back to indexKarl Friston: Neuroscience and the Free Energy Principle | Lex Fridman Podcast #99
link |
The following is a conversation with Carl Friston,
link |
one of the greatest neuroscientists in history.
link |
Cited over 245,000 times,
link |
known for many influential ideas in brain imaging,
link |
neuroscience, and theoretical neurobiology,
link |
including especially the fascinating idea
link |
of the free energy principle for action and perception.
link |
Carl's mix of humor, brilliance, and kindness,
link |
to me, are inspiring and captivating.
link |
This was a huge honor and a pleasure.
link |
This is the Artificial Intelligence Podcast.
link |
If you enjoy it, subscribe on YouTube,
link |
review it with five stars on Apple Podcast,
link |
support it on Patreon,
link |
or simply connect with me on Twitter,
link |
at Lex Friedman, spelled F R I D M A N.
link |
As usual, I'll do a few minutes of ads now,
link |
and never any ads in the middle
link |
that can break the flow of the conversation.
link |
I hope that works for you,
link |
and doesn't hurt the listening experience.
link |
This show is presented by Cash App,
link |
the number one finance app in the App Store.
link |
When you get it, use code LEXPODCAST.
link |
Cash App lets you send money to friends by Bitcoin,
link |
and invest in the stock market with as little as $1.
link |
Since Cash App allows you to send
link |
and receive money digitally,
link |
let me mention a surprising fact related to physical money.
link |
Of all the currency in the world,
link |
roughly 8% of it is actual physical money.
link |
The other 92% of money only exists digitally.
link |
So again, if you get Cash App from the App Store,
link |
Google Play, and use the code LEXPODCAST, you get $10,
link |
and Cash App will also donate $10 to FIRST,
link |
an organization that is helping to advance robotics
link |
and STEM education for young people around the world.
link |
And now, here's my conversation with Carl Friston.
link |
How much of the human brain do we understand
link |
from the low level of neuronal communication
link |
to the functional level to the highest level,
link |
maybe the psychiatric disorder level?
link |
Well, we're certainly in a better position
link |
than we were last century.
link |
How far we've got to go, I think,
link |
is almost an unanswerable question.
link |
So you'd have to set the parameters,
link |
you know, what constitutes understanding, what level
link |
of understanding do you want?
link |
I think we've made enormous progress
link |
in terms of broad brush principles.
link |
Whether that affords a detailed cartography
link |
of the functional anatomy of the brain and what it does,
link |
right down to the microcircuitry and the neurons,
link |
that's probably out of reach at the present time.
link |
So the cartography, so mapping the brain,
link |
do you think mapping of the brain,
link |
the detailed, perfect imaging of it,
link |
does that get us closer to understanding
link |
of the mind, of the brain?
link |
So how far does it get us if we have
link |
that perfect cartography of the brain?
link |
I think there are lower bounds on that.
link |
It's a really interesting question.
link |
And it would determine the sort of scientific career
link |
you'd pursue if you believe that knowing
link |
every dendritic connection, every sort of microscopic,
link |
synaptic structure right down to the molecular level
link |
was gonna give you the right kind of information
link |
to understand the computational anatomy,
link |
then you'd choose to be a microscopist
link |
and you would study little cubic millimeters of brain
link |
for the rest of your life.
link |
If on the other hand you were interested
link |
in holistic functions and a sort of functional anatomy
link |
of the sort that a neuropsychologist would understand,
link |
you'd study brain lesions and strokes,
link |
just looking at the whole person.
link |
So again, it comes back to at what level
link |
do you want understanding?
link |
I think there are principled reasons not to go too far.
link |
If you commit to a view of the brain
link |
as a machine that's performing a form of inference
link |
and representing things, that level of understanding
link |
is necessarily cast in terms of probability densities
link |
and ensemble densities, distributions.
link |
And what that tells you is that you don't really want
link |
to look at the atoms to understand the thermodynamics
link |
of probabilistic descriptions of how the brain works.
link |
So I personally wouldn't look at the molecules
link |
or indeed the single neurons in the same way
link |
if I wanted to understand the thermodynamics
link |
of some non equilibrium steady state of a gas
link |
or an active material, I wouldn't spend my life
link |
looking at the individual molecules
link |
that constitute that ensemble.
link |
I'd look at their collective behavior.
link |
On the other hand, if you go too coarse grain,
link |
you're gonna miss some basic canonical principles
link |
of connectivity and architectures.
link |
I'm thinking here this bit colloquial,
link |
but this current excitement about high field
link |
magnetic resonance imaging at seven Tesla, why?
link |
Well, it gives us for the first time the opportunity
link |
to look at the brain in action at the level
link |
of a few millimeters that distinguish
link |
between different layers of the cortex
link |
that may be very important in terms of evincing
link |
generic principles of conical microcircuitry
link |
that are replicated throughout the brain
link |
that may tell us something fundamental
link |
about message passing in the brain
link |
and these density dynamics or neuronal
link |
or some more population dynamics
link |
that underwrite our brain function.
link |
So somewhere between a millimeter and a meter.
link |
Lingering for a bit on the big questions if you allow me,
link |
what to use the most beautiful or surprising characteristic
link |
of the human brain?
link |
I think it's hierarchical and recursive aspect.
link |
It's recurrent aspect.
link |
Of the structure or of the actual
link |
representation of power of the brain?
link |
Well, I think one speaks to the other.
link |
I was actually answering in a dull minded way
link |
from the point of view of purely its anatomy
link |
and its structural aspects.
link |
I mean, there are many marvelous organs in the body.
link |
Let's take your liver for example.
link |
Without it, you wouldn't be around for very long
link |
and it does some beautiful and delicate biochemistry
link |
and homeostasis and evolved with a finesse
link |
that would easily parallel the brain
link |
but it doesn't have a beautiful anatomy.
link |
It has a simple anatomy which is attractive
link |
in a minimalist sense but it doesn't have
link |
that crafted structure of sparse connectivity
link |
and that recurrence and that specialization
link |
that the brain has.
link |
So you said a lot of interesting terms here.
link |
So the recurrence, the sparsity,
link |
but you also started by saying hierarchical.
link |
So I've never thought of our brain as hierarchical.
link |
Sort of I always thought it's just like a giant mess,
link |
interconnected mess where it's very difficult
link |
to figure anything out.
link |
But in what sense do you see the brain as hierarchical?
link |
Well, I see it, it's not a magic soup.
link |
Which of course is what I used to think
link |
before I studied medicine and the like.
link |
So a lot of those terms imply each other.
link |
So hierarchies, if you just think about
link |
the nature of a hierarchy,
link |
how would you actually build one?
link |
And what you would have to do is basically
link |
carefully remove the right connections
link |
that destroy the completely connected soups
link |
that you might have in mind.
link |
So a hierarchy is in and of itself defined
link |
by a sparse and particular connectivity structure.
link |
I'm not committing to any particular form of hierarchy.
link |
But your sense is there is some.
link |
Oh, absolutely, yeah.
link |
In virtue of the fact that there is a sparsity
link |
of connectivity, not necessarily of a qualitative sort,
link |
but certainly of a quantitative sort.
link |
So it is demonstrably so that the further apart
link |
two parts of the brain are,
link |
the less likely they are to be wired,
link |
to possess axonal processes, neuronal processes
link |
that directly communicate one message
link |
or messages from one part of that brain
link |
to the other part of the brain.
link |
So we know there's a sparse connectivity.
link |
And furthermore, on the basis of anatomical connectivity
link |
in traces studies, we know that that sparsity
link |
underwrites a hierarchical and very structured
link |
sort of connectivity that might be best understood
link |
like a little bit like an onion.
link |
There is a concentric, sometimes referred to as centripetal
link |
by people like Marcel Masulam,
link |
hierarchical organization to the brain.
link |
So you can think of the brain as in a rough sense,
link |
like an onion, and all the sensory information
link |
and all the afferent outgoing messages
link |
that supply commands to your muscles
link |
or to your secretory organs come from the surface.
link |
So there's a massive exchange interface
link |
with the world out there on the surface.
link |
And then underneath, there's a little layer
link |
that sits and looks at the exchange on the surface.
link |
And then underneath that, there's a layer
link |
right the way down to the very center,
link |
to the deepest part of the onion.
link |
That's what I mean by a hierarchical organization.
link |
There's a discernible structure defined
link |
by the sparsity of connections
link |
that lends the architecture a hierarchical structure
link |
that tells one a lot about the kinds of representations
link |
Coming back to your earlier question,
link |
is this about the representational capacity
link |
or is it about the anatomy?
link |
Well, one underwrites the other.
link |
If one simply thinks of the brain
link |
as a message passing machine,
link |
a process that is in the service of doing something,
link |
then the circuitry and the connectivity
link |
that shape that message passing also dictate its function.
link |
So you've done a lot of amazing work
link |
in a lot of directions.
link |
So let's look at one aspect of that,
link |
of looking into the brain
link |
and trying to study this onion structure.
link |
What can we learn about the brain by imaging it?
link |
Which is one way to sort of look at the anatomy of it.
link |
Broadly speaking, what are the methods of imaging,
link |
but even bigger, what can we learn about it?
link |
Right, so well, most human neuroimaging
link |
that you might see in science journals
link |
that speaks to the way the brain works,
link |
measures brain activity over time.
link |
So that's the first thing to say,
link |
that we're effectively looking at fluctuations
link |
in neuronal responses,
link |
usually in response to some sensory input
link |
or some instruction, some task.
link |
Not necessarily, there's a lot of interest
link |
in just looking at the brain
link |
in terms of resting state, endogenous,
link |
or intrinsic activity.
link |
But crucially, at every point,
link |
looking at these fluctuations,
link |
either induced or intrinsic in the neural activity,
link |
and understanding them at two levels.
link |
So normally, people would recourse
link |
to two principles of brain organization
link |
that are complementary.
link |
One, functional specialization or segregation.
link |
So what does that mean?
link |
It simply means that there are certain parts of the brain
link |
that may be specialized for certain kinds of processing.
link |
For example, visual motion,
link |
our ability to recognize or to perceive movement
link |
in the visual world.
link |
And furthermore, that specialized processing
link |
may be spatially or anatomically segregated,
link |
leading to functional segregation.
link |
Which means that if I were to compare your brain activity
link |
during a period of viewing a static image,
link |
and then compare that to the responses of fluctuations
link |
in the brain when you were exposed to a moving image,
link |
say a flying bird,
link |
we'd expect to see
link |
restricted, segregated differences in activity.
link |
And those are basically the hotspots
link |
that you see in the statistical parametric maps
link |
that test for the significance of the responses
link |
that are circumscribed.
link |
So now, basically, we're talking about
link |
some people have perhaps unkindly called a neocartography.
link |
This is a phrenology augmented by modern day neuroimaging,
link |
basically finding blobs or bumps on the brain
link |
that do this or do that,
link |
and trying to understand the cartography
link |
of that functional specialization.
link |
So how much is there such,
link |
this is such a beautiful sort of ideal to strive for.
link |
We humans, scientists, would like this,
link |
to hope that there's a beautiful structure to this
link |
where it's, like you said, there's segregated regions
link |
that are responsible for the different function.
link |
How much hope is there to find such regions
link |
in terms of looking at the progress of studying the brain?
link |
Oh, I think enormous progress has been made
link |
in the past 20 or 30 years.
link |
So this is beyond incremental.
link |
At the advent of brain imaging,
link |
the very notion of functional segregation
link |
was just a hypothesis based upon a century,
link |
if not more, of careful neuropsychology,
link |
looking at people who had lost via insult
link |
or traumatic brain injury particular parts of the brain,
link |
and then saying, well, they can't do this
link |
or they can't do that.
link |
For example, losing the visual cortex
link |
and not being able to see,
link |
or losing particular parts of the visual cortex
link |
or regions known as V5
link |
or the middle temporal region, MT,
link |
and noticing that they selectively
link |
could not see moving things.
link |
And so that created the hypothesis
link |
that perhaps visual movement processing
link |
was located in this functionally segregated area.
link |
And you could then go and put invasive electrodes
link |
in animal models and say, yes, indeed,
link |
we can excite activity here.
link |
We can form receptive fields that are sensitive to
link |
or defined in terms of visual motion.
link |
But at no point could you exclude the possibility
link |
that everywhere else in the brain
link |
was also very interested in visual motion.
link |
By the way, I apologize to interrupt,
link |
but a tiny little tangent.
link |
You said animal models, just out of curiosity,
link |
from your perspective, how different is the human brain
link |
versus the other animals
link |
in terms of our ability to study the brain?
link |
Well, clearly, the further away you go from a human brain,
link |
the greater the differences,
link |
but not as remarkable as you might think.
link |
So people will choose their level of approximation
link |
to the human brain,
link |
depending upon the kinds of questions
link |
that they want to answer.
link |
So if you're talking about sort of canonical principles
link |
of microcircuitry, it might be perfectly okay
link |
to look at a mouse, indeed.
link |
You could even look at flies, worms.
link |
If, on the other hand, you wanted to look at
link |
the finer details of organization of visual cortex
link |
and V1, V2, these are designated patches of cortex
link |
that may do different things, indeed, do.
link |
You'd probably want to use a primate
link |
that looked a little bit more like a human,
link |
because there are lots of ethical issues
link |
in terms of the use of nonhuman primates
link |
to answer questions about human anatomy.
link |
But I think most people assume
link |
that most of the important principles are conserved
link |
in a continuous way, right from, well, yes,
link |
worms right through to you and me.
link |
So now returning to, so that was the early sort of ideas
link |
of studying the functional regions of the brain
link |
by if there's some damage to it,
link |
to try to infer that that part of the brain
link |
might be somewhat responsible for this type of function.
link |
So where does that lead us?
link |
What are the next steps beyond that?
link |
Right, well, I'll just actually just reverse a bit,
link |
come back to your sort of notion
link |
that the brain is a magic soup.
link |
That was actually a very prominent idea at one point,
link |
notions such as Lashley's law of mass action
link |
inherited from the observation that for certain animals,
link |
if you just took out spoonfuls of the brain,
link |
it didn't matter where you took these spoonfuls out,
link |
they always showed the same kinds of deficits.
link |
So it was very difficult to infer functional specialization
link |
purely on the basis of lesion deficit studies.
link |
But once we had the opportunity
link |
to look at the brain lighting up
link |
and it's literally it's sort of excitement, neuronal excitement
link |
when looking at this versus that,
link |
one was able to say, yes, indeed,
link |
these functionally specialized responses
link |
are very restricted and they're here or they're over there.
link |
If I do this, then this part of the brain lights up.
link |
And that became doable in the early 90s.
link |
In fact, shortly before with the advent
link |
of positron emission tomography.
link |
And then functional magnetic resonance imaging
link |
came along in the early 90s.
link |
And since that time, there has been an explosion
link |
of discovery, refinement, confirmation.
link |
There are people who believe that it's all in the anatomy.
link |
If you understand the anatomy,
link |
then you understand the function at some level.
link |
And many, many hypotheses were predicated
link |
on a deep understanding of the anatomy and the connectivity,
link |
but they were all confirmed
link |
and taken much further with neuroimaging.
link |
So that's what I meant by we've made an enormous amount
link |
of progress in this century indeed,
link |
and in relation to the previous century,
link |
by looking at these functionally selective responses.
link |
But that wasn't the whole story.
link |
So there's this sort of near phrenology,
link |
but finding bumps and hot spots in the brain
link |
that did this or that.
link |
The bigger question was, of course,
link |
the functional integration.
link |
How all of these regionally specific responses
link |
were orchestrated, how they were distributed,
link |
how did they relate to distributed processing
link |
and indeed representations in the brain.
link |
So then you turn to the more challenging issue
link |
of the integration, the connectivity.
link |
And then we come back to this beautiful,
link |
sparse, recurrent, hierarchical connectivity
link |
that seems characteristic of the brain
link |
and probably not many other organs.
link |
But nevertheless, we come back to this challenge
link |
of trying to figure out how everything is integrated.
link |
But what's your feeling?
link |
What's the general consensus?
link |
Have we moved away from the magic soup view of the brain?
link |
So there is a deep structure to it.
link |
And then maybe a further question.
link |
You said some people believe that the structure
link |
is most of it, that you can really get
link |
at the core of the function
link |
by just deeply understanding the structure.
link |
Where do you sit on that, do you?
link |
I think it's got some mileage to it, yes, yeah.
link |
So it's a worthy pursuit of going,
link |
of studying through imaging and all the different methods
link |
to actually study the structure.
link |
No, absolutely, yeah, yeah.
link |
Sorry, I'm just noting, you were accusing me
link |
of using lots of long words
link |
and then you introduced one there, which is deep,
link |
which is interesting.
link |
Because deep is the sort of millennial equivalent
link |
So if you've put deep in front of anything,
link |
not only are you very millennial and very trending,
link |
but you're also implying a hierarchical architecture.
link |
So it is a depth, which is, for me, the beautiful thing.
link |
That's right, the word deep kind of,
link |
yeah, exactly, it implies hierarchy.
link |
I didn't even think about that.
link |
That indeed, the implicit meaning
link |
of the word deep is hierarchy.
link |
So deep inside the onion is the center of your soul.
link |
Maybe briefly, if you could paint a picture
link |
of the kind of methods of neuroimaging,
link |
maybe the history which you were a part of,
link |
from statistical parametric mapping.
link |
I mean, just what's out there that's interesting
link |
for people maybe outside the field
link |
to understand of what are the actual methodologies
link |
of looking inside the human brain?
link |
Right, well, you can answer that question
link |
from two perspectives.
link |
Basically, it's the modality.
link |
What kind of signal are you measuring?
link |
And they can range from,
link |
and let's limit ourselves
link |
to sort of imaging based noninvasive techniques.
link |
So you've essentially got brain scanners,
link |
and brain scanners can either measure
link |
the structural attributes, the amount of water,
link |
the amount of fat, or the amount of iron
link |
in different parts of the brain,
link |
and you can make lots of inferences
link |
about the structure of the organ of the sort
link |
that you might have produced from an X ray,
link |
but a very nuanced X ray that is looking
link |
at this kind of property or that kind of property.
link |
So looking at the anatomy noninvasively
link |
would be the first sort of neuroimaging
link |
that people might want to employ.
link |
Then you move on to the kinds of measurements
link |
that reflect dynamic function,
link |
and the most prevalent of those fall into two camps.
link |
You've got these metabolic, sometimes hemodynamic,
link |
blood related signals.
link |
So these metabolic and or hemodynamic signals
link |
are basically proxies for elevated activity
link |
and message passing and neuronal dynamics
link |
in particular parts of the brain.
link |
Characteristically though, the time constants
link |
of these hemodynamic or metabolic responses
link |
to neural activity are much longer
link |
than the neural activity itself.
link |
And this is referring,
link |
forgive me for the dumb questions,
link |
but this would be referring to blood,
link |
like the flow of blood.
link |
Absolutely, absolutely.
link |
So there's a ton of,
link |
it seems like there's a ton of blood vessels in the brain.
link |
So what's the interaction between the flow of blood
link |
and the function of the neurons?
link |
Is there an interplay there or?
link |
Yup, yup, and that interplay accounts for several careers
link |
of world renowned scientists, yes, absolutely.
link |
So this is known as neurovascular coupling,
link |
is exactly what you said.
link |
It's how does the neural activity,
link |
the neuronal infrastructure, the actual message passing
link |
that we think underlies our capacity to perceive and act,
link |
how is that coupled to the vascular responses
link |
that supply the energy for that neural processing?
link |
So there's a delicate web of large vessels,
link |
arteries and veins, that gets progressively finer
link |
and finer in detail until it perfuses
link |
at a microscopic level,
link |
the machinery where little neurons lie.
link |
So coming back to this sort of onion perspective,
link |
we were talking before using the onion as a metaphor
link |
for a deep hierarchical structure,
link |
but also I think it's just anatomically quite
link |
a useful metaphor.
link |
All the action, all the heavy lifting
link |
in terms of neural computation is done
link |
on the surface of the brain,
link |
and then the interior of the brain is constituted
link |
by fatty wires, essentially, axonal processes
link |
that are enshrouded by myelin sheaths.
link |
And these, when you dissect them, they look fatty and white,
link |
and so it's called white matter,
link |
as opposed to the actual neuro pill,
link |
which does the computation constituted largely by neurons,
link |
and that's known as gray matter.
link |
So the gray matter is a surface or a skin
link |
that sits on top of this big ball,
link |
now we are talking magic soup,
link |
but a big ball of connections like spaghetti,
link |
very carefully structured with sparse connectivity
link |
that preserves this deep hierarchical structure,
link |
but all the action takes place on the surface,
link |
on the cortex of the onion, and that means
link |
that you have to supply the right amount of blood flow,
link |
the right amount of nutrient,
link |
which is rapidly absorbed and used by neural cells
link |
that don't have the same capacity
link |
that your leg muscles would have
link |
to basically spend their energy budget
link |
and then claim it back later.
link |
So one peculiar thing about cerebral metabolism,
link |
brain metabolism, is it really needs to be driven
link |
in the moment, which means you basically
link |
have to turn on the taps.
link |
So if there's lots of neural activity
link |
in one part of the brain, a little patch
link |
of a few millimeters, even less possibly,
link |
you really do have to water that piece
link |
of the garden now and quickly,
link |
and by quickly I mean within a couple of seconds.
link |
So that contains a lot of, hence the imaging
link |
could tell you a story of what's happening.
link |
Absolutely, but it is slightly compromised
link |
in terms of the resolution.
link |
So the deployment of these little microvessels
link |
that water the garden to enable the neural activity
link |
to play out, the spatial resolution
link |
is in order of a few millimeters,
link |
and crucially, the temporal resolution
link |
is the order of a few seconds.
link |
So you can't get right down and dirty
link |
into the actual spatial and temporal scale
link |
of neural activity in and of itself.
link |
To do that, you'd have to turn
link |
to the other big imaging modality,
link |
which is the recording of electromagnetic signals
link |
as they're generated in real time.
link |
So here, the temporal bandwidth, if you like,
link |
or the low limit on the temporal resolution
link |
is incredibly small, talking about milliseconds.
link |
And then you can get into the phasic fast responses
link |
that is in and of itself the neural activity,
link |
and start to see the succession or cascade
link |
of hierarchical recurrent message passing
link |
evoked by a particular stimulus.
link |
But the problem is you're looking
link |
at electromagnetic signals that have passed
link |
through an enormous amount of magic soup
link |
or spaghetti of collectivity, and through the scalp
link |
and the skull, and it's become spatially very diffuse.
link |
So it's very difficult to know where you are.
link |
So you've got this sort of catch 22.
link |
You can either use an imaging modality
link |
that tells you within millimeters
link |
which part of the brain is activated,
link |
but you don't know when,
link |
or you've got these electromagnetic EEG, MEG setups
link |
that tell you to within a few milliseconds
link |
when something has responded, but you're not aware.
link |
So you've got these two complementary measures,
link |
either indirect via the blood flow,
link |
or direct via the electromagnetic signals
link |
caused by neural activity.
link |
These are the two big imaging devices.
link |
And then the second level of responding to your question,
link |
what are the, from the outside,
link |
what are the big ways of using this technology?
link |
So once you've chosen the kind of neural imaging
link |
that you want to use to answer your set questions,
link |
and sometimes it would have to be both,
link |
then you've got a whole raft of analyses,
link |
time series analyses usually, that you can bring to bear
link |
in order to answer your questions
link |
or address your hypothesis about those data.
link |
And interestingly, they both fall
link |
into the same two camps we were talking about before,
link |
this dialectic between specialization and integration,
link |
differentiation and integration.
link |
So it's the cartography, the blobology analyses.
link |
I apologize, I probably shouldn't interrupt so much,
link |
but just heard a fun word, the blah.
link |
It's a neologism, which means the study of blobs.
link |
Are you being witty and humorous,
link |
or does the word blobology ever appear
link |
in a textbook somewhere?
link |
It would appear in a popular book.
link |
It would not appear in a worthy specialist journal.
link |
Yeah, I thought so.
link |
It's the fond word for the study of literally little blobs
link |
on brain maps showing activations.
link |
So the kind of thing that you'd see in the newspapers
link |
on ABC or BBC reporting the latest finding
link |
from brain imaging.
link |
Interestingly though, the maths involved
link |
in that stream of analysis does actually call upon
link |
the mathematics of blobs.
link |
So seriously, they're actually called Euler characteristics
link |
and they have a lot of fancy names in mathematics.
link |
We'll talk about it, about your ideas
link |
in free energy principle.
link |
I mean, there's echoes of blobs there
link |
when you consider sort of entities,
link |
mathematically speaking.
link |
Well, circumscribed, well defined,
link |
you entities of, well, from the free energy point of view,
link |
entities of anything, but from the point of view
link |
of the analysis, the cartography of the brain,
link |
these are the entities that constitute the evidence
link |
for this functional segregation.
link |
You have segregated this function in this blob
link |
and it is not outside of the blob.
link |
And that's basically the, if you were a map maker
link |
of America and you did not know its structure,
link |
the first thing were you doing constituting
link |
or creating a map would be to identify the cities,
link |
for example, or the mountains or the rivers.
link |
All of these uniquely spatially localizable features,
link |
possibly topological features have to be placed somewhere
link |
because that requires a mathematics of identifying
link |
what does a city look like on a satellite image
link |
or what does a river look like
link |
or what does a mountain look like?
link |
What would it, you know, what data features
link |
would evidence that particular top,
link |
you know, that particular thing
link |
that you wanted to put on the map?
link |
And they normally are characterized
link |
in terms of literally these blobs
link |
or these sort of, another way of looking at this
link |
is that a certain statistical measure
link |
of the degree of activation crosses a threshold
link |
and in crossing that threshold
link |
in the spatially restricted part of the brain,
link |
it creates a blob.
link |
And that's basically what statistical parametric mapping does.
link |
It's basically mathematically finessed blobology.
link |
Okay, so those are the,
link |
you kind of described these two methodologies for,
link |
one is temporally noisy, one is spatially noisy
link |
and you kind of have to play and figure out
link |
what can be useful.
link |
It'd be great if you can sort of comment.
link |
I got a chance recently to spend a day
link |
at a company called Neuralink
link |
that uses brain computer interfaces
link |
and their dream is to, well,
link |
there's a bunch of sort of dreams,
link |
but one of them is to understand the brain
link |
by sort of, you know, getting in there
link |
past the so called sort of factory wall,
link |
getting in there and be able to listen,
link |
communicate both directions.
link |
What are your thoughts about this,
link |
the future of this kind of technology
link |
of brain computer interfaces
link |
to be able to now have a window
link |
or direct contact within the brain
link |
to be able to measure some of the signals,
link |
to be able to sense signals,
link |
to understand some of the functionality of the brain?
link |
Ambivalent, my sense is ambivalent.
link |
So it's a mixture of good and bad
link |
and I acknowledge that freely.
link |
So the good bits, if you just look at the legacy
link |
of that kind of reciprocal but invasive
link |
your brain stimulation,
link |
I didn't paint a complete picture
link |
when I was talking about sort of the ways
link |
we understand the brain prior to neuroimaging.
link |
It wasn't just lesion deficit studies.
link |
Some of the early work, in fact,
link |
literally 100 years from where we're sitting
link |
at the institution of neurology,
link |
was done by stimulating the brain of say dogs
link |
and looking at how they responded
link |
with their muscles or with their salivation
link |
and imputing what that part of the brain must be doing.
link |
If I stimulate it and I vote this kind of response,
link |
then that tells me quite a lot
link |
about the functional specialization.
link |
So there's a long history of brain stimulation
link |
which continues to enjoy a lot of attention nowadays.
link |
Positive attention.
link |
Oh yes, absolutely.
link |
You know, deep brain stimulation for Parkinson's disease
link |
is now a standard treatment
link |
and also a wonderful vehicle
link |
to try and understand the neuronal dynamics
link |
underlying movement disorders like Parkinson's disease.
link |
Even interest in magnetic stimulation,
link |
stimulating the magnetic fields
link |
and will it work in people who are depressed, for example.
link |
Quite a crude level of understanding what you're doing,
link |
but there is historical evidence
link |
that these kinds of brute force interventions
link |
They, you know, it's a little bit like banging the TV
link |
when the valves aren't working properly,
link |
but it's still, it works.
link |
So, you know, there is a long history.
link |
Brain computer interfacing or BCI,
link |
I think is a beautiful example of that.
link |
It's sort of carved out its own niche
link |
and its own aspirations
link |
and there've been enormous advances within limits.
link |
Advances in terms of our ability to understand
link |
how the brain, the embodied brain,
link |
engages with the world.
link |
I'm thinking here of sensory substitution,
link |
the augmenting our sensory capacities
link |
by giving ourselves extra ways of sensing
link |
and sampling the world,
link |
ranging from sort of trying to replace lost visual signals
link |
through to giving people completely new signals.
link |
So, one of the, I think, most engaging examples of this
link |
is equipping people with a sense of magnetic fields.
link |
So you can actually give them magnetic sensors
link |
that enable them to feel,
link |
should we say, tactile pressure around their tummy,
link |
where they are in relation to the magnetic field of the Earth.
link |
And after a few weeks, they take it for granted.
link |
They integrate it, they embody this,
link |
simulate this new sensory information
link |
into the way that they literally feel their world,
link |
but now equipped with this sense of magnetic direction.
link |
So that tells you something
link |
about the brain's plastic potential
link |
to remodel and its plastic capacity
link |
to suddenly try to explain the sensory data at hand
link |
by augmenting the sensory sphere
link |
and the kinds of things that you can measure.
link |
Clearly, that's purely for entertainment
link |
and understanding the nature and the power of our brains.
link |
I would imagine that most BCI is pitched
link |
at solving clinical and human problems
link |
such as locked in syndrome, such as paraplegia,
link |
or replacing lost sensory capacities
link |
like blindness and deafness.
link |
So then we come to the negative part of my ambivalence,
link |
the other side of it.
link |
So I don't want to be deflationary
link |
because much of my deflationary comments
link |
is probably large out of ignorance than anything else.
link |
But generally speaking, the bandwidth
link |
and the bit rates that you get
link |
from brain computer interfaces as we currently know them,
link |
we're talking about bits per second.
link |
So that would be like me only being able to communicate
link |
with any world or with you using very, very, very slow Morse code.
link |
And it is not even within an order of magnitude
link |
near what we actually need for an inactive realization
link |
of what people aspire to when they think about
link |
sort of curing people with paraplegia or replacing sight
link |
despite heroic efforts.
link |
So one has to ask, is there a lower bound
link |
on the kinds of recurrent information exchange
link |
between a brain and some augmented or artificial interface?
link |
And then we come back to, interestingly,
link |
what I was talking about before,
link |
which is if you're talking about function
link |
in terms of inference, and I presume we'll get to that
link |
later on in terms of the free energy principle,
link |
then at the moment, there may be fundamental reasons
link |
to assume that is the case.
link |
We're talking about ensemble activity.
link |
We're talking about basically, for example,
link |
let's paint the challenge facing brain computer interfacing
link |
in terms of controlling another system
link |
that is highly and deeply structured,
link |
very relevant to our lives, very nonlinear,
link |
that rests upon the kind of nonequilibrium
link |
steady states and dynamics that the brain does,
link |
the weather, all right?
link |
So imagine you had some very aggressive satellites
link |
that could produce signals that could perturb
link |
some little parts of the weather system.
link |
And then what you're asking now is,
link |
can I meaningfully get into the weather
link |
and change it meaningfully and make the weather respond
link |
in a way that I want it to?
link |
You're talking about chaos control on a scale
link |
which is almost unimaginable.
link |
So there may be fundamental reasons
link |
why BCI, as you might read about it in a science fiction novel,
link |
aspirational BCI may never actually work
link |
in the sense that to really be integrated
link |
and be part of the system is a requirement
link |
that requires you to have evolved with that system,
link |
that you have to be part of a very delicately structured,
link |
deeply structured, dynamic, ensemble activity
link |
that is not like rewiring a broken computer
link |
or plugging in a peripheral interface adapter.
link |
It is much more like getting into the weather patterns
link |
or, come back to your magic soup,
link |
getting into the active matter
link |
and meaningfully relate that to the outside world.
link |
So I think there are enormous challenges there.
link |
So I think the example of the weather is a brilliant one.
link |
And I think you paint a really interesting picture
link |
and it wasn't as negative as I thought.
link |
It's essentially saying that it might be
link |
incredibly challenging, including the low bound
link |
of the bandwidth and so on.
link |
I kind of, so just to full disclosure,
link |
I come from the machine learning world.
link |
So my natural thought is the hardest part
link |
is the engineering challenge of controlling the weather,
link |
of getting those satellites up and running and so on.
link |
And once they are, then the rest is fundamentally
link |
the same approaches that allow you to win in the game of Go
link |
will allow you to potentially play in this soup,
link |
So I have a hope that sort of machine learning methods
link |
will help us play in this soup.
link |
But perhaps you're right that it is a biology
link |
and the brain is just an incredible system
link |
that may be almost impossible to get in.
link |
But for me, what seems impossible
link |
is the incredible mess of blood vessels
link |
that you also described without,
link |
we also value the brain.
link |
You can't make any mistakes, you can't damage things.
link |
So to me, that engineering challenge seems nearly impossible.
link |
One of the things I was really impressed by at Neuralink
link |
is just talking to brilliant neurosurgeons
link |
and the roboticists that made me realize
link |
that even though it seems impossible,
link |
if anyone can do it, it's some of these world class
link |
engineers that are trying to take it on.
link |
So I think the conclusion of our discussion here
link |
of this part is basically that the problem is really hard
link |
but hopefully not impossible.
link |
So if it's okay, let's start with the basics.
link |
So you've also formulated a fascinating principle,
link |
the free energy principle.
link |
Could we maybe start at the basics
link |
and what is the free energy principle?
link |
Well, in fact, the free energy principle
link |
inherits a lot from the building
link |
of these data analytic approaches
link |
to these very high dimensional time series
link |
you get from the brain.
link |
So I think it's interesting to acknowledge that.
link |
And in particular, the analysis tools
link |
that try to address the other side,
link |
which is a functional integration,
link |
so the connectivity analyses.
link |
So on the one hand, but I should also acknowledge
link |
it inherits an awful lot from machine learning as well.
link |
So the free energy principle is just a formal statement
link |
that the existential imperatives for any system
link |
that manages to survive in a changing world
link |
can be cast as an inference problem
link |
in the sense that you can interpret
link |
the probability of existing as the evidence that you exist.
link |
And if you can write down that problem of existence
link |
as a statistical problem,
link |
then you can use all the maths that has been developed
link |
for inference to understand and characterize
link |
the ensemble dynamics that must be in play
link |
in the service of that inference.
link |
So technically, what that means is
link |
you can always interpret anything that exists
link |
in virtue of being separate from the environment
link |
in which it exists as trying to minimize
link |
variational free energy.
link |
And if you're from the machine learning community,
link |
you will know that as a negative evidence lower bound
link |
or a negative elbow, which is the same as saying
link |
you're trying to maximize or it will look as if
link |
all your dynamics are trying to maximize
link |
the complement of that which is the marginal likelihood
link |
or the evidence for your own existence.
link |
So that's basically the free energy principle.
link |
But to even take a sort of a small step backwards,
link |
you said the existential imperative.
link |
There's a lot of beautiful poetic words here,
link |
but to put it crudely, it's a fascinating idea
link |
of basically just of trying to describe
link |
if you're looking at a blob,
link |
how do you know this thing is alive?
link |
What does it mean to be alive?
link |
What does it mean to exist?
link |
And so you can look at the brain,
link |
you can look at parts of the brain,
link |
or this is just a general principle
link |
that applies to almost any system.
link |
That's just a fascinating sort of philosophically
link |
at every level question and a methodology
link |
to try to answer that question.
link |
What does it mean to be alive?
link |
So that's a huge endeavor and it's nice
link |
that there's at least some,
link |
from some perspective, a clean answer.
link |
So maybe can you talk about that optimization view of it?
link |
So what's trying to be minimized, maximized?
link |
A system that's alive, what is it trying to minimize?
link |
Right, you've made a big move there.
link |
First of all, it's good to make big moves.
link |
But you've assumed that the thing exists
link |
in a state that could be living or nonliving.
link |
So I may ask you, what licenses you
link |
to say that something exists?
link |
That's why I use the word existential.
link |
It's beyond living, it's just existence.
link |
So if you drill down onto the definition
link |
of things that exist, then they have certain properties
link |
if you borrow the maths
link |
from nonequilibrium steady state physics
link |
that enable you to interpret their existence
link |
in terms of this optimization procedure.
link |
So it's good you introduced the word optimization.
link |
So what the free energy principle
link |
in its sort of most ambitious,
link |
but also most deflationary and simplest, says
link |
is that if something exists,
link |
then it must, by the mathematics
link |
of nonequilibrium steady state,
link |
exhibit properties that make it look
link |
as if it is optimizing a particular quantity.
link |
And it turns out that particular quantity
link |
happens to be exactly the same
link |
as the evidence lower bound in machine learning
link |
or Bayesian model evidence in Bayesian statistics.
link |
Or, and then I can list a whole other list
link |
of ways of understanding this key quantity,
link |
which is a bound on surprise or self information
link |
if you have information theory.
link |
There are a number of different perspectives
link |
It's just basically the log probability
link |
of being in a particular state.
link |
I'm telling this story as an honest,
link |
an attempt to answer your question.
link |
And I'm answering it as if I was pretending
link |
to be a physicist who was trying to understand
link |
the fundaments of nonequilibrium steady state.
link |
And I shouldn't really be doing that
link |
because the last time I was taught physics,
link |
What kind of systems,
link |
when you think about the free energy principle,
link |
what kind of systems are you imagining
link |
as a sort of more specific kind of case study?
link |
Yeah, I'm imagining a range of systems,
link |
but at its simplest, a single celled organism
link |
that can be identified from its eco niche
link |
or its environment.
link |
So at its simplest, that's basically
link |
what I always imagined in my head.
link |
And you may ask, well, is there any,
link |
how on earth can you even elaborate questions
link |
about the existence of a single drop of oil, for example?
link |
But there are deep questions there.
link |
Why doesn't the oil, why doesn't the thing,
link |
the interface between the drop of oil
link |
that contains an interior
link |
and the thing that is not the drop of oil,
link |
which is the solvent in which it is immersed,
link |
how does that interface persist over time?
link |
Why doesn't the oil just dissolve into solvent?
link |
So what special properties of the exchange
link |
between the surface of the oil drop
link |
and the external states in which it's immersed,
link |
if you're a physicist, say it would be the heat bath.
link |
You've got a physical system, an ensemble again,
link |
we're talking about density dynamics, ensemble dynamics,
link |
an ensemble of atoms or molecules immersed in the heat bath.
link |
But the question is, how did the heat bath get there?
link |
And why does it not dissolve?
link |
How is it maintaining itself?
link |
What actions is it?
link |
I mean, it's such a fascinating idea of a drop of oil
link |
and I guess it would dissolve in water,
link |
it wouldn't dissolve in water.
link |
Precisely, so why not?
link |
And how do you mathematically describe,
link |
I mean, it's such a beautiful idea.
link |
And also the idea of like, where does the thing,
link |
where does the drop of oil end and where does it begin?
link |
Right, so I mean, you're asking deep questions,
link |
deep in a nonmillennial sense here.
link |
In a hierarchical sense.
link |
But what you can do, so this is the deflationary part of it.
link |
Can I just qualify my answer by saying that normally
link |
when I'm asked this question,
link |
I answer from the point of view of a psychologist,
link |
we talk about predictive processing and predictive coding
link |
and the brain as an inference machine,
link |
but you haven't asked me from that perspective,
link |
I'm answering from the point of view of a physicist.
link |
So the question is not so much why,
link |
but if it exists, what properties must it display?
link |
So that's the deflationary part of the free energy principle.
link |
The free energy principle does not supply an answer
link |
as to why, it's saying if something exists,
link |
then it must display these properties.
link |
That's the sort of thing that's on offer.
link |
And it so happens that these properties it must display
link |
are actually intriguing and have this inferential gloss,
link |
this sort of self evidencing gloss that inherits on the fact
link |
that the very preservation of the boundary
link |
between the oil drop and the not oil drop
link |
requires an optimization of a particular function
link |
or a functional that defines the presence
link |
or the existence of this oil drop,
link |
which is why I started with existential imperatives.
link |
It is a necessary condition for existence
link |
that this must occur because the boundary
link |
basically defines the thing that's existing.
link |
So it is that self assembly aspect
link |
it's that you were hinting at in biology,
link |
sometimes known as autopoiesis
link |
in computational chemistry with self assembly.
link |
It's the, what does it look like?
link |
Sorry, how would you describe things
link |
that configure themselves out of nothing?
link |
The way they clearly demarcate themselves
link |
from the states or the soup in which they are immersed.
link |
So from the point of view of computational chemistry,
link |
for example, you would just understand that
link |
as a configuration of a macro molecule
link |
to minimize its free energy, its thermodynamic free energy.
link |
It's exactly the same principle that we've been talking about
link |
that thermodynamic free energy is just the negative elbow.
link |
It's the same mathematical construct.
link |
So the very emergence of existence, of structure, of form
link |
that can be distinguished from the environment
link |
or the thing that is not the thing
link |
necessitates the existence of an objective function
link |
that it looks as if it is minimizing.
link |
It's finding a free energy minima.
link |
And so just to clarify, I'm trying to wrap my head around.
link |
So the free energy principle says that if something exists,
link |
these are the properties it should display.
link |
So what that means is we can't just look,
link |
we can't just go into a soup and there's no mechanism.
link |
Free energy principle doesn't give us a mechanism
link |
to find the things that exist.
link |
Is that what's implying, is being implied
link |
that you can kind of use it to reason,
link |
to think about like, study a particular system
link |
and say, does this exhibit these qualities?
link |
That's an excellent question.
link |
But to answer that, I'd have to return
link |
to your previous question about what's the difference
link |
between living and nonliving things.
link |
Yes, well, actually, sorry.
link |
So yeah, maybe we can go there.
link |
Maybe we can go there, you kind of drew a line
link |
and forgive me for the stupid questions,
link |
but you kind of drew a line between living and existing.
link |
Is there an interesting sort of distinction?
link |
Yeah, I think there is.
link |
So things do exist, grains of sand,
link |
rocks on the moon, trees, you.
link |
So all of these things can be separated from the environment
link |
in which they are immersed.
link |
And therefore, they must at some level
link |
be optimizing their free energy,
link |
taking this sort of model evidence interpretation
link |
of this quantity that basically means
link |
they're self evidencing.
link |
Another nice little twist of phrase here
link |
is that you are your own existence proof,
link |
statistically speaking, which I don't think
link |
I said that, somebody did, but I love that phrase.
link |
You are your own existence proof.
link |
Yeah, so it's so existential, isn't it?
link |
I'm gonna have to think about that for a few days.
link |
That's a beautiful line.
link |
So the step through to answer your question
link |
about what's it good for,
link |
we'll go along the following lines.
link |
First of all, you have to define what it means
link |
to exist, which now, as you've rightly pointed out,
link |
you have to define what probabilistic properties
link |
must the states of something possess
link |
so it knows where it finishes.
link |
And then you write that down in terms
link |
of statistical dependencies, again, sparsity.
link |
Again, it's not what's connected or what's correlated
link |
or what depends upon, it's what's not correlated
link |
and what doesn't depend upon something.
link |
Again, it comes down to the deep structures,
link |
not in this instance, hierarchical,
link |
but the structures that emerge
link |
from removing connectivity and dependency.
link |
And in this instance, basically being able
link |
to identify the surface of the oil drop
link |
from the water in which it is immersed.
link |
And when you do that, you start to realize,
link |
well, there are actually four kinds of states
link |
in any given universe that contains anything.
link |
The things that are internal to the surface,
link |
the things that are external to the surface
link |
and the surface in and of itself,
link |
which is why I use a metaphor,
link |
a little single celled organism
link |
that has an interior and exterior
link |
and then the surface of the cell.
link |
And that's mathematically a Markov blanket.
link |
Just to pause, I'm in awe of this concept
link |
that there's the stuff outside the surface,
link |
stuff inside the surface and the surface itself,
link |
the Markov blanket.
link |
It's just the most beautiful kind of notion
link |
about trying to explore what it means
link |
to exist mathematically.
link |
I apologize, it's just a beautiful idea.
link |
But it came out of California, so that's.
link |
I changed my mind.
link |
I take it all back.
link |
So anyway, so you were just talking
link |
about the surface, about the Markov blanket.
link |
So this surface or these blanket states
link |
that are the, because they are now defined
link |
in relation to these independencies
link |
and what different states internal blanket
link |
or external states can,
link |
which ones can influence each other
link |
and which cannot influence each other.
link |
You can now apply standard results
link |
that you would find in non equilibrium physics
link |
or steady state or thermodynamics or hydrodynamics,
link |
usually out of equilibrium solutions
link |
and apply them to this partition.
link |
And what it looks like is if all the normal gradient flows
link |
that you would associate with any non equilibrium system
link |
apply in such a way that part of the Markov blanket
link |
and the internal states seem to be hill climbing
link |
or doing a gradient descent on the same quantity.
link |
And that means that you can now describe
link |
the very existence of this oil drop.
link |
You can write down the existence of this oil drop
link |
in terms of flows, dynamics, equations of motion,
link |
where the blanket states or part of them,
link |
we call them active states and the internal states
link |
now seem to be and must be trying to look
link |
as if they're minimizing the same function,
link |
which is a low probability of occupying these states.
link |
Interesting thing is that what would they be called
link |
if you were trying to describe these things?
link |
So what we're talking about are internal states,
link |
external states and blanket states.
link |
Now let's carve the blanket states
link |
into two sensory states and active states.
link |
Operationally, it has to be the case
link |
that in order for this carving up
link |
into different sets of states to exist,
link |
the active states, the Markov blanket
link |
cannot be influenced by the external states.
link |
And we already know that the internal states
link |
can't be influenced by the external states
link |
because the blanket separates them.
link |
So what does that mean?
link |
Well, it means the active states, the internal states
link |
are now jointly not influenced by external states.
link |
They only have autonomous dynamics.
link |
So now you've got a picture of an oil drop
link |
that has autonomy, it has autonomous states,
link |
it has autonomous states in the sense
link |
that there must be some parts of the surface of the oil drop
link |
that are not influenced by the external states
link |
and all the interior.
link |
And together, those two states endow
link |
even a little oil drop with autonomous states
link |
that look as if they are optimizing
link |
their variational free energy or their negative elbow,
link |
their moral evidence.
link |
And that would be an interesting intellectual exercise.
link |
And you could say, you could even go into the realms
link |
of panpsychism, that everything that exists
link |
is implicitly making inferences on self evidencing.
link |
Now we make the next move, but what about living things?
link |
I mean, so let me ask you,
link |
what's the difference between an oil drop
link |
and a little tadpole or a little lava or a plankton?
link |
The picture was just painted of an oil drop.
link |
Just immediately in a matter of minutes
link |
took me into the world of panpsychism,
link |
where you've just convinced me,
link |
made me feel like an oil drop is a living,
link |
it's certainly an autonomous system,
link |
but almost a living system.
link |
So it has sensory capabilities and acting capabilities
link |
and it maintains something.
link |
So what is the difference between that
link |
and something that we traditionally
link |
think of as a living system?
link |
That it could die or it can't,
link |
I mean, yeah, mortality, I'm not exactly sure.
link |
I'm not sure what the right answer there is
link |
because they can move,
link |
like movement seems like an essential element
link |
to being able to act in the environment,
link |
but the oil drop is doing that.
link |
The oil drop will be moved,
link |
but does it in and of itself move autonomously?
link |
Well, the surface is performing actions
link |
that maintain its structure.
link |
Yeah, you're being too clever.
link |
I was, I had in mind a passive little oil drop
link |
that's sitting there at the bottom
link |
on the top of a glass of water.
link |
What I'm trying to say is you're absolutely right.
link |
So where does that movement come from?
link |
If it comes from the inside,
link |
then you've got, I think, something that's living.
link |
What do you mean from the inside?
link |
What I mean is that the internal states
link |
that can influence the active states,
link |
where the active states can influence,
link |
but they're not influenced by the external states,
link |
can cause movement.
link |
So there are two types of oil drops, if you like.
link |
There are oil drops where the internal states
link |
are so random that they average themselves away,
link |
and the thing cannot, on average,
link |
when you do the averaging, move.
link |
So a nice example of that would be the Sun.
link |
The Sun certainly has internal states.
link |
There's lots of intrinsic autonomous activity going on,
link |
but because it's not coordinated,
link |
because it doesn't have the deep, in the millennial sense,
link |
the hierarchical structure that the brain does,
link |
there is no overall mode or pattern or organization
link |
that expresses itself on the surface
link |
that allows it to actually swim.
link |
It can certainly have a very active surface,
link |
but en masse, at the scale of the actual surface of the Sun,
link |
the average position of that surface cannot, in itself, move,
link |
because the internal dynamics are more like a hot gas.
link |
They are literally like a hot gas,
link |
whereas your internal dynamics are much more structured
link |
and deeply structured,
link |
and now you can express on your active states
link |
with your muscles and your secretory organs,
link |
your autonomic nervous system and its effectors,
link |
you can actually move, and that's all you can do.
link |
And that's something which,
link |
if you haven't thought of it like this before,
link |
I think it's nice to just realize
link |
there is no other way that you can change the universe
link |
other than simply moving.
link |
Whether that moving is articulating with my voice box
link |
or walking around or squeezing juices
link |
out of my secretory organs,
link |
there's only one way you can change the universe.
link |
And the fact that you do so nonrandomly makes you alive.
link |
Yeah, so it's that nonrandomness.
link |
And that would be manifested,
link |
we realize, in terms of essentially swimming,
link |
essentially moving, changing one's shape,
link |
a morphogenesis that is dynamic and possibly adaptive.
link |
So that's what I was trying to get at
link |
between the difference between the oil drop
link |
and the little tadpole.
link |
The tadpole is moving around.
link |
Its active states are actually changing the external states.
link |
And there's now a cycle,
link |
an action perception cycle, if you like,
link |
a recurrent dynamic that's going on
link |
that depends upon this deeply structured autonomous behavior
link |
that rests upon internal dynamics
link |
that are not only modeling
link |
the data impressed upon their surface or the blanket states,
link |
but they are actively resampling those data by moving.
link |
They're moving towards chemical gradients and chemotaxis.
link |
So they've gone beyond just being good little models
link |
of the kind of world they live in.
link |
For example, an oil droplet could, in a panpsychic sense,
link |
be construed as a little being
link |
that has now perfectly inferred.
link |
It's a passive, nonliving oil drop
link |
living in a bowl of water.
link |
But to now equip that oil drop with the ability to go out
link |
and test that hypothesis about different states of beings.
link |
So it can actually push its surface over there, over there,
link |
and test for chemical gradients,
link |
or then you start to move to a much more lifelike form.
link |
This is all fun, theoretically interesting,
link |
but it actually is quite important
link |
in terms of reflecting what I have seen
link |
since the turn of the millennium,
link |
which is this move towards an inactive
link |
and embodied understanding of intelligence.
link |
And you say you're from machine learning.
link |
So what that means,
link |
the central importance of movement,
link |
I think has yet to really hit machine learning.
link |
It certainly has now diffused itself throughout robotics.
link |
And perhaps you could say certain problems in active vision
link |
where you actually have to move the camera
link |
to sample this and that.
link |
But machine learning of the data mining deep learning sort
link |
simply hasn't contended with this issue.
link |
What it's done, instead of dealing with the movement problem
link |
and the active sampling of data,
link |
it's just said, we don't need to worry about,
link |
we can see all the data because we've got big data.
link |
So we can ignore movement.
link |
So that for me is an important omission
link |
in current machine learning.
link |
The current machine learning is much more like the oil drop.
link |
But an oil drop that enjoys exposure
link |
to nearly all the data that it will ever need to be exposed to,
link |
as opposed to the tadpoles swimming out
link |
to find the right data.
link |
For example, it likes food.
link |
That's a good hypothesis.
link |
Let's test it out.
link |
Let's go and move and ingest food, for example,
link |
and see is that evidence that I'm the kind of thing
link |
that likes this kind of food.
link |
So the next natural question, and forgive this question,
link |
but if we think of sort of even artificial intelligence
link |
systems, which I just painted a beautiful picture
link |
of existence and life.
link |
So do you ascribe, do you find within this framework
link |
a possibility of defining consciousness
link |
or exploring the idea of consciousness?
link |
Like what, you know, self awareness
link |
and expand it to consciousness?
link |
How can we start to think about consciousness
link |
within this framework?
link |
Well, yeah, I think it's possible to think about it,
link |
whether you'll get it.
link |
Get anywhere is another question.
link |
And again, I'm not sure that I'm licensed
link |
to answer that question.
link |
I think you'd have to speak to a qualified philosopher
link |
to get a definitive answer to that.
link |
But certainly, there's a lot of interest
link |
in using not just these ideas, but related ideas
link |
from information theory to try and tie down
link |
the maths and the calculus and the geometry of consciousness,
link |
either in terms of sort of a minimal consciousness,
link |
even less than a minimal selfhood.
link |
And what I'm talking about is the ability, effectively,
link |
to plan, to have agency.
link |
So you could argue that a virus does have a form of agency
link |
in virtue of the way that it selectively
link |
finds hosts and cells to live in and moves around.
link |
But you wouldn't endow it with the capacity
link |
to think about planning and moving in a purposeful way
link |
where it countenances the future.
link |
Whereas you might an ant.
link |
You might think an ant's not quite as unconscious
link |
It certainly seems to have a purpose.
link |
It talks to its friends en route during its foraging.
link |
It has a different kind of autonomy, which is biotic,
link |
but beyond a virus.
link |
So there's something about, so there's
link |
some line that has to do with the complexity of planning
link |
that may contain an answer.
link |
I mean, it would be beautiful if we
link |
can find a line beyond which we could say a being is conscious.
link |
These are wonderful lines that we've drawn with existence,
link |
life, and consciousness.
link |
Yes, it will be very nice.
link |
One little wrinkle there, and this
link |
is something I've only learned in the past few months,
link |
is the philosophical notion of vagueness.
link |
So you're saying it would be wonderful to draw a line.
link |
I had always assumed that that line at some point
link |
would be drawn until about four months ago,
link |
and the philosopher taught me about vagueness.
link |
So I don't know if you've come across this,
link |
but it's a technical concept.
link |
And I think most revealingly illustrated with,
link |
at what point does a pile of sand become a pile?
link |
Is it one grain, two grains, three grains, or four grains?
link |
So at what point would you draw the line
link |
between being a pile of sand and a collection of grains of sand?
link |
In the same way, is it right to ask,
link |
where would I draw the line between conscious
link |
And it might be a vague concept.
link |
Having said that, I agree with you entirely.
link |
Systems that have the ability to plan.
link |
So just technically, what that means
link |
is your inferential self evidencing,
link |
by which I simply mean the thermodynamics and gradient
link |
flows that underwrite the preservation of your oil
link |
droplet like form, can be described
link |
as an optimization of log Bayesian model
link |
evidence, your elbow.
link |
That self evidencing must be evidence
link |
for a model of what's causing the sensory impressions
link |
on the sensory part of your surface or your Markov
link |
If that model is capable of planning,
link |
it must include a model of the future consequences
link |
of your active states or your action, just planning.
link |
So we're now in the game of planning as inference.
link |
Now notice what we've made, though.
link |
We've made quite a big move away from big data and machine
link |
learning, because again, it's the consequences of moving.
link |
It's the consequences of selecting those data or those
link |
data or looking over there.
link |
And that tells you immediately that even
link |
to be a contender for a conscious artifact or a strong
link |
AI or generalized, I don't know what that's called nowadays,
link |
then you've got to have movement in the game.
link |
And furthermore, you've got to have a generative model
link |
of the sort you might find in, say, a variational auto
link |
encoder that is thinking about the future conditioned
link |
upon different courses of action.
link |
Now that brings a number of things to the table, which
link |
now you start to think, well, those
link |
have got all the right ingredients
link |
to talk about consciousness.
link |
I've now got to select among a number of different courses
link |
of action into the future as part of planning.
link |
I've now got free will.
link |
The act of selecting this course of action or that policy
link |
or that policy or that action suddenly
link |
makes me into an inference machine,
link |
a self evidencing artifact that now
link |
looks as if it's selecting amongst different alternative
link |
ways forward, as I actively swim here or swim there
link |
or look over here, look over there.
link |
So I think you've now got to a situation
link |
if there is planning in the mix.
link |
You're now getting much closer to that line
link |
if that line were ever to exist.
link |
I don't think it gets you quite as far as self aware, though.
link |
And then you have to, I think, grapple with the question,
link |
how would formally write down a calculus or a maths
link |
of self awareness?
link |
I don't think it's impossible to do.
link |
But I think there would be pressure on you
link |
to actually commit to a formal definition of what
link |
you mean by self awareness.
link |
I think most people that I know would probably
link |
say that a goldfish, a pet fish, was not self aware.
link |
They would probably argue about their favorite cat,
link |
but would be quite happy to say that their mom was self aware.
link |
I mean, but that might very well connect
link |
to some level of complexity with planning.
link |
It seems like self awareness is essential for complex planning.
link |
Do you want to take that further?
link |
Because I think you're absolutely right.
link |
Again, the line is unclear, but it
link |
seems like integrating yourself into the world,
link |
into your planning, is essential for constructing complex plans.
link |
So mathematically describing that in the same elegant way
link |
as you have with the free energy principle might be difficult.
link |
I don't think that, well, perhaps we should just,
link |
can we just go back?
link |
That's a very important answer you gave.
link |
And I think if I just unpacked it,
link |
you'd see the truisms that you've just exposed for us.
link |
But let me, sorry, I'm mindful that I didn't answer
link |
your question before.
link |
Well, what's the free energy principle good for?
link |
Is it just a pretty theoretical exercise
link |
to explain nonequilibrium steady states?
link |
It does nothing more for you than that.
link |
It can be regarded, it's going to sound very arrogant,
link |
but it is of the sort of theory of natural selection,
link |
or a hypothesis of natural selection.
link |
Beautiful, undeniably true, but tells you
link |
absolutely nothing about why you have legs and eyes.
link |
It tells you nothing about the actual phenotype,
link |
and it wouldn't allow you to build something.
link |
So the free energy principle by itself
link |
is as vacuous as most tautological theories.
link |
And by tautological, of course,
link |
I'm talking to the theory of natural,
link |
the survival of the fittest.
link |
What's the fittest of those that survive?
link |
Why do they cycle?
link |
It just goes around in circles.
link |
In a sense, the free energy principle has that same
link |
deflationary tautology under the hood.
link |
It's a characteristic of things that exist.
link |
Why do they exist?
link |
Because they minimize their free energy.
link |
Why do they minimize their free energy?
link |
Because they exist.
link |
And you just keep on going round and round and round.
link |
But the practical thing,
link |
which you don't get from natural selection,
link |
but you could say has now manifest in things
link |
like differential evolution or genetic algorithms
link |
and MCMC, for example, in machine learning.
link |
The practical thing you can get is,
link |
if it looks as if things that exist
link |
are trying to have density dynamics
link |
and look as though they're optimizing
link |
a variational free energy,
link |
and a variational free energy has to be
link |
a functional of a generative model,
link |
a probabilistic description of causes and consequences,
link |
causes out there, consequences in the sensorium
link |
on the sensory parts of the Markov blanket,
link |
then it should, in theory, be possible
link |
to write down the generative model,
link |
work out the gradients,
link |
and then cause it to autonomously self evidence.
link |
So you should be able to write down oil droplets.
link |
You should be able to create artifacts
link |
where you have supplied the objective function
link |
that supplies the gradients,
link |
that supplies the self organizing dynamics
link |
to non equilibrium steady state.
link |
So there is actually a practical application
link |
of the free energy principle
link |
when you can write down your required evidence
link |
in terms of, well, when you can write down
link |
the generative model that is the thing
link |
that has the evidence.
link |
The probability of these sensory data
link |
or this data, given that model,
link |
is effectively the thing that the elbow
link |
or the variational free energy bounds or approximates.
link |
That means that you can actually write down the model
link |
and the kind of thing that you want to engineer,
link |
the kind of AGI or artificial general intelligence
link |
that you want to manifest probabilistically,
link |
and then you engineer, a lot of hard work,
link |
but you would engineer a robot and a computer
link |
to perform a gradient descent on that objective function.
link |
So it does have a practical implication.
link |
Now, why am I wittering on about that?
link |
It did seem relevant to, yes.
link |
So what kinds of, so the answer to,
link |
would it be easier or would it be hard?
link |
Well, mathematically, it's easy.
link |
I've just told you all you need to do
link |
is write down your perfect artifact,
link |
probabilistically, in the form
link |
of a probabilistic generative model,
link |
a probability distribution over the causes
link |
and consequences of the world
link |
in which this thing is immersed.
link |
And then you just engineer a computer and a robot
link |
to perform a gradient descent on that objective function.
link |
But of course, the big problem
link |
is writing down the generative model.
link |
So that's where the heavy lifting comes in.
link |
So it's the form and the structure of that generative model
link |
which basically defines the artifact that you will create
link |
or, indeed, the kind of artifact that has self awareness.
link |
So that's where all the hard work comes,
link |
very much like natural selection doesn't tell you
link |
in the slightest why you have eyes.
link |
So you have to drill down on the actual phenotype,
link |
the actual generative model.
link |
So with that in mind, what did you tell me
link |
that tells me immediately the kinds of generative models
link |
I would have to write down in order to have self awareness?
link |
What you said to me was I have to have a model
link |
that is effectively fit for purpose
link |
for this kind of world in which I operate.
link |
And if I now make the observation
link |
that this kind of world is effectively largely populated
link |
by other things like me, i.e. you,
link |
then it makes enormous sense
link |
that if I can develop a hypothesis
link |
that we are similar kinds of creatures,
link |
in fact, the same kind of creature,
link |
but I am me and you are you,
link |
then it becomes, again, mandated to have a sense of self.
link |
So if I live in a world
link |
that is constituted by things like me,
link |
basically a social world, a community,
link |
then it becomes necessary now for me to infer
link |
that it's me talking and not you talking.
link |
I wouldn't need that if I was on Mars by myself
link |
or if I was in the jungle as a feral child.
link |
If there was nothing like me around,
link |
there would be no need to have an inference
link |
at a hypothesis, oh yes, it is me
link |
that is experiencing or causing these sounds
link |
and it is not you.
link |
It's only when there's ambiguity in play
link |
induced by the fact that there are others in that world.
link |
So I think that the special thing about self aware artifacts
link |
is that they have learned to, or they have acquired,
link |
or at least are equipped with, possibly by evolution,
link |
generative models that allow for the fact
link |
there are lots of copies of things like them around,
link |
and therefore they have to work out it's you and not me.
link |
I've never thought of that.
link |
I never thought of that, that the purpose
link |
of the really usefulness of consciousness
link |
or self awareness in the context of planning
link |
existing in the world is so you can operate
link |
with other things like you, and like you could,
link |
it doesn't have to necessarily be human.
link |
It could be other kind of similar creatures.
link |
Absolutely, well, we view a lot of our attributes
link |
into our pets, don't we?
link |
Or we try to make our robots humanoid.
link |
And I think there's a deep reason for that,
link |
that it's just much easier to read the world
link |
if you can make the simplifying assumption
link |
that basically you're me, and it's just your turn to talk.
link |
I mean, when we talk about planning,
link |
when you talk specifically about planning,
link |
the highest, if you like, manifestation or realization
link |
of that planning is what we're doing now.
link |
I mean, the human condition doesn't get any higher
link |
than this talking about the philosophy of existence
link |
and the conversation.
link |
But in that conversation, there is a beautiful art
link |
of turn taking and mutual inference, theory of mind.
link |
I have to know when you wanna listen.
link |
I have to know when you want to interrupt.
link |
I have to make sure that you're online.
link |
I have to have a model in my head
link |
of your model in your head.
link |
That's the highest, the most sophisticated form
link |
of generative model, where the generative model
link |
actually has a generative model
link |
of somebody else's generative model.
link |
And I think that, and what we are doing now evinces
link |
the kinds of generative models
link |
that would support self awareness,
link |
because without that, we'd both be talking over each other,
link |
or we'd be singing together in a choir.
link |
That's not a brilliant analogy for what I'm trying to say,
link |
but yeah, we wouldn't have this discourse.
link |
We wouldn't have this.
link |
Yeah, the dance of it.
link |
Yeah, that's right.
link |
As I interrupt, I mean, that's beautifully put.
link |
I'll re listen to this conversation many times.
link |
There's so much poetry in this, and mathematics.
link |
Let me ask the silliest, or perhaps the biggest question
link |
as a last kind of question.
link |
We've talked about living in existence
link |
and the objective function under which
link |
these objects would operate.
link |
What do you think is the objective function
link |
What's the meaning of life?
link |
What do you think is the, for you, perhaps,
link |
the purpose, the source of fulfillment,
link |
the source of meaning for your existence,
link |
as one blob in this soup?
link |
I'm tempted to answer that, again, as a physicist,
link |
until it's the free energy I expect
link |
consequent upon my behavior.
link |
So technically, we could get a really interesting
link |
conversation about what that comprises
link |
in terms of searching for information,
link |
resolving uncertainty about the kind of thing that I am.
link |
But I suspect that you want a slightly more personal
link |
and fun answer, but which can be consistent with that.
link |
And I think it's reassuringly simple
link |
and hops back to what you were taught as a child,
link |
that you have certain beliefs about the kind of creature
link |
and the kind of person you are.
link |
And all that self evidencing means,
link |
all that minimizing variational free energy
link |
in an inactive and embodied way,
link |
means is fulfilling the beliefs about
link |
what kind of thing you are.
link |
And of course, we're all given those scripts,
link |
those narratives, at a very early age,
link |
usually in the form of bedtime stories or fairy stories
link |
that I'm a princess and I'm gonna meet a beast
link |
who's gonna transform and he's gonna be a prince.
link |
And so the narratives are all around you
link |
from your parents to the friends
link |
to the society feeds these stories.
link |
And then your objective function is to fulfill.
link |
Exactly, that narrative that has been encultured
link |
by your immediate family, but as you say,
link |
also the sort of the culture in which you grew up
link |
and you create for yourself.
link |
I mean, again, because of this active inference,
link |
this inactive aspect of self evidencing,
link |
not only am I modeling my environment,
link |
my eco niche, my external states out there,
link |
but I'm actively changing them all the time
link |
and doing the same back, we're doing it together.
link |
So there's a synchrony that means that I'm creating
link |
my own culture over different timescales.
link |
So the question now is for me being very selfish,
link |
what scripts were I given?
link |
It basically was a mixture between Einstein and shark homes.
link |
So I smoke as heavily as possible,
link |
try to avoid too much interpersonal contact,
link |
enjoy the fantasy that you're a popular scientist
link |
who's gonna make a difference in a slightly quirky way.
link |
So that's what I grew up on.
link |
My father was an engineer and loved science
link |
and he loved sort of things like Sir Arthur Edmonds,
link |
Spacetime and Gravitation, which was the first
link |
understandable version of general relativity.
link |
So all the fairy stories I was told as I was growing up
link |
were all about these characters.
link |
I'm keeping the Hobbit out of this
link |
because that doesn't quite fit my narrative.
link |
There's a journey of exploration, I suppose, of sorts.
link |
So yeah, I've just grown up to be what I imagine
link |
a mild mannered Sherlock Holmes slash Albert Einstein
link |
would do in my shoes.
link |
And you did it elegantly and beautifully.
link |
Carl was a huge honor talking today, it was fun.
link |
Thank you so much for your time.
link |
No, thank you. Appreciate it.
link |
Thank you for listening to this conversation
link |
with Carl Friston and thank you
link |
to our presenting sponsor, Cash App.
link |
Please consider supporting the podcast
link |
by downloading Cash App and using code LexPodcast.
link |
If you enjoy this podcast, subscribe on YouTube,
link |
review it with five stars on Apple Podcast,
link |
support on Patreon, or simply connect with me on Twitter
link |
And now let me leave you with some words from Carl Friston.
link |
Your arm moves because you predict it will
link |
and your motor system seeks to minimize prediction error.
link |
Thank you for listening and hope to see you next time.