.
link |
I can make that prediction and somehow,
link |
so it's not just the location.
link |
There's an orientation component of this as well.
link |
This is known in the old parts of the brain too.
link |
There's things called head direction cells,
link |
which way the rat is facing.
link |
It's the same kind of basic idea.
link |
So if my finger were a rat, you know, in three dimensions,
link |
I have a three dimensional orientation
link |
and I have a three dimensional location.
link |
If I was a rat, I would have a,
link |
you might think of it as a two dimensional location,
link |
a two dimensional orientation,
link |
a one dimensional orientation,
link |
like just which way is it facing?
link |
So how the two components work together,
link |
how it is that I combine orientation,
link |
the orientation of my sensor,
link |
as well as the location is a tricky problem.
link |
And I think I've made progress on it.
link |
So at a bigger version of that,
link |
so perspective is super interesting, but super specific.
link |
Yeah, I warned you.
link |
No, no, no, that's really good,
link |
but there's a more general version of that.
link |
Do you think context matters,
link |
the fact that we're in a building in North America,
link |
that we, in the day and age where we have mugs?
link |
I mean, there's all this extra information
link |
that you bring to the table about everything else
link |
in the room that's outside of just the coffee cup.
link |
How does it get connected, do you think?
link |
Yeah, and that is another really interesting question.
link |
I'm gonna throw that under the rubric
link |
or the name of attentional problems.
link |
First of all, we have this model,
link |
I have many, many models.
link |
And also the question, does it matter?
link |
Well, it matters for certain things, of course it does.
link |
Maybe what we think of that as a coffee cup
link |
in another part of the world
link |
is viewed as something completely different.
link |
Or maybe our logo, which is very benign
link |
in this part of the world,
link |
it means something very different
link |
in another part of the world.
link |
So those things do matter.
link |
I think the way to think about it is the following,
link |
one way to think about it,
link |
is we have all these models of the world, okay?
link |
And we model everything.
link |
And as I said earlier, I kind of snuck it in there,
link |
our models are actually, we build composite structure.
link |
So every object is composed of other objects,
link |
which are composed of other objects,
link |
and they become members of other objects.
link |
So this room has chairs and a table and a room
link |
and walls and so on.
link |
Now we can just arrange these things in a certain way
link |
and go, oh, that's the nomenclature conference room.
link |
So, and what we do is when we go around the world
link |
and we experience the world,
link |
by walking into a room, for example,
link |
the first thing I do is I can say,
link |
oh, I'm in this room, do I recognize the room?
link |
Then I can say, oh, look, there's a table here.
link |
And by attending to the table,
link |
I'm then assigning this table in the context of the room.
link |
Then I can say, oh, on the table, there's a coffee cup.
link |
Oh, and on the table, there's a logo.
link |
And in the logo, there's the word Nementa.
link |
Oh, and look in the logo, there's the letter E.
link |
Oh, and look, it has an unusual serif.
link |
And it doesn't actually, but I pretended to serif.
link |
So the point is your attention is kind of drilling
link |
deep in and out of these nested structures.
link |
And I can pop back up and I can pop back down.
link |
I can pop back up and I can pop back down.
link |
So when I attend to the coffee cup,
link |
I haven't lost the context of everything else,
link |
but it's sort of, there's this sort of nested structure.
link |
So the attention filters the reference frame information
link |
for that particular period of time?
link |
Yes, it basically, moment to moment,
link |
you attend the sub components,
link |
and then you can attend the sub components
link |
to sub components.
link |
And you can move up and down.
link |
You can move up and down.
link |
We do that all the time.
link |
You're not even, now that I'm aware of it,
link |
I'm very conscious of it.
link |
But until, but most people don't even think about this.
link |
You just walk in a room and you don't say,
link |
oh, I looked at the chair and I looked at the board
link |
and looked at that word on the board
link |
and I looked over here, what's going on, right?
link |
So what percent of your day are you deeply aware of this?
link |
And what part can you actually relax and just be Jeff?
link |
Me personally, like my personal day?
link |
Unfortunately, I'm afflicted with too much of the former.
link |
Well, unfortunately or unfortunately.
link |
You don't think it's useful?
link |
Oh, it is useful, totally useful.
link |
I think about this stuff almost all the time.
link |
And one of my primary ways of thinking
link |
is when I'm in sleep at night,
link |
I always wake up in the middle of the night.
link |
And then I stay awake for at least an hour
link |
with my eyes shut in sort of a half sleep state
link |
thinking about these things.
link |
I come up with answers to problems very often
link |
in that sort of half sleeping state.
link |
I think about it on my bike ride, I think about it on walks.
link |
I'm just constantly thinking about this.
link |
I have to almost schedule time
link |
to not think about this stuff
link |
because it's very, it's mentally taxing.
link |
Are you, when you're thinking about this stuff,
link |
are you thinking introspectively,
link |
like almost taking a step outside of yourself
link |
and trying to figure out what is your mind doing right now?
link |
I do that all the time, but that's not all I do.
link |
I'm constantly observing myself.
link |
So as soon as I started thinking about grid cells,
link |
for example, and getting into that,
link |
I started saying, oh, well, grid cells
link |
can have my place of sense in the world.
link |
That's where you know where you are.
link |
And it's interesting, we always have a sense
link |
of where we are unless we're lost.
link |
And so I started at night when I got up
link |
to go to the bathroom, I would start trying to do it
link |
completely with my eyes closed all the time.
link |
And I would test my sense of grid cells.
link |
I would walk five feet and say, okay, I think I'm here.
link |
Am I really there?
link |
And then I would calculate my error again
link |
and see how the errors could accumulate.
link |
So even something as simple as getting up
link |
in the middle of the night to go to the bathroom,
link |
I'm testing these theories out.
link |
I mean, the coffee cup is an example of that too.
link |
So I find that these sort of everyday introspections
link |
are actually quite helpful.
link |
It doesn't mean you can ignore the science.
link |
I mean, I spend hours every day
link |
reading ridiculously complex papers.
link |
That's not nearly as much fun,
link |
but you have to sort of build up those constraints
link |
and the knowledge about the field and who's doing what
link |
and what exactly they think is happening here.
link |
And then you can sit back and say,
link |
okay, let's try to piece this all together.
link |
Let's come up with some, I'm very,
link |
in this group here, people, they know they do,
link |
I do this all the time.
link |
I come in with these introspective ideas and say,
link |
well, have you ever thought about this?
link |
Now watch, well, let's all do this together.
link |
It's not, as long as you don't,
link |
all you did was that, then you're just making up stuff.
link |
But if you're constraining it by the reality
link |
of the neuroscience, then it's really helpful.
link |
So let's talk a little bit about deep learning
link |
and the successes in the applied space of neural networks,
link |
ideas of training model on data
link |
and these simple computational units,
link |
artificial neurons that with backpropagation,
link |
statistical ways of being able to generalize
link |
from the training set onto data
link |
that's similar to that training set.
link |
So where do you think are the limitations
link |
of those approaches?
link |
What do you think are its strengths
link |
relative to your major efforts
link |
of constructing a theory of human intelligence?
link |
Well, I'm not an expert in this field.
link |
I'm somewhat knowledgeable.
link |
Some of it is in just your intuition.
link |
Well, I have a little bit more than intuition,
link |
but I just want to say like,
link |
you know, one of the things that you asked me,
link |
do I spend all my time thinking about neuroscience?
link |
That's to the exclusion of thinking about things
link |
like convolutional neural networks.
link |
But I try to stay current.
link |
So look, I think it's great, the progress they've made.
link |
And as I mentioned earlier,
link |
it's very highly useful for many things.
link |
The models that we have today are actually derived
link |
from a lot of neuroscience principles.
link |
There are distributed processing systems
link |
and distributed memory systems,
link |
and that's how the brain works.
link |
They use things that we might call them neurons,
link |
but they're really not neurons at all.
link |
So we can just, they're not really neurons.
link |
So they're distributed processing systems.
link |
And that nature of hierarchy,
link |
that came also from neuroscience.
link |
And so there's a lot of things,
link |
the learning rules, basically,
link |
not back prop, but other, you know,
link |
sort of heavy on top of that.
link |
I'd be curious to say they're not neurons at all.
link |
Can you describe in which way?
link |
I mean, some of it is obvious,
link |
but I'd be curious if you have specific ways
link |
in which you think are the biggest differences.
link |
Yeah, we had a paper in 2016 called
link |
Why Neurons Have Thousands of Synapses.
link |
And if you read that paper,
link |
you'll know what I'm talking about here.
link |
A real neuron in the brain is a complex thing.
link |
And let's just start with the synapses on it,
link |
which is a connection between neurons.
link |
Real neurons can have everywhere
link |
from five to 30,000 synapses on them.
link |
The ones near the cell body,
link |
the ones that are close to the soma of the cell body,
link |
those are like the ones that people model
link |
in artificial neurons.
link |
There is a few hundred of those.
link |
Maybe they can affect the cell.
link |
They can make the cell become active.
link |
95% of the synapses can't do that.
link |
They're too far away.
link |
So if you activate one of those synapses,
link |
it just doesn't affect the cell body enough
link |
to make any difference.
link |
Any one of them individually.
link |
Any one of them individually,
link |
or even if you do a mass of them.
link |
What real neurons do is the following.
link |
If you activate or you get 10 to 20 of them
link |
active at the same time,
link |
meaning they're all receiving an input at the same time,
link |
and those 10 to 20 synapses or 40 synapses
link |
within a very short distance on the dendrite,
link |
like 40 microns, a very small area.
link |
So if you activate a bunch of these
link |
right next to each other at some distant place,
link |
what happens is it creates
link |
what's called the dendritic spike.
link |
And the dendritic spike travels through the dendrites
link |
and can reach the soma or the cell body.
link |
Now, when it gets there, it changes the voltage,
link |
which is sort of like gonna make the cell fire,
link |
but never enough to make the cell fire.
link |
It's sort of what we call, it says we depolarize the cell,
link |
you raise the voltage a little bit,
link |
but not enough to do anything.
link |
It's like, well, what good is that?
link |
And then it goes back down again.
link |
So we propose a theory,
link |
which I'm very confident in basics are,
link |
is that what's happening there is
link |
those 95% of the synapses are recognizing
link |
dozens to hundreds of unique patterns.
link |
They can write about 10, 20 synapses at a time,
link |
and they're acting like predictions.
link |
So the neuron actually is a predictive engine on its own.
link |
It can fire when it gets enough,
link |
what they call proximal input
link |
from those ones near the cell fire,
link |
but it can get ready to fire from dozens to hundreds
link |
of patterns that it recognizes from the other guys.
link |
And the advantage of this to the neuron
link |
is that when it actually does produce a spike
link |
in action potential,
link |
it does so slightly sooner than it would have otherwise.
link |
And so what could is slightly sooner?
link |
Well, the slightly sooner part is it,
link |
all the excitatory neurons in the brain
link |
are surrounded by these inhibitory neurons,
link |
and they're very fast, the inhibitory neurons,
link |
these basket cells.
link |
And if I get my spike out
link |
a little bit sooner than someone else,
link |
I inhibit all my neighbors around me, right?
link |
And what you end up with is a different representation.
link |
You end up with a reputation that matches your prediction.
link |
It's a sparser representation,
link |
meaning fewer neurons are active,
link |
but it's much more specific.
link |
And so we showed how networks of these neurons
link |
can do very sophisticated temporal prediction, basically.
link |
So this, summarize this,
link |
real neurons in the brain are time based prediction engines,
link |
and there's no concept of this at all
link |
in artificial, what we call point neurons.
link |
I don't think you can build a brain without them.
link |
I don't think you can build intelligence without them,
link |
because it's where a large part of the time comes from.
link |
These are predictive models, and the time is,
link |
there's a prior and a prediction and an action,
link |
and it's inherent through every neuron in the neocortex.
link |
So I would say that point neurons sort of model
link |
a piece of that, and not very well at that either.
link |
But like for example, synapses are very unreliable,
link |
and you cannot assign any precision to them.
link |
So even one digit of precision is not possible.
link |
So the way real neurons work is they don't add these,
link |
they don't change these weights accurately
link |
like artificial neural networks do.
link |
They basically form new synapses,
link |
and so what you're trying to always do is
link |
detect the presence of some 10 to 20
link |
active synapses at the same time,
link |
as opposed, and they're almost binary.
link |
It's like, because you can't really represent
link |
anything much finer than that.
link |
So these are the kind of,
link |
and I think that's actually another essential component,
link |
because the brain works on sparse patterns,
link |
and all that mechanism is based on sparse patterns,
link |
and I don't actually think you could build real brains
link |
or machine intelligence without
link |
incorporating some of those ideas.
link |
It's hard to even think about the complexity
link |
that emerges from the fact that
link |
the timing of the firing matters in the brain,
link |
the fact that you form new synapses,
link |
and I mean, everything you just mentioned
link |
in the past couple minutes.
link |
Trust me, if you spend time on it,
link |
you can get your mind around it.
link |
It's not like, it's no longer a mystery to me.
link |
No, but sorry, as a function, in a mathematical way,
link |
can you start getting an intuition about
link |
what gets it excited, what not,
link |
and what kind of representation?
link |
Yeah, it's not as easy as,
link |
there's many other types of neural networks
link |
that are more amenable to pure analysis,
link |
especially very simple networks.
link |
Oh, I have four neurons, and they're doing this.
link |
Can we describe to them mathematically
link |
what they're doing type of thing?
link |
Even the complexity of convolutional neural networks today,
link |
it's sort of a mystery.
link |
They can't really describe the whole system.
link |
And so it's different.
link |
My colleague Subitai Ahmad, he did a nice paper on this.
link |
You can get all this stuff on our website
link |
if you're interested,
link |
talking about sort of the mathematical properties
link |
of sparse representations.
link |
And so what we can do is we can show mathematically,
link |
for example, why 10 to 20 synapses to recognize a pattern
link |
is the correct number, is the right number you'd wanna use.
link |
And by the way, that matches biology.
link |
We can show mathematically some of these concepts
link |
about the show why the brain is so robust
link |
to noise and error and fallout and so on.
link |
We can show that mathematically
link |
as well as empirically in simulations.
link |
But the system can't be analyzed completely.
link |
Any complex system can't, and so that's out of the realm.
link |
But there is mathematical benefits and intuitions
link |
that can be derived from mathematics.
link |
And we try to do that as well.
link |
Most of our papers have a section about that.
link |
So I think it's refreshing and useful for me
link |
to be talking to you about deep neural networks,
link |
because your intuition basically says
link |
that we can't achieve anything like intelligence
link |
with artificial neural networks.
link |
Well, not in the current form.
link |
Not in the current form.
link |
I'm sure we can do it in the ultimate form, sure.
link |
So let me dig into it
link |
and see what your thoughts are there a little bit.
link |
So I'm not sure if you read this little blog post
link |
called Bitter Lesson by Rich Sutton recently.
link |
He's a reinforcement learning pioneer.
link |
I'm not sure if you're familiar with him.
link |
His basic idea is that all the stuff we've done in AI
link |
in the past 70 years, he's one of the old school guys.
link |
The biggest lesson learned is that all the tricky things
link |
we've done, they benefit in the short term,
link |
but in the long term, what wins out
link |
is a simple general method that just relies on Moore's law,
link |
on computation getting faster and faster.
link |
This is what he's saying.
link |
This is what has worked up to now.
link |
This is what has worked up to now.
link |
If you're trying to build a system,
link |
if we're talking about,
link |
he's not concerned about intelligence.
link |
He's concerned about a system that works
link |
in terms of making predictions
link |
on applied narrow AI problems, right?
link |
That's what this discussion is about.
link |
That you just try to go as general as possible
link |
and wait years or decades for the computation
link |
to make it actually.
link |
Is he saying that as a criticism
link |
or is he saying this is a prescription
link |
of what we ought to be doing?
link |
Well, it's very difficult.
link |
He's saying this is what has worked
link |
and yes, a prescription, but it's a difficult prescription
link |
because it says all the fun things
link |
you guys are trying to do, we are trying to do.
link |
He's part of the community.
link |
He's saying it's only going to be short term gains.
link |
So this all leads up to a question, I guess,
link |
on artificial neural networks
link |
and maybe our own biological neural networks
link |
is do you think if we just scale things up significantly,
link |
so take these dumb artificial neurons,
link |
the point neurons, I like that term.
link |
If we just have a lot more of them,
link |
do you think some of the elements
link |
that we see in the brain may start emerging?
link |
No, I don't think so.
link |
We can do bigger problems of the same type.
link |
I mean, it's been pointed out by many people
link |
that today's convolutional neural networks
link |
aren't really much different
link |
than the ones we had quite a while ago.
link |
They're bigger and train more
link |
and we have more labeled data and so on.
link |
But I don't think you can get to the kind of things
link |
I know the brain can do and that we think about
link |
as intelligence by just scaling it up.
link |
So that may be, it's a good description
link |
of what's happened in the past,
link |
what's happened recently with the reemergence
link |
of artificial neural networks.
link |
It may be a good prescription
link |
for what's gonna happen in the short term.
link |
But I don't think that's the path.
link |
I've said that earlier.
link |
There's an alternate path.
link |
I should mention to you, by the way,
link |
that we've made sufficient progress
link |
on the whole cortical theory in the last few years
link |
that last year we decided to start actively pursuing
link |
how do we get these ideas embedded into machine learning?
link |
Well, that's, again, being led by my colleague,
link |
Subed Tariman, and he's more of a machine learning guy.
link |
I'm more of a neuroscience guy.
link |
So this is now, I wouldn't say our focus,
link |
but it is now an equal focus here
link |
because we need to proselytize what we've learned
link |
and we need to show how it's beneficial
link |
to the machine learning layer.
link |
So we're putting, we have a plan in place right now.
link |
In fact, we just did our first paper on this.
link |
I can tell you about that.
link |
But one of the reasons I wanna talk to you
link |
is because I'm trying to get more people
link |
in the machine learning community to say,
link |
I need to learn about this stuff.
link |
And maybe we should just think about this a bit more
link |
about what we've learned about the brain
link |
and what are those team at Nimenta, what have they done?
link |
Is that useful for us?
link |
Yeah, so is there elements of all the cortical theory
link |
that things we've been talking about
link |
that may be useful in the short term?
link |
Yes, in the short term, yes.
link |
This is the, sorry to interrupt,
link |
but the open question is,
link |
it certainly feels from my perspective
link |
that in the long term,
link |
some of the ideas we've been talking about
link |
will be extremely useful.
link |
The question is whether in the short term.
link |
Well, this is always what I would call
link |
the entrepreneur's dilemma.
link |
So you have this long term vision,
link |
oh, we're gonna all be driving electric cars
link |
or we're all gonna have computers
link |
or we're all gonna, whatever.
link |
And you're at some point in time and you say,
link |
I can see that long term vision,
link |
I'm sure it's gonna happen.
link |
How do I get there without killing myself?
link |
Without going out of business, right?
link |
That's the challenge.
link |
That's the dilemma.
link |
That's the really difficult thing to do.
link |
So we're facing that right now.
link |
So ideally what you'd wanna do
link |
is find some steps along the way
link |
that you can get there incrementally.
link |
You don't have to like throw it all out
link |
and start over again.
link |
The first thing that we've done
link |
is we focus on the sparse representations.
link |
So just in case you don't know what that means
link |
or some of the listeners don't know what that means,
link |
in the brain, if I have like 10,000 neurons,
link |
what you would see is maybe 2% of them active at a time.
link |
You don't see 50%, you don't see 30%,
link |
And it's always like that.
link |
For any set of sensory inputs?
link |
It doesn't matter if anything,
link |
doesn't matter any part of the brain.
link |
But which neurons differs?
link |
Which neurons are active?
link |
Yeah, so let's say I take 10,000 neurons
link |
that are representing something.
link |
They're sitting there in a little block together.
link |
It's a teeny little block of neurons, 10,000 neurons.
link |
And they're representing a location,
link |
they're representing a cup,
link |
they're representing the input from my sensors.
link |
I don't know, it doesn't matter.
link |
It's representing something.
link |
The way the representations occur,
link |
it's always a sparse representation.
link |
Meaning it's a population code.
link |
So which 200 cells are active tells me what's going on.
link |
It's not, individual cells aren't that important at all.
link |
It's the population code that matters.
link |
And when you have sparse population codes,
link |
then all kinds of beautiful properties come out of them.
link |
So the brain uses sparse population codes.
link |
We've written and described these benefits
link |
in some of our papers.
link |
So they give this tremendous robustness to the systems.
link |
Brains are incredibly robust.
link |
Neurons are dying all the time and spasming
link |
and synapses are falling apart all the time.
link |
And it keeps working.
link |
So what Sibutai and Louise, one of our other engineers here
link |
have done, have shown they're introducing sparseness
link |
into convolutional neural networks.
link |
Now other people are thinking along these lines,
link |
but we're going about it in a more principled way, I think.
link |
And we're showing that if you enforce sparseness
link |
throughout these convolutional neural networks
link |
in both the act, which sort of,
link |
which neurons are active and the connections between them,
link |
that you get some very desirable properties.
link |
So one of the current hot topics in deep learning right now
link |
are these adversarial examples.
link |
So, you know, you give me any deep learning network
link |
and I can give you a picture that looks perfect
link |
and you're going to call it, you know,
link |
you're going to say the monkey is, you know, an airplane.
link |
So that's a problem.
link |
And DARPA just announced some big thing.
link |
They're trying to, you know, have some contest for this.
link |
But if you enforce sparse representations here,
link |
many of these problems go away.
link |
They're much more robust and they're not easy to fool.
link |
So we've already shown some of those results,
link |
just literally in January or February,
link |
just like last month we did that.
link |
And you can, I think it's on bioRxiv right now,
link |
or on iRxiv, you can read about it.
link |
But, so that's like a baby step, okay?
link |
That's taking something from the brain.
link |
We know about sparseness.
link |
We know why it's important.
link |
We know what it gives the brain.
link |
So let's try to enforce that onto this.
link |
What's your intuition why sparsity leads to robustness?
link |
Because it feels like it would be less robust.
link |
Why would you feel the rest robust to you?
link |
So it just feels like if the fewer neurons are involved,
link |
the more fragile the representation.
link |
But I didn't say there was lots of few neurons.
link |
I said, let's say 200.
link |
There's still a lot, it's just.
link |
So here's an intuition for it.
link |
This is a bit technical, so for engineers,
link |
machine learning people, this will be easy,
link |
but all the listeners, maybe not.
link |
If you're trying to classify something,
link |
you're trying to divide some very high dimensional space
link |
into different pieces, A and B.
link |
And you're trying to create some point where you say,
link |
all these points in this high dimensional space are A,
link |
and all these points in this high dimensional space are B.
link |
And if you have points that are close to that line,
link |
it's not very robust.
link |
It works for all the points you know about,
link |
but it's not very robust,
link |
because you can just move a little bit
link |
and you've crossed over the line.
link |
When you have sparse representations,
link |
imagine I pick, I'm gonna pick 200 cells active
link |
out of 10,000, okay?
link |
So I have 200 cells active.
link |
Now let's say I pick randomly another,
link |
a different representation, 200.
link |
The overlap between those is gonna be very small,
link |
I can pick millions of samples randomly of 200 neurons,
link |
and not one of them will overlap more than just a few.
link |
So one way to think about it is,
link |
if I wanna fool one of these representations
link |
to look like one of those other representations,
link |
I can't move just one cell, or two cells,
link |
or three cells, or four cells.
link |
I have to move 100 cells.
link |
And that makes them robust.
link |
In terms of further, so you mentioned sparsity.
link |
What would be the next thing?
link |
Okay, so we have, we picked one.
link |
We don't know if it's gonna work well yet.
link |
So again, we're trying to come up with incremental ways
link |
to moving from brain theory to add pieces
link |
to machine learning, current machine learning world,
link |
and one step at a time.
link |
So the next thing we're gonna try to do
link |
is sort of incorporate some of the ideas
link |
of the thousand brains theory,
link |
that you have many, many models that are voting.
link |
Now that idea is not new.
link |
There's a mixture of models that's been around
link |
But the way the brain does it is a little different.
link |
And the way it votes is different.
link |
And the kind of way it represents uncertainty
link |
So we're just starting this work,
link |
but we're gonna try to see if we can sort of incorporate
link |
some of the principles of voting,
link |
or principles of the thousand brain theory.
link |
Like lots of simple models that talk to each other
link |
And can we build more machines, systems that learn faster
link |
and also, well mostly are multimodal
link |
and robust to multimodal type of issues.
link |
So one of the challenges there
link |
is the machine learning computer vision community
link |
has certain sets of benchmarks,
link |
sets of tests based on which they compete.
link |
And I would argue, especially from your perspective,
link |
that those benchmarks aren't that useful
link |
for testing the aspects that the brain is good at,
link |
They're not really testing intelligence.
link |
They're very fine.
link |
And it's been extremely useful
link |
for developing specific mathematical models,
link |
but it's not useful in the long term
link |
for creating intelligence.
link |
So you think you also have a role in proposing
link |
Yeah, this is a very,
link |
you've identified a very serious problem.
link |
First of all, the tests that they have
link |
are the tests that they want.
link |
Not the tests of the other things
link |
that we're trying to do, right?
link |
You know, what are the, so on.
link |
The second thing is sometimes these,
link |
to be competitive in these tests,
link |
you have to have huge data sets and huge computing power.
link |
And so, you know, and we don't have that here.
link |
We don't have it as well as other big teams
link |
that big companies do.
link |
So there's numerous issues there.
link |
You know, we come out, you know,
link |
where our approach to this is all based on,
link |
in some sense, you might argue, elegance.
link |
We're coming at it from like a theoretical base
link |
that we think, oh my God, this is so clearly elegant.
link |
This is how brains work.
link |
This is what intelligence is.
link |
But the machine learning world has gotten in this phase
link |
where they think it doesn't matter.
link |
Doesn't matter what you think,
link |
as long as you do, you know, 0.1% better on this benchmark,
link |
that's what, that's all that matters.
link |
And that's a problem.
link |
You know, we have to figure out how to get around that.
link |
That's a challenge for us.
link |
That's one of the challenges that we have to deal with.
link |
So I agree, you've identified a big issue.
link |
It's difficult for those reasons.
link |
But you know, part of the reasons I'm talking to you here
link |
today is I hope I'm gonna get some machine learning people
link |
to say, I'm gonna read those papers.
link |
Those might be some interesting ideas.
link |
I'm tired of doing this 0.1% improvement stuff, you know?
link |
Well, that's why I'm here as well,
link |
because I think machine learning now as a community
link |
is at a place where the next step needs to be orthogonal
link |
to what has received success in the past.
link |
Well, you see other leaders saying this,
link |
machine learning leaders, you know,
link |
Jeff Hinton with his capsules idea.
link |
Many people have gotten up to say, you know,
link |
we're gonna hit road map, maybe we should look at the brain,
link |
you know, things like that.
link |
So hopefully that thinking will occur organically.
link |
And then we're in a nice position for people to come
link |
and look at our work and say,
link |
well, what can we learn from these guys?
link |
Yeah, MIT is launching a billion dollar computing college
link |
that's centered around this idea, so.
link |
Is it on this idea of what?
link |
Well, the idea that, you know,
link |
the humanities, psychology, and neuroscience
link |
have to work all together to get to build the S.
link |
Yeah, I mean, Stanford just did
link |
this Human Centered AI Center.
link |
I'm a little disappointed in these initiatives
link |
because, you know, they're focusing
link |
on sort of the human side of it,
link |
and it could very easily slip into
link |
how humans interact with intelligent machines,
link |
which is nothing wrong with that,
link |
but that's not, that is orthogonal
link |
to what we're trying to do.
link |
We're trying to say, like,
link |
what is the essence of intelligence?
link |
In fact, I wanna build intelligent machines
link |
that aren't emotional, that don't smile at you,
link |
that, you know, that aren't trying to tuck you in at night.
link |
Yeah, there is that pattern that you,
link |
when you talk about understanding humans
link |
is important for understanding intelligence,
link |
that you start slipping into topics of ethics
link |
or, yeah, like you said,
link |
the interactive elements as opposed to,
link |
no, no, no, we have to zoom in on the brain,
link |
study what the human brain, the baby, the...
link |
Let's study what a brain does.
link |
And then we can decide which parts of that
link |
we wanna recreate in some system,
link |
but until you have that theory about what the brain does,
link |
what's the point, you know, it's just,
link |
you're gonna be wasting time, I think.
link |
Right, just to break it down
link |
on the artificial neural network side,
link |
maybe you could speak to this
link |
on the biological neural network side,
link |
the process of learning versus the process of inference.
link |
Maybe you can explain to me,
link |
is there a difference between,
link |
you know, in artificial neural networks,
link |
there's a difference between the learning stage
link |
and the inference stage.
link |
Do you see the brain as something different?
link |
One of the big distinctions that people often say,
link |
I don't know how correct it is,
link |
is artificial neural networks need a lot of data.
link |
They're very inefficient learning.
link |
Do you see that as a correct distinction
link |
from the biology of the human brain,
link |
that the human brain is very efficient,
link |
or is that just something we deceive ourselves?
link |
No, it is efficient, obviously.
link |
We can learn new things almost instantly.
link |
And so what elements do you think are useful?
link |
Yeah, I can talk about that.
link |
You brought up two issues there.
link |
So remember I talked early about the constraints
link |
we always feel, well, one of those constraints
link |
is the fact that brains are continually learning.
link |
That's not something we said, oh, we can add that later.
link |
That's something that was upfront,
link |
had to be there from the start,
link |
made our problems harder.
link |
But we showed, going back to the 2016 paper
link |
on sequence memory, we showed how that happens,
link |
how the brains infer and learn at the same time.
link |
And our models do that.
link |
And they're not two separate phases,
link |
or two separate sets of time.
link |
I think that's a big, big problem in AI,
link |
at least for many applications, not for all.
link |
So I can talk about that.
link |
There are some, it gets detailed,
link |
there are some parts of the neocortex in the brain
link |
where actually what's going on,
link |
there's these cycles of activity in the brain.
link |
And there's very strong evidence
link |
that you're doing more of inference
link |
on one part of the phase,
link |
and more of learning on the other part of the phase.
link |
So the brain can actually sort of separate
link |
different populations of cells
link |
or going back and forth like this.
link |
But in general, I would say that's an important problem.
link |
We have all of our networks that we've come up with do both.
link |
And they're continuous learning networks.
link |
And you mentioned benchmarks earlier.
link |
Well, there are no benchmarks about that.
link |
So we have to, we get in our little soapbox,
link |
and hey, by the way, this is important,
link |
and here's a mechanism for doing that.
link |
But until you can prove it to someone
link |
in some commercial system or something, it's a little harder.
link |
So yeah, one of the things I had to linger on that
link |
is in some ways to learn the concept of a coffee cup,
link |
you only need this one coffee cup
link |
and maybe some time alone in a room with it.
link |
Well, the first thing is,
link |
imagine I reach my hand into a black box
link |
and I'm reaching, I'm trying to touch something.
link |
I don't know upfront if it's something I already know
link |
or if it's a new thing.
link |
And I have to, I'm doing both at the same time.
link |
I don't say, oh, let's see if it's a new thing.
link |
Oh, let's see if it's an old thing.
link |
As I go, my brain says, oh, it's new or it's not new.
link |
And if it's new, I start learning what it is.
link |
And by the way, it starts learning from the get go,
link |
even if it's gonna recognize it.
link |
So they're not separate problems.
link |
And so that's the thing there.
link |
The other thing you mentioned was the fast learning.
link |
So I was just talking about continuous learning,
link |
but there's also fast learning.
link |
Literally, I can show you this coffee cup
link |
and I say, here's a new coffee cup.
link |
It's got the logo on it.
link |
Take a look at it, done, you're done.
link |
You can predict what it's gonna look like,
link |
you know, in different positions.
link |
So I can talk about that too.
link |
In the brain, the way learning occurs,
link |
I mentioned this earlier, but I'll mention it again.
link |
The way learning occurs,
link |
imagine I am a section of a dendrite of a neuron,
link |
and I'm gonna learn something new.
link |
Doesn't matter what it is.
link |
I'm just gonna learn something new.
link |
I need to recognize a new pattern.
link |
So what I'm gonna do is I'm gonna form new synapses.
link |
New synapses, we're gonna rewire the brain
link |
onto that section of the dendrite.
link |
Once I've done that, everything else that neuron has learned
link |
is not affected by it.
link |
That's because it's isolated
link |
to that small section of the dendrite.
link |
They're not all being added together, like a point neuron.
link |
So if I learn something new on this segment here,
link |
it doesn't change any of the learning
link |
that occur anywhere else in that neuron.
link |
So I can add something without affecting previous learning.
link |
And I can do it quickly.
link |
Now let's talk, we can talk about the quickness,
link |
how it's done in real neurons.
link |
You might say, well, doesn't it take time to form synapses?
link |
Yes, it can take maybe an hour to form a new synapse.
link |
We can form memories quicker than that,
link |
and I can explain that how it happens too, if you want.
link |
But it's getting a bit neurosciencey.
link |
That's great, but is there an understanding
link |
of these mechanisms at every level?
link |
So from the short term memories and the forming.
link |
So this idea of synaptogenesis, the growth of new synapses,
link |
that's well described, it's well understood.
link |
And that's an essential part of learning.
link |
Going back many, many years,
link |
people, you know, it was, what's his name,
link |
the psychologist who proposed, Hebb, Donald Hebb.
link |
He proposed that learning was the modification
link |
of the strength of a connection between two neurons.
link |
People interpreted that as the modification
link |
of the strength of a synapse.
link |
He didn't say that.
link |
He just said there's a modification
link |
between the effect of one neuron and another.
link |
So synaptogenesis is totally consistent
link |
with what Donald Hebb said.
link |
But anyway, there's these mechanisms,
link |
the growth of new synapses.
link |
You can go online, you can watch a video
link |
of a synapse growing in real time.
link |
It's literally, you can see this little thing going boop.
link |
It's pretty impressive.
link |
So those mechanisms are known.
link |
Now there's another thing that we've speculated
link |
and we've written about,
link |
which is consistent with known neuroscience,
link |
but it's less proven.
link |
And this is the idea, how do I form a memory
link |
really, really quickly?
link |
Like instantaneous.
link |
If it takes an hour to grow a synapse,
link |
like that's not instantaneous.
link |
So there are types of synapses called silent synapses.
link |
They look like a synapse, but they don't do anything.
link |
They're just sitting there.
link |
It's like if an action potential comes in,
link |
it doesn't release any neurotransmitter.
link |
Some parts of the brain have more of these than others.
link |
For example, the hippocampus has a lot of them,
link |
which is where we associate most short term memory with.
link |
So what we speculated, again, in that 2016 paper,
link |
we proposed that the way we form very quick memories,
link |
very short term memories, or quick memories,
link |
is that we convert silent synapses into active synapses.
link |
It's like saying a synapse has a zero weight
link |
but the longterm memory has to be formed by synaptogenesis.
link |
So you can remember something really quickly
link |
by just flipping a bunch of these guys from silent to active.
link |
It's not from 0.1 to 0.15.
link |
It's like, it doesn't do anything
link |
till it releases transmitter.
link |
And if I do that over a bunch of these,
link |
I've got a very quick short term memory.
link |
So I guess the lesson behind this
link |
is that most neural networks today are fully connected.
link |
Every neuron connects every other neuron
link |
from layer to layer.
link |
That's not correct in the brain.
link |
We don't want that.
link |
We actually don't want that.
link |
You want a very sparse connectivity
link |
so that any neuron connects to some subset of the neurons
link |
in the other layer.
link |
And it does so on a dendrite by dendrite segment basis.
link |
So it's a very some parcelated out type of thing.
link |
And that then learning is not adjusting all these weights,
link |
but learning is just saying,
link |
okay, connect to these 10 cells here right now.
link |
In that process, you know, with artificial neural networks,
link |
it's a very simple process of backpropagation
link |
that adjusts the weights.
link |
The process of synaptogenesis.
link |
Backpropagation requires something
link |
that really can't happen in brains.
link |
This backpropagation of this error signal,
link |
that really can't happen.
link |
People are trying to make it happen in brains,
link |
but it doesn't happen in brains.
link |
This is pure Hebbian learning.
link |
Well, synaptogenesis is pure Hebbian learning.
link |
It's basically saying,
link |
there's a population of cells over here
link |
that are active right now.
link |
And there's a population of cells over here
link |
How do I form connections between those active cells?
link |
And it's literally saying this guy became active.
link |
These 100 neurons here became active
link |
before this neuron became active.
link |
So form connections to those ones.
link |
There's no propagation of error, nothing.
link |
All the networks we do,
link |
all the models we have work on almost completely on
link |
but on dendritic segments
link |
and multiple synapses at the same time.
link |
So now let's sort of turn the question
link |
that you already answered,
link |
and maybe you can answer it again.
link |
If you look at the history of artificial intelligence,
link |
where do you think we stand?
link |
How far are we from solving intelligence?
link |
You said you were very optimistic.
link |
Can you elaborate on that?
link |
Yeah, it's always the crazy question to ask
link |
because no one can predict the future.
link |
So I'll tell you a story.
link |
I used to run a different neuroscience institute
link |
called the Redwood Neuroscience Institute,
link |
and we would hold these symposiums
link |
and we'd get like 35 scientists
link |
from around the world to come together.
link |
And I used to ask them all the same question.
link |
I would say, well, how long do you think it'll be
link |
before we understand how the neocortex works?
link |
And everyone went around the room
link |
and they had introduced the name
link |
and they have to answer that question.
link |
So I got, the typical answer was 50 to 100 years.
link |
Some people would say 500 years.
link |
Some people said never.
link |
I said, why are you a neuroscientist?
link |
It's never gonna, it's a good pay.
link |
So, you know, but it doesn't work like that.
link |
As I mentioned earlier, these are not,
link |
these are step functions.
link |
Things happen and then bingo, they happen.
link |
You can't predict that.
link |
I feel I've already passed a step function.
link |
So if I can do my job correctly over the next five years,
link |
then, meaning I can proselytize these ideas.
link |
I can convince other people they're right.
link |
We can show that other people,
link |
machine learning people should pay attention
link |
Then we're definitely in an under 20 year timeframe.
link |
If I can do those things, if I'm not successful in that,
link |
and this is the last time anyone talks to me
link |
and no one reads our papers and you know,
link |
and I'm wrong or something like that,
link |
then I don't know.
link |
But it's not 50 years.
link |
Think about electric cars.
link |
How quickly are they gonna populate the world?
link |
It probably takes about a 20 year span.
link |
It'll be something like that.
link |
But I think if I can do what I said, we're starting it.
link |
And of course there could be other,
link |
you said step functions.
link |
It could be everybody gives up on your ideas for 20 years
link |
and then all of a sudden somebody picks it up again.
link |
Wait, that guy was onto something.
link |
Yeah, so that would be a failure on my part, right?
link |
Think about Charles Babbage.
link |
Charles Babbage, he's the guy who invented the computer
link |
back in the 18 something, 1800s.
link |
And everyone forgot about it until 100 years later.
link |
And say, hey, this guy figured this stuff out
link |
But he was ahead of his time.
link |
I don't think, as I said,
link |
I recognize this is part of any entrepreneur's challenge.
link |
I use entrepreneur broadly in this case.
link |
I'm not meaning like I'm building a business
link |
or trying to sell something.
link |
I mean, I'm trying to sell ideas.
link |
And this is the challenge as to how you get people
link |
to pay attention to you, how do you get them
link |
to give you positive or negative feedback,
link |
how do you get the people to act differently
link |
based on your ideas.
link |
So we'll see how well we do on that.
link |
So you know that there's a lot of hype
link |
behind artificial intelligence currently.
link |
Do you, as you look to spread the ideas
link |
that are of neocortical theory, the things you're working on,
link |
do you think there's some possibility
link |
we'll hit an AI winter once again?
link |
Yeah, it's certainly a possibility.
link |
No question about it.
link |
Is that something you worry about?
link |
Yeah, well, I guess, do I worry about it?
link |
I haven't decided yet if that's good or bad for my mission.
link |
That's true, that's very true.
link |
Because it's almost like you need the winter
link |
to refresh the palette.
link |
Yeah, it's like, I want, here's what you wanna have it is.
link |
You want, like to the extent that everyone is so thrilled
link |
about the current state of machine learning and AI
link |
and they don't imagine they need anything else,
link |
it makes my job harder.
link |
If everything crashed completely
link |
and every student left the field
link |
and there was no money for anybody to do anything
link |
and it became an embarrassment
link |
to talk about machine intelligence and AI,
link |
that wouldn't be good for us either.
link |
You want sort of the soft landing approach, right?
link |
You want enough people, the senior people in AI
link |
and machine learning to say, you know,
link |
we need other approaches.
link |
We really need other approaches.
link |
Damn, we need other approaches.
link |
Maybe we should look to the brain.
link |
Okay, let's look to the brain.
link |
Who's got some brain ideas?
link |
Okay, let's start a little project on the side here
link |
trying to do brain idea related stuff.
link |
That's the ideal outcome we would want.
link |
So I don't want a total winter
link |
and yet I don't want it to be sunny all the time either.
link |
So what do you think it takes to build a system
link |
with human level intelligence
link |
where once demonstrated you would be very impressed?
link |
So does it have to have a body?
link |
Does it have to have the C word we used before,
link |
consciousness as an entirety in a holistic sense?
link |
First of all, I don't think the goal
link |
is to create a machine that is human level intelligence.
link |
I think it's a false goal.
link |
Back to Turing, I think it was a false statement.
link |
We want to understand what intelligence is
link |
and then we can build intelligent machines
link |
of all different scales, all different capabilities.
link |
A dog is intelligent.
link |
I don't need, that'd be pretty good to have a dog.
link |
But what about something that doesn't look
link |
like an animal at all, in different spaces?
link |
So my thinking about this is that
link |
we want to define what intelligence is,
link |
agree upon what makes an intelligent system.
link |
We can then say, okay, we're now gonna build systems
link |
that work on those principles or some subset of them
link |
and we can apply them to all different types of problems.
link |
And the kind, the idea, it's not computing.
link |
We don't ask, if I take a little one chip computer,
link |
I don't say, well, that's not a computer
link |
because it's not as powerful as this big server over here.
link |
No, no, because we know that what the principles
link |
of computing are and I can apply those principles
link |
to a small problem or into a big problem.
link |
And same, intelligence needs to get there.
link |
We have to say, these are the principles.
link |
I can make a small one, a big one.
link |
I can make them distributed.
link |
I can put them on different sensors.
link |
They don't have to be human like at all.
link |
Now, you did bring up a very interesting question
link |
Does it have to have a body?
link |
It has to have some concept of movement.
link |
It has to be able to move through these reference frames
link |
I talked about earlier.
link |
Whether it's physically moving,
link |
like I need, if I'm gonna have an AI
link |
that understands coffee cups,
link |
it's gonna have to pick up the coffee cup
link |
and touch it and look at it with its eyes and hands
link |
or something equivalent to that.
link |
If I have a mathematical AI,
link |
maybe it needs to move through mathematical spaces.
link |
I could have a virtual AI that lives in the internet
link |
and its movements are traversing links
link |
and digging into files,
link |
but it's got a location that it's traveling
link |
through some space.
link |
You can't have an AI that just take some flash thing input.
link |
We call it flash inference.
link |
Here's a pattern, done.
link |
No, it's movement pattern, movement pattern,
link |
movement pattern, attention, digging, building structure,
link |
figuring out the model of the world.
link |
So some sort of embodiment,
link |
whether it's physical or not, has to be part of it.
link |
So self awareness and the way to be able to answer
link |
Well, you're bringing up self,
link |
that's a different topic, self awareness.
link |
No, the very narrow definition of self,
link |
meaning knowing a sense of self enough to know
link |
where am I in the space where it's actually.
link |
Yeah, basically the system needs to know its location
link |
or each component of the system needs to know
link |
where it is in the world at that point in time.
link |
So self awareness and consciousness.
link |
Do you think one, from the perspective of neuroscience
link |
and neurocortex, these are interesting topics,
link |
Do you have any ideas of why the heck it is
link |
that we have a subjective experience at all?
link |
Yeah, I have a lot of thoughts on that.
link |
And is it useful or is it just a side effect of us?
link |
It's interesting to think about.
link |
I don't think it's useful as a means to figure out
link |
how to build intelligent machines.
link |
It's something that systems do
link |
and we can talk about what it is that are like,
link |
well, if I build a system like this,
link |
then it would be self aware.
link |
Or if I build it like this, it wouldn't be self aware.
link |
So that's a choice I can have.
link |
It's not like, oh my God, it's self aware.
link |
I can't turn, I heard an interview recently
link |
with this philosopher from Yale,
link |
I can't remember his name, I apologize for that.
link |
But he was talking about,
link |
well, if these computers are self aware,
link |
then it would be a crime to unplug them.
link |
And I'm like, oh, come on, that's not,
link |
I unplug myself every night, I go to sleep.
link |
I plug myself in again in the morning and there I am.
link |
So people get kind of bent out of shape about this.
link |
I have very definite, very detailed understanding
link |
or opinions about what it means to be conscious
link |
and what it means to be self aware.
link |
I don't think it's that interesting a problem.
link |
You've talked to Christoph Koch.
link |
He thinks that's the only problem.
link |
I didn't actually listen to your interview with him,
link |
but I know him and I know that's the thing he cares about.
link |
He also thinks intelligence and consciousness are disjoint.
link |
So I mean, it's not, you don't have to have one or the other.
link |
I disagree with that.
link |
I just totally disagree with that.
link |
So where's your thoughts and consciousness,
link |
where does it emerge from?
link |
So then we have to break it down to the two parts, okay?
link |
Because consciousness isn't one thing.
link |
That's part of the problem with that term
link |
is it means different things to different people
link |
and there's different components of it.
link |
There is a concept of self awareness, okay?
link |
That can be very easily explained.
link |
You have a model of your own body.
link |
The neocortex models things in the world
link |
and it also models your own body.
link |
And then it has a memory.
link |
It can remember what you've done, okay?
link |
So it can remember what you did this morning,
link |
can remember what you had for breakfast and so on.
link |
And so I can say to you, okay, Lex,
link |
were you conscious this morning when you had your bagel?
link |
And you'd say, yes, I was conscious.
link |
Now what if I could take your brain
link |
and revert all the synapses back
link |
to the state they were this morning?
link |
And then I said to you, Lex,
link |
were you conscious when you ate the bagel?
link |
And you said, no, I wasn't conscious.
link |
I said, here's a video of eating the bagel.
link |
And you said, I wasn't there.
link |
That's not possible
link |
because I must've been unconscious at that time.
link |
So we can just make this one to one correlation
link |
between memory of your body's trajectory through the world
link |
over some period of time,
link |
a memory and the ability to recall that memory
link |
is what you would call conscious.
link |
I was conscious of that, it's a self awareness.
link |
And any system that can recall,
link |
memorize what it's done recently
link |
and bring that back and invoke it again
link |
would say, yeah, I'm aware.
link |
I remember what I did.
link |
All right, I got it.
link |
That's an easy one.
link |
Although some people think that's a hard one.
link |
The more challenging part of consciousness
link |
is this one that's sometimes used
link |
going by the word of qualia,
link |
which is, why does an object seem red?
link |
And why does pain feel like something?
link |
Why do I feel redness?
link |
Or why do I feel painness?
link |
And then I could say, well,
link |
why does sight seems different than hearing?
link |
It's the same problem.
link |
It's really, these are all just neurons.
link |
And so how is it that,
link |
why does looking at you feel different than hearing you?
link |
It feels different, but there's just neurons in my head.
link |
They're all doing the same thing.
link |
So that's an interesting question.
link |
The best treatise I've read about this
link |
is by a guy named Oregon.
link |
He wrote a book called,
link |
Why Red Doesn't Sound Like a Bell.
link |
It's a little, it's not a trade book, easy to read,
link |
but it, and it's an interesting question.
link |
Take something like color.
link |
Color really doesn't exist in the world.
link |
It's not a property of the world.
link |
Property of the world that exists is light frequency.
link |
And that gets turned into,
link |
we have certain cells in the retina
link |
that respond to different frequencies
link |
different than others.
link |
And so when they enter the brain,
link |
you just have a bunch of axons
link |
that are firing at different rates.
link |
And from that, we perceive color.
link |
But there is no color in the brain.
link |
I mean, there's no color coming in on those synapses.
link |
It's just a correlation between some axons
link |
and some property of frequency.
link |
And that isn't even color itself.
link |
Frequency doesn't have a color.
link |
It's just what it is.
link |
So then the question is,
link |
well, why does it even appear to have a color at all?
link |
Just as you're describing it,
link |
there seems to be a connection to those ideas
link |
of reference frames.
link |
I mean, it just feels like consciousness
link |
having the subject,
link |
assigning the feeling of red to the actual color
link |
or to the wavelength is useful for intelligence.
link |
Yeah, I think that's a good way of putting it.
link |
It's useful as a predictive mechanism
link |
or useful as a generalization idea.
link |
It's a way of grouping things together to say,
link |
it's useful to have a model like this.
link |
So think about the well known syndrome
link |
that people who've lost a limb experience
link |
called phantom limbs.
link |
And what they claim is they can have their arm is removed,
link |
but they feel their arm.
link |
That not only feel it, they know it's there.
link |
It's there, I know it's there.
link |
They'll swear to you that it's there.
link |
And then they can feel pain in their arm
link |
and they'll feel pain in their finger.
link |
And if they move their non existent arm behind their back,
link |
then they feel the pain behind their back.
link |
So this whole idea that your arm exists
link |
is a model of your brain.
link |
It may or may not really exist.
link |
And just like, but it's useful to have a model of something
link |
that sort of correlates to things in the world.
link |
So you can make predictions about what would happen
link |
when those things occur.
link |
It's a little bit of a fuzzy,
link |
but I think you're getting quite towards the answer there.
link |
It's useful for the model to express things certain ways
link |
that we can then map them into these reference frames
link |
and make predictions about them.
link |
I need to spend more time on this topic.
link |
It doesn't bother me.
link |
Do you really need to spend more time?
link |
It does feel special that we have subjective experience,
link |
but I'm yet to know why.
link |
I'm just personally curious.
link |
It's not necessary for the work we're doing here.
link |
I don't think I need to solve that problem
link |
to build intelligent machines at all, not at all.
link |
But there is sort of the silly notion
link |
that you described briefly
link |
that doesn't seem so silly to us humans is,
link |
if you're successful building intelligent machines,
link |
it feels wrong to then turn them off.
link |
Because if you're able to build a lot of them,
link |
it feels wrong to then be able to turn off the...
link |
Let's break that down a bit.
link |
As humans, why do we fear death?
link |
There's two reasons we fear death.
link |
Well, first of all, I'll say,
link |
when you're dead, it doesn't matter at all.
link |
So why do we fear death?
link |
We fear death for two reasons.
link |
One is because we are programmed genetically to fear death.
link |
That's a survival and pop beginning of the genes thing.
link |
And we also are programmed to feel sad
link |
when people we know die.
link |
We don't feel sad for someone we don't know dies.
link |
There's people dying right now,
link |
they're only just gonna say,
link |
I don't feel bad about them,
link |
because I don't know them.
link |
But if I knew them, I'd feel really bad.
link |
So again, these are old brain,
link |
genetically embedded things that we fear death.
link |
It's outside of those uncomfortable feelings.
link |
There's nothing else to worry about.
link |
Well, wait, hold on a second.
link |
Do you know the denial of death by Becker?
link |
There's a thought that death is,
link |
our whole conception of our world model
link |
kind of assumes immortality.
link |
And then death is this terror that underlies it all.
link |
Some people's world model, not mine.
link |
But, okay, so what Becker would say
link |
is that you're just living in an illusion.
link |
You've constructed an illusion for yourself
link |
because it's such a terrible terror,
link |
the fact that this...
link |
What's the illusion?
link |
The illusion that death doesn't matter.
link |
You're still not coming to grips with...
link |
The illusion of what?
link |
Oh, like it's not gonna happen?
link |
You're actually operating.
link |
You haven't, even though you said you've accepted it,
link |
you haven't really accepted the notion that you're gonna die
link |
So it sounds like you disagree with that notion.
link |
Yeah, yeah, totally.
link |
I literally, every night I go to bed, it's like dying.
link |
Like little deaths.
link |
It's little deaths.
link |
And if I didn't wake up, it wouldn't matter to me.
link |
Only if I knew that was gonna happen would it be bothersome.
link |
If I didn't know it was gonna happen, how would I know?
link |
Then I would worry about my wife.
link |
So imagine I was a loner and I lived in Alaska
link |
and I lived out there and there was no animals.
link |
Nobody knew I existed.
link |
I was just eating these roots all the time.
link |
And nobody knew I was there.
link |
And one day I didn't wake up.
link |
What pain in the world would there exist?
link |
Well, so most people that think about this problem
link |
would say that you're just deeply enlightened
link |
or are completely delusional.
link |
But I would say that's a very enlightened way
link |
That's the rational one as well.
link |
It's rational, that's right.
link |
But the fact is we don't,
link |
I mean, we really don't have an understanding
link |
of why the heck it is we're born and why we die
link |
and what happens after we die.
link |
Well, maybe there isn't a reason, maybe there is.
link |
So I'm interested in those big problems too, right?
link |
You interviewed Max Tegmark,
link |
and there's people like that, right?
link |
I'm interested in those big problems as well.
link |
And in fact, when I was young,
link |
I made a list of the biggest problems I could think of.
link |
First, why does anything exist?
link |
Second, why do we have the laws of physics that we have?
link |
Third, is life inevitable?
link |
And why is it here?
link |
Fourth, is intelligence inevitable?
link |
And why is it here?
link |
I stopped there because I figured
link |
if you can make a truly intelligent system,
link |
that will be the quickest way
link |
to answer the first three questions.
link |
And so I said, my mission, you asked me earlier,
link |
my first mission is to understand the brain,
link |
but I felt that is the shortest way
link |
to get to true machine intelligence.
link |
And I wanna get to true machine intelligence
link |
because even if it doesn't occur in my lifetime,
link |
other people will benefit from it
link |
because I think it'll occur in my lifetime,
link |
but 20 years, you never know.
link |
But that will be the quickest way for us to,
link |
we can make super mathematicians,
link |
we can make super space explorers,
link |
we can make super physicist brains that do these things
link |
and that can run experiments that we can't run.
link |
We don't have the abilities to manipulate things and so on,
link |
but we can build intelligent machines that do all those things
link |
with the ultimate goal of finding out the answers
link |
to the other questions.
link |
Let me ask you another depressing and difficult question,
link |
which is once we achieve that goal of creating,
link |
no, of understanding intelligence,
link |
do you think we would be happier,
link |
more fulfilled as a species?
link |
The understanding intelligence
link |
or understanding the answers to the big questions?
link |
Understanding intelligence.
link |
Oh, totally, totally.
link |
It would be far more fun place to live.
link |
I mean, just put aside this terminator nonsense
link |
and just think about, you can think about,
link |
we can talk about the risks of AI if you want.
link |
I'd love to, so let's talk about.
link |
But I think the world would be far better knowing things.
link |
We're always better than know things.
link |
Do you think it's better, is it a better place to live in
link |
that I know that our planet is one of many
link |
in the solar system and the solar system's one of many
link |
I think it's a more, I dread, I sometimes think like,
link |
God, what would it be like to live 300 years ago?
link |
I'd be looking up at the sky, I can't understand anything.
link |
Oh my God, I'd be like going to bed every night going,
link |
what's going on here?
link |
Well, I mean, in some sense I agree with you,
link |
but I'm not exactly sure.
link |
So I'm also a scientist, so I share your views,
link |
but I'm not, we're like rolling down the hill together.
link |
What's down the hill?
link |
I feel like we're climbing a hill.
link |
We're getting closer to enlightenment
link |
and you're going down the hill.
link |
We're climbing, we're getting pulled up a hill
link |
Our curiosity is, we're pulling ourselves up the hill
link |
Yeah, Sisyphus was doing the same thing with the rock.
link |
Yeah, yeah, yeah, yeah.
link |
But okay, our happiness aside, do you have concerns
link |
about, you talk about Sam Harris, Elon Musk,
link |
of existential threats of intelligent systems?
link |
No, I'm not worried about existential threats at all.
link |
There are some things we really do need to worry about.
link |
Even today's AI, we have things we have to worry about.
link |
We have to worry about privacy
link |
and about how it impacts false beliefs in the world.
link |
And we have real problems and things to worry about
link |
And that will continue as we create more intelligent systems.
link |
There's no question, the whole issue
link |
about making intelligent armaments and weapons
link |
is something that really we have to think about carefully.
link |
I don't think of those as existential threats.
link |
I think those are the kind of threats we always face
link |
and we'll have to face them here
link |
and we'll have to deal with them.
link |
We could talk about what people think
link |
are the existential threats,
link |
but when I hear people talking about them,
link |
they all sound hollow to me.
link |
They're based on ideas, they're based on people
link |
who really have no idea what intelligence is.
link |
And if they knew what intelligence was,
link |
they wouldn't say those things.
link |
So those are not experts in the field.
link |
Yeah, so there's two, right?
link |
So one is like super intelligence.
link |
So a system that becomes far, far superior
link |
in reasoning ability than us humans.
link |
How is that an existential threat?
link |
Then, so there's a lot of ways in which it could be.
link |
One way is us humans are actually irrational, inefficient
link |
and get in the way of, not happiness,
link |
but whatever the objective function is
link |
of maximizing that objective function.
link |
Super intelligent.
link |
The paperclip problem and things like that.
link |
So the paperclip problem but with the super intelligent.
link |
Yeah, yeah, yeah, yeah.
link |
So we already face this threat in some sense.
link |
They're called bacteria.
link |
These are organisms in the world
link |
that would like to turn everything into bacteria.
link |
And they're constantly morphing,
link |
they're constantly changing to evade our protections.
link |
And in the past, they have killed huge swaths
link |
of populations of humans on this planet.
link |
So if you wanna worry about something
link |
that's gonna multiply endlessly, we have it.
link |
And I'm far more worried in that regard.
link |
I'm far more worried that some scientists in the laboratory
link |
will create a super virus or a super bacteria
link |
that we cannot control.
link |
That is a more of an existential threat.
link |
Putting an intelligence thing on top of it
link |
actually seems to make it less existential to me.
link |
It's like, it limits its power.
link |
It limits where it can go.
link |
It limits the number of things it can do in many ways.
link |
A bacteria is something you can't even see.
link |
So that's only one of those problems.
link |
So the other one, just in your intuition about intelligence,
link |
when you think about intelligence of us humans,
link |
do you think of that as something,
link |
if you look at intelligence on a spectrum
link |
from zero to us humans,
link |
do you think you can scale that to something far,
link |
far superior to all the mechanisms we've been talking about?
link |
I wanna make another point here, Lex, before I get there.
link |
Intelligence is the neocortex.
link |
It is not the entire brain.
link |
The goal is not to make a human.
link |
The goal is not to make an emotional system.
link |
The goal is not to make a system
link |
that wants to have sex and reproduce.
link |
Why would I build that?
link |
If I wanna have a system that wants to reproduce
link |
and have sex, make bacteria, make computer viruses.
link |
Those are bad things, don't do that.
link |
Those are really bad, don't do those things.
link |
But if I just say I want an intelligent system,
link |
why does it have to have any of the human like emotions?
link |
Why does it even care if it lives?
link |
Why does it even care if it has food?
link |
It doesn't care about those things.
link |
It's just, you know, it's just in a trance
link |
thinking about mathematics or it's out there
link |
just trying to build the space for it on Mars.
link |
That's a choice we make.
link |
Don't make human like things,
link |
don't make replicating things,
link |
don't make things that have emotions,
link |
just stick to the neocortex.
link |
So that's a view actually that I share
link |
but not everybody shares in the sense that
link |
you have faith and optimism about us as engineers of systems,
link |
humans as builders of systems to not put in stupid, not.
link |
So this is why I mentioned the bacteria one.
link |
Because you might say, well, some person's gonna do that.
link |
Well, some person today could create a bacteria
link |
that's resistant to all the known antibacterial agents.
link |
So we already have that threat.
link |
We already know this is going on.
link |
It's not a new threat.
link |
So just accept that and then we have to deal with it, right?
link |
Yeah, so my point is nothing to do with intelligence.
link |
Intelligence is a separate component
link |
that you might apply to a system
link |
that wants to reproduce and do stupid things.
link |
Let's not do that.
link |
Yeah, in fact, it is a mystery
link |
why people haven't done that yet.
link |
My dad is a physicist, believes that the reason,
link |
he says, for example, nuclear weapons haven't proliferated
link |
amongst evil people.
link |
So one belief that I share is that
link |
there's not that many evil people in the world
link |
that would use, whether it's bacteria or nuclear weapons
link |
or maybe the future AI systems to do bad.
link |
So the fraction is small.
link |
And the second is that it's actually really hard,
link |
technically, so the intersection between evil
link |
and competent is small in terms of, and that's the.
link |
And by the way, to really annihilate humanity,
link |
you'd have to have sort of the nuclear winter phenomenon,
link |
which is not one person shooting or even 10 bombs.
link |
You'd have to have some automated system
link |
that detonates a million bombs
link |
or whatever many thousands we have.
link |
So extreme evil combined with extreme competence.
link |
And to start with building some stupid system
link |
that would automatically, Dr. Strangelove type of thing,
link |
you know, I mean, look, we could have
link |
some nuclear bomb go off in some major city in the world.
link |
I think that's actually quite likely, even in my lifetime.
link |
I don't think that's an unlikely thing.
link |
And it'd be a tragedy.
link |
But it won't be an existential threat.
link |
And it's the same as, you know, the virus of 1917,
link |
whatever it was, you know, the influenza.
link |
These bad things can happen and the plague and so on.
link |
We can't always prevent them.
link |
We always try, but we can't.
link |
But they're not existential threats
link |
until we combine all those crazy things together.
link |
So on the spectrum of intelligence from zero to human,
link |
do you have a sense of whether it's possible
link |
to create several orders of magnitude
link |
or at least double that of human intelligence?
link |
Talking about neuro context.
link |
I think it's the wrong thing to say double the intelligence.
link |
Break it down into different components.
link |
Can I make something that's a million times fast
link |
than a human brain?
link |
Yes, I can do that.
link |
Could I make something that is,
link |
has a lot more storage than the human brain?
link |
Yes, I could do that.
link |
More common, more copies of common.
link |
Can I make something that att