Anil Seth的《意識的神經科學》。 (The Neuroscience of Consciousness with Anil Seth)

Subtitles section Play video

[APPLAUSE]
- Two out of the three fundamental mysteries
about our place in the universe have already been resolved.
The first is literally about our place in the universe.
Many years ago Copernicus told us that we were not at its
centre, that we were just a tiny dot suspended in the abyss.
This is an image of the earth taken from the probe Voyager 1
as it was leaving the solar system from about six
billion kilometres away.
All of human history, all of the history of life on Earth,
has taken place on that pale blue dot.
The second mystery, Darwin then revealed
that we humans are just one branch, or one twig,
of a beautifully rich and delicate evolutionary tree.
And that much of the machinery of life
is shared even with the lowliest of our fellow creatures.
The third mystery is that of consciousness,
our inner universe.
Now earlier this year, for the third time in my life,
I ceased to exist.
As the propofol anaesthetic flowed from the cannula
in my wrist into my bloodstream and then into my brain,
there was a falling apart.
A blackness.
An absence.
And then, I was back.
Drowsy and disoriented, but definitely there.
And when you wake from a deep sleep,
you might be confused what time it is, especially
in flying somewhere, but you'll know that some time has passed.
There seems to be some basic continuity
between your consciousness then, and your consciousness now.
But coming around from a general anaesthetic,
it could have been five minutes.
It could have been five hours.
It could have been five days, or five years.
I was simply not there.
A premonition of the oblivion of death.
And general anaesthesia doesn't just work on your brain.
It doesn't just work on your mind.
It works on your consciousness.
By altering the delicate electrochemical circuitry
inside your head, the basic ground state
of what it is to be is temporarily abolished.
And in this process lies one of the greatest remaining
mysteries in science and philosophy.
How does consciousness happen?
Why is life in the first person?
It is going away, and coming back.
The modern incarnation of this problem
is usually traced to Descartes, who in the 17th century
distinguished between matter stuff, res extensa, the stuff
that these desks are made of, that clothes are made of.
But also the brains and bodies and made of, material stuff.
And res cogitans, the stuff of thought, of feelings.
The stuff of consciousness.
And in making this distinction, he gave rise
to the now infamous mind/body problem,
and life has never been simple ever since.
But Descartes actually generated even more mischief
with his doctrine of the beast machine,
which I'm going to mention now, because it anticipates where
I'm going to end up as the bell rings when I finish in an hour.
Before Descartes, people commonly believed in something
called the great chain of being, with rocks and plants
at one end, and other non-human animals, a bit higher
up than humans, and then angels and gods at the very top.
And this great scale of being was also
a scale of moral virtue, so that humans had more moral virtues
than animals and plants, and then rocks and so on.
Now Descartes, in making this division between mind
and matter, argued that only humans had minds, and therefore
moral status, while other animals didn't have minds.
They were merely physiological machines, or beast machines,
morally equivalent to plants, and to rocks.
And in this view, the physiological mechanisms
that give rise to the property of being alive
were not relevant to the presence
of mind or consciousness.
Now I'm going to propose, at the end of this talk, the opposite.
That our conscious sense of self arises because of, and not
in spite of, the fact that we, too, are beast machines.
So to get there, let's return to the apparent mystery
of consciousness.
Now as recently as 1989, which is quite a while ago, but not
that long ago, Stuart Sutherland,
who was founding professor of experimental psychology
at my university of Sussex, had this to say.
"Consciousness is a fascinating but elusive phenomenon.
It is impossible to specify what it is, what it does,
or why it evolved.
Nothing worth reading has been written on it."
[LAUGHTER]
It's quite a pessimistic point of view.
And that may have been true then.
I don't think it was true then, but in any case
things have changed a lot since.
And more or less, about the time that Sutherland
made these remarks, we can see the birth, or the rebirth,
of the study of consciousness within the neurosciences.
And a good landmark is this paper
by Francis Crick and Christof Koch, published in 1990.
And they start their paper by saying that it is remarkable
that most of the work in cognitive sciences,
and the neurosciences, makes no reference to consciousness
or awareness at all.
And then they go on to propose their own theory
of what the neural correlates of consciousness are.
What it is in the brain that is responsible for being
conscious.
And since then, over the last 25 years,
there's been first a trickle, and now a deluge of research
on the brain basis of conscious experience.
Some of this work is being carried out
in my laboratory, the Sackler Centre, the consciousness
science, which was founded six years ago with Hugo
Critchley, my co-director.
And there are now even specialised academic journals,
The Neuroscience of Consciousness,
which I started last year with Oxford University Press.
And this is a real change of the tide.
When I started out more than 20 years ago,
it was thought to be a very-- it was thought
to be career suicide to want to study consciousness,
scientifically.
And it may still be, we don't know.
Let's see.
So while the brain basis of consciousness
is still a mystery, it is, in some sense,
an accessible mystery.
And the author, Mark Haddon, put this very nicely, I think.
He said the raw material of consciousness
is not on the other side of the universe.
It didn't happen 14 billion years ago.
And it's not squirrelled away deep inside an atom.
The raw material of consciousness
is right here, inside your head, and you can
hold the brain in your hands.
But the brain won't deliver its secrets very easily.
What's extraordinary about the brain
is not so much the number of neurons,
though there are about 90 billion.
It's not even the number of connections,
though there are so many that if you counted one every second,
it would take you about three million years
to finish counting.
What's truly extraordinary are the patterns
of connectivity, which to a large extent,
are still not known, but within which are inscribed everything
that makes you, you.
The challenge is then this, at least the the way I see it.
How can the structure and dynamics
of the brain, in connection with the body and the environment,
account for the subjective phenomenological properties
of consciousness?
And considering things this way, we
come up against what the philosopher David
Chalmers has often called the hard problem of consciousness.
And the idea is this.
There is an easy problem.
The easy problem is to understand how the combined
operations of the brain and the body give rise to perception,
to cognition, to thinking, to learning, to behaviour.
How the brain works, in other words.
The hard problem is to understand
why and how any of this should have anything
to do with consciousness at all.
Why aren't we just robots, or philosophical zombies,
without any in a universe?
Now there's a tempting intuition here,
which is that, even if we solve the hard problem, even if we
solve the easy problem, the hard problem would still remain
as mysterious as it seems now.
But this just seems wrong-headed to me.
It may not be necessary to explain why consciousness
exists at all, in order to make progress in understanding
its material basis.
And this for me, is the real problem of consciousness;
how to account for its various properties in terms
of biological mechanisms without pretending that it doesn't
exist at all, as you do if you solve the easy problem,
and without trying to account for why it's
parts of the universe in the first place, which
is the hard problem.
And in the history of science, we've
been somewhere similar before.
It's hard to say if it's exactly the same situation.
But in our understanding of life,
eminent biochemists of the time found it entirely mysterious
how biological mechanisms could give rise
to the property of being alive.
And there were proposed of things,
like elan vital and essence vital,
and all sorts of other stuff.
And although we don't yet understand everything
about life, this initial sense of mystery about life
has, to a large extent, dissolved
as biologists have just got on with the business
of understanding the properties of living systems in terms
of mechanisms.
An important part of this story was
the realisation that life is not just one thing,
but rather a constellation of partially dependent, partially
separable, processes, like metabolism, homeostasis,
and reproduction.
In the same way, to make progress
on the real problem of consciousness,
it can be useful to distinguish different aspects or dimensions
of what it is to be conscious.
The space of possible minds, if you like.
And one simple classification is into conscious level,
which is the property of being conscious at all.
For example, the difference between being
in a dreamless sleep, or under general anaesthesia,
and being awake and conscious as you are now.
And the conscious content, when you are conscious,
you're conscious of something.
The myriad of sights, sounds, smells, emotions, feelings,
and beliefs that populate your inner universe at any one time.
And one thing you are conscious of when you are conscious,
is the experience, the specific experience, of being you,
and this is conscious self.
And it's the third dimension of consciousness.
Now I don't claim these distinctions mark
completely independent aspects of what it is to be conscious,
but they're a pragmatically useful way of breaking down
the problem a bit.
So let's start with conscious level.
What are the fundamental brain mechanisms
that underlie our ability to be conscious at all?
And we can think of this, at least a first approximation,
as a scale from being completely unconscious,
as if you were in a coma, or under general anaesthesia,
to being awake, alert, and fully conscious as you are now.
And there's various states in between being drowsy,
being mildly sedated and so on.
What's important is that, while being conscious and being awake
often go together, this is not always the case.
For instance, when you are dreaming you are asleep,
but you are having conscious experiences.
The conscious experience of your dreams.
And on the other side of this diagram,
there are pathological states, like the vegetative state,
where physiologically you will go through sleep/wake cycles,
but there is nobody at home.
There is no consciousness happening.
So what are the specific mechanisms
that underlie being conscious and not simply being
physiologically awake?
Well there are a number of possibilities.
Is it the number of neurons?
Well actually, probably not.
There are more neurons in your cerebellum,
this bit at the back of your brain,
than in the rest of your brain put together.
In fact are about four times more neurons in your cerebellum
than in the rest of your cortex.
But if you have damage to your cerebellum, yeah,
you'll have some problems with coordination
and other things, some cognitive problems,
but you won't lose consciousness.
It's not just the number of neurons.
Doesn't seem to be any particular region.
In fact, there are regions that, if you suffer damage,
you will permanently lose consciousness;
in thalamina, the nuclei, and the thalamus
deep inside the brain.
But these seems to be more like on/off switches
than actual generators of conscious experience.
It's not even neural activity, at least not
simple kinds of neural activity.
Your brain is still highly active
doing unconscious states, during the sleep.
And even if your brain is highly synchronised,
one of the first theories of consciousness
was it depended on neurons firing in synchrony
with each other.
If your brain is too synchronised,
you will lose consciousness, and this happens
in states of absence epilepsy.
What seems to be the case is that, being conscious it all,
depends on how different brain regions talk
to each other in specific ways.
And this was some groundbreaking work by Marcello Massimini
in Milan about 10 years ago.
And what he did here, was he stimulated
the cortex of the brain with a brief pulse
of electromagnetic energy, using a technique called
transcranial magnetic stimulation or TMS.
And then he used EEG electroencephalography
to listen to the brain's echos.
A little bit like banging on the brain
and listening to its electrical response.
And what he noticed when you do this,
and you can see on the left is asleep,
and on the right is awake.
And this is very much slowed down.
When you stimulate the brain in a sleep condition,
there is still a response.
There's still an echo, but the echo stays very localised
to the point of stimulation.
It doesn't travel around very much,
and it doesn't last very long.
But when you stimulate a conscious brain,
there's a spatial temporally complex response.
This echo bounces around all over the cortex
in very interesting ways.
What's more, the complexity of this echo can be quantified.
You can apply some simple algorithms
to describe how complex, how rich,
this pattern of interactivity is.
This is also from the Milan group.
And what they've done here is, they basically look at the echo
as it moves around the brain.
And they see the extent to which you could describe it,
the minimum description length.
How much can you compress the image of that echo?
Much the same way that algorithms make
compressed files from digital images in your phone.
And they came up with an index called
the perturbational complexity index.
And what you find is, you now have
a number that you can attach to how conscious you are.
This is, I think, really intriguing,
because it's a first step towards having
an actual measurement of conscious level.
This graph on the bottom shows this measure
applied to a variety of conscious states,
ranging from pathological conscious states,
like the vegetative state, where there
is no consciousness at all, all the way through locked
in syndrome, and then healthy waking.
And you can immediately see that techniques
like this might already have clinical value in diagnosing
potential for consciousness patients
might have after severe brain injury.
Now at Sussex, we are continuing work along these lines.
We actually look, instead of bashing
on the brain with a sharp pulse of energy,
we want to see whether we can get something similar just
by recording the spontaneous activity of the brain.
So we look at spontaneous dynamics
from, in this case, waking states and anaesthesia.
This is work with my PhD students Michael Schartner
and Adam Barrett.
We measure its complexity, and indeed, we
find that we can distinguish different levels
of consciousness just by the spontaneous activity
of the brain.
In a way this isn't that surprising,
because we know various things change.
The balance of different frequencies
of your brain activity changes when you lose consciousness.
But this doesn't have to do with that.
This is independent of that.
There's something specific that is being detected
by these changes in complexity.
More recently we've applied the same measures to sleep,
in this case taking advantage with colleagues
in Milan of recordings taken from directly
within the human cerebral cortex.
These are implanted electrodes.
And we see much the same story.
If you compare where the two areas are,
you compare the complexity of wakeful rest,
and early non-rem sleep, where you are not dreaming very much.
You see that complexity fools a great deal.
What's interesting here is, if you compare wakeful rest to REM
sleep, where people will often report dreams if you wake them
up, the level of complexity is very much
as it is during the wakeful state.
There's something else going on here,
which is that the complexity in the frontal part of the brain
seems to be higher than in other parts of the brain.
And that's something we still don't understand fully, yet.
I just wanted to give you something
hot off the press, so to speak, which is where you've also
been applying these measures now to data taken
from people under the influence of psychedelic drugs;
psilocybin, ketamine, and LSD.
And what we find, at least in our hands to start with here,
is that the level of complexity actually
increases as compared to the baseline state, which is not
something we've seen before in any other application
of these measures.
So what's important about this way of looking at things
is that, it's grounded in a theory that
tries to explain why certain sorts of brain dynamics
go along with being conscious.
And put very simply, the idea is this--
and it goes back to Guilio Tononi and Gerald Edelman,
people that I went to work with in America about 18 years ago--
the idea is very simple.
Consciousness is extremely informative.
Every conscious experience you have, or have had,
or will have, is different from every other conscious
experience you have had, are having, or will have.
Even the experience of pure darkness
rules out a vast repertoire of alternative possible
experiences that you could have, or might have in the future.
There's a huge amount of information for the organism
in any conscious experience.
At the same time every experience that you have
is highly integrated.
Every conscious scene is experienced, as all of a piece,
is bound together.
We don't experience colours and shapes separately in any way.
It's conscious experiences at the level of phenomenology
combine these properties.
They are the one hand highly informative, composed
of many different parts.
On the other, bound together in a unified whole.
And this motivates us to search for mathematical measures which
have the same property, which are neither
lacking in information.
On the left, you see a system which
is all connected together, so it can't enter
very many different states.
On the right is a system which is completely dissociated,
so they can enter states, but it's not a single system.
We want measures that track this middle ground of systems,
that combine both integration and differentiation.
And a number of these measures now exist.
There are some equations here, which
we can talk about later if you like,
that try to target this middle ground.
And time will tell whether, by applying these more
phenomenologically grounded measures,
we come up with even more precise practical measures
of consciousness.
Now why is this business of measurement important?
And I want to make a general point here,
which is that, if you're trying to naturalise
a phenomenon which seems mysterious,
the ability to measure it is usually one of the most
important steps you can take.
And we've seen numerous examples of this.
The history of our understanding of heat and temperature
is one very good example Here's an early thermometer
from the 18th century, which used the expansion of air.
But of course there are many problems
in generating a reliable thermometer
and a scale of temperature, if you don't already
have fixed points.
And if you don't know what heat is you
get trapped in a kind of vicious circle
that took a long time to break out of.
But people did break out of this,
and the development of the thermometers
catalysed our understanding of heat
from being something that flowed in and out of objects, to being
something that was identical to a physical property.
The mean molecular kinetic energy of molecules
in a substance.
And having that concept of heat now
allows us to talk about temperature
far beyond the realms of human experience.
We can talk about the temperature
on the surface of the sun, in a sense,
the way we can talk about the temperature of interstellar
space, close to absolute zero.
None of these things make any sense
and in our phenomenological experience of hot and cold.
So this brings me to my first take-home message.
Measurement is important, and consciousness, conscious level,
depends on a complex balance of differentiation and integration
in brain dynamics, reflecting the fact
that conscious experiences themselves
are both highly informative and always integrated.
Now when we are conscious, we are conscious of something.
So what are the brain mechanisms that determine
the content of consciousness?
And the hero for this part of the story
is the German physicist and physiologist Hermann Von
Helmholtz.
And he proposed the idea that the brain is
a kind of prediction machine.
That what we see, hear, and feel are
nothing other than the brain's best guess about the causes
of sensory inputs.
And the basic idea is, again, quite simple.
The brain is locked inside its bony skull home,
and has very indirect access to the external world.
All it receives are ambiguous and noisy sensory signals,
which are highly and directly related
to this external world of objects, and so on,
if there is an external world of objects out there at all.
They know about that.
Perception in this view is, by necessity,
a process of inference in which the brain interprets
these ambiguous and noisy sensory signals with respect
to some prior expectations or beliefs
about the way the world is.
And this forms the brain's best guess
of the causes of the sensory signals that
are impacting our sensory surfaces all the time.
What we see is the brain's best guess of what's out there.
I want to give you a couple of examples
that illustrate this process.
It's quite easy to do, in a way.
This first example is a well-known visual illusion
called Edelstein's Checkerboard.
Now here, you'll see two patches.
You'll see patches A and B. And I hope to you they look to be
different shades of grey.
Do they?
They look to be different shades of grey.
Of course they are exactly the same shade of grey.
I can illustrate that by putting an alternative image,
and joining up those two patches.
You'll see that's it's the same shade of grey.
You may not believe me, so what I'll do
is, I'll shift it along, and you'll see even more clearly.
There are no sharp edges.
It's the same shade of grey.
What's going on here, of course, is
that the brain is unconsciously applying its prior knowledge
that a cast shadow dims the appearance of surfaces
that it casts onto.
So we therefore see the patch B as being
lighter than it really is, in order
to account for that effect.
And this is of course an illustration
of the success of the visual system, not its failure.
The visual system is a very bad physical light metre,
but that's not what it's supposed to do.
It's supposed to, or one thing it's supposed to do,
is to interpret the causes of sensory signals
in terms of meaningful objects in the world.
It's also an example of what we sometimes
call cognitive impenetrability.
Even if you know the patches are the same shade of grey,
when I take that bar away, they again look different.
Can't do much about that.
The second example just shows you
how quickly the brain can take in new prior information
to change the nature of conscious perception.
This is a so-called Mooney image.
And if you haven't seen it before, hopefully what
you will see here is just a passing
of black and white splotches.
Does everybody kind of get that?
Black and white splotches?
Some of you might have seen this before.
And now what I'm going to do is fill it in,
and you see something very different.
What you'll see is a very meaningful scene here involving
at least two objects, a woman, a hat and a horse.
Now if you stare at this for a while,
I won't leave it up for too long--
but if you just look at it for a little bit,
and then I take that image away again,
you should still be able to see the objects within that image.
Now for me this is quite remarkable,
because the sensory information hasn't changed at all.
All that's changed are your brain's prior expectations
about what that sensory data signifies.
And this changes what you consciously see.
Now this also works in the auditory domain.
Here are two spectrograms.
This is something called sine wave speech,
and what you see here are two time frequency representations
of speech sounds.
The one on the top has all the sharp acoustical features that
provide normal speech removed.
A little bit like thresholding an image.
And the bottom is something else.
So I'm going to play the top first,
and let's see what it sounds like.
[STRANGE BEEPS AND NOISES]
And now I'll play you something else.
(BOTTOM SOUND - A MAN'S VOICE): Jazz and swing fans
like fast music.
- So I hope you all understood that piece of sage advice.
And now if I play the original sound again--
[BEEPS AND WHISTLES THAT SOUND LIKE THE SENTENCE]
- Yeah?
This is exactly the same.
Again, all this change is what we
expect that sound to signify.
[SAME SOUND PLAYED AGAIN]
- One more time, just for luck.
It's not just a bunch of noisy whistles, it's speech.
Now this typical framework for thinking
about these kinds of effects is Bayesian Inference.
And this is a form of probabilistic reasoning, which
is applicable in all sorts of domains,
not just in neuroscience, in medical diagnosis,
and all sorts of things, like finding lost submarines.
But in neuroscience, we talk about the Bayesian brain.
And it's a way of formalising Helmholtz's idea
that perception is a form of best guessing.
And the idea is that sensory signals and prior beliefs
can be represented as probability distribution.
So for instance, this yellow curve
is the probability of something being the case,
maybe that you've got a brief glimpse of an object moving
to the right.
The sensory data may say something different.
It may have a probability that peaks
at a different angle of movement.
Maybe it's drifting in a different direction.
And the optimal combination of the prior,
and the likelihood, the yellow curve and the red curve,
is this green curve, which we will call the posterior
distribution.
And that represents the best optimal combination
of these two sorts of evidence.
And the idea is, well that's what we perceive.
Thinking about perception in this way
does something rather strange to the way,
classically in neuroscience, people
have thought about perception.
The classical view is that the brain
processes sensory information in a bottom-up,
or feed-forward direction.
This is a picture of the visual system of the monkey,
and the idea is that information comes in through the retina,
then goes through the thalamus.
It then goes to the back of the brain.
And as the sensory signals percolate deeper and deeper
and deeper into the brain, they encode
or represent progressively more sophisticated features
of objects.
So you start out at early levels the visual cortex
with response to luminance, and edges,
and then higher up to objects, including other monkeys.
What's important here is the perceptual heavy lifting
is done by information flowing in this bottom-up or
feed-forward direction.
Now the Bayesian brain idea says something very different.
It says that what's really important are the top-down
or inside-out connections that flow from the centre
of the brain back out.
And we've known for a long time there's
a large number, a very large number of these connections,
and some descriptions more than flow the other way around.
But the function has been rather mysterious.
Thinking about the Bayesian brain
gives us a nice way to interpret this.
Which is that it's exactly these top-down or inside-out
connections that convey predictions
from high levels of the brain to lower levels, to lower levels,
back out to the sensory surfaces.
So these blue arrows convey the brain's predictions
about the causes of sensory signals.
And then what flows in the feed-forward or bottom-up
direction, from the outside-in, that's just the prediction
area, the difference between what the brain expects
and what it gets at each level of description.
So this is often called predictive coding,
or predictive processing, in some formal frameworks.
And the idea is that minimization of prediction
error occurs across all levels of this hierarchy
at the same time, and what we then
perceive is the consequence of this joint minimization
of prediction error.
So you can think of perception as a sort
of controlled hallucination, in which
our perceptual predictions are being reined in at all points
by sensory information from the world and the body.
Now there's quite a lot of experiments
that show that something like this
is probably going on in the brain.
These are a couple of examples.
And since they're--
I was looking for the best example
so they don't come from my lab at all.
[LAUGHTER]
This is from Lars Muckli in Glasgow.
He's shown using advanced brain reading techniques, which
I won't describe, that you can decode the context of what
a person is seeing from parts of the visual cortex that
isn't even receiving any input.
And what's more, you can decode better
when you decode from the top part of the cortex, which
is supposed to receive predictions from higher levels.
So that suggests there are predictions being fed back.
And another study by Andre Bastos and Pascal Fries
in Germany, they used a method called Granger causality, which
is sensitive to information flow in systems.
And they find that top-down signals and bottom-up signals
are conveyed in different frequency bands
in the cortex, which is what you'd
predict from predictive coding.
One last experiment which I find particularly interesting
is an experiment from a Japanese group of Masanori Murayama.
And they used to optigenetics, which
is a way of using lights to selectively turn
on or off neural circuits.
And in this experiment they showed
that by just deactivating top superficial levels
of somatosensory cortex in a mouse
brain, the part of the mouse brain
that's sensitive to touch, they could
affect how well that mouse was able to do
tactile discriminations.
Those top-down connections were coming from a motor cortex.
So there's a lot of evidence that top-down connections
in the brain are important for perception,
is the basic message there.
But what's rather strange, and what I'm going to tell
you next is that all this stuff is all very good,
but predictive processing is not a theory of consciousness.
Nothing I've said has anything to do with consciousness,
at all.
It has to do-- it's a very general theory of how
brains do what they do.
How they do perception, how they do
cognition, how they do action.
So somewhat counter-intuitively, I
think this is exactly why it's a great theory of consciousness.
And the reason I think this is because it
allows us to ask all sorts of questions
about the real problem.
About what it is, what happens in brains
that underlies what you happen to be conscious of right now,
without getting sucked into the metaphysical pluckhole of why
you are conscious in the first place.
In other words, it provides a powerful approach
to looking for neural correlates of consciousness,
those things in the brain that go along with being conscious.
Because we can now take advantage
of a very general theory of how brains do what
they do, rather than just looking
at this region or that region.
So what does predictive processing,
or the Bayesian brain say about consciousness, specifically?
Well many years ago, some influential experiments
revealed a very strong connection
between top-down signalling and conscious contents.
In this example by Alvaro Pascual-Leone and Vincent
Walsh, what they did was they had
people look at visual motion, examples of visual motion.
And they used TMS, this interventional technique
where you can zap the brain very briefly.
I mentioned it before.
But they used it here, specifically
to interrupt the top-M down signalling
that was evoked by this perception of visual motion.
And the result was that, if you interrupted specifically
the top-down feedback, you would abolish
the conscious perception of visual motion,
even if you left the bottom-up signalling intact.
So that was an early key.
Now, more recently, in our lab and in many other labs
all over the place, we've been asking some other questions
about the relationship between what you expect
and what you consciously experience.
One of the most basic questions you can ask
is, do we consciously see what we expect to see?
Or do we see what violates our expectations of what we expect?
And a recent study from our group, led by Yair Pinto,
used a method called continuous flash suppression
to address this question.
It's illustrated here.
You see different images in the different eyes.
In one eye you see this rapidly changing Mondrian pattern
of squares.
And in the other eye, you either see a face or a house.
And they change contrast like this.
So initially, the person would just see this random pattern,
and then they'll see either a house or a face.
And simply, you just ask them to expect
to see-- you just tell them a face is more likely
or a house is more likely.
And what we find over a number of studies
is that we see faces more quickly when that's
what we're expecting to see.
It may seem obvious, but it could be the other way around.
At least in these studies, we see what we expect to see,
not what violates our expectations.
That's the data.
And the same goes for houses.
These kinds of studies support the idea
that it's the top-down predictions that
are really important for determining
what we're conscious of.
There's another experiment which I will just mention.
We did pretty much the same thing.
This is called motion induced blindness.
If you are in a lab rather than in a lecture theatre,
and you stare at this central point here,
then this red dot might disappear from time to time.
And what we did was after it disappeared,
we changed its colour and we led people to expect the colour
change to be one thing or another.
And again, it reappeared more quickly if it changed colour
in the way you were expecting.
Again, I am confirming that once your expectations were
validated, then that accelerated your conscious awareness
of something in the world.
Now that's just behavioural evidence.
That's just asking people what they see and when they see it.
We've also been interested in the brain mechanisms that
underlie and shape how our expectations change,
what we consciously see.
And we've been particularly interested in something
called the alpha rhythm.
Now the alpha rhythm is an oscillation
of about 10 hertz or 10 cycles per second.
That's especially prominent in the visual cortex,
across the back of the brain.
In one study, led by in this case a PhD student,
Maxine Sherman, with Ryota Kanai, in Sussex.
What we did here, we manipulated people's expectations
of what they we're likely to see.
And it's a very boring experiment, this was.
The only thing they could see was what we call Gabor patches.
They're just very dim patches of lines that
are blurry around the edges.
But the visual system loves these kinds of things.
They activate early visual cortex very, very well.
And people were expecting either that a patch
was there, or that it wasn't there, in different conditions.
And while doing this we measured brain activity.
And to cut a long story short, what we found
was that there were certain phases, certain parts
of the cycle, this 10 Hertz cycle,
at which your expectation had a greater effect on what
you said that you saw.
So there was part of this cycle, as the alpha rhythm,
there was part of it where your expectations dominated
your perception.
And there was another part of it which
was the opposite, where your sensory signals were
more important in determining what you saw at that point.
So this suggests that this oscillation
in the back of the brain is orchestrating
this exchange of predictions and prediction errors.
And that's the sort of cycle that
might be the neural mechanism for conscious vision.
And other theories about what the alpha rhythm is doing,
there are many.
One is that it's doing nothing, it's just the brain
idling away, and I think this is at least a more interesting way
to think about it.
Another experiment we've done with another PhD student
Asa-Chang, we showed people these very fast
changing luminance sequences.
And it turns out that your brains
will learn to predict the specific changes
in these sequences that change you very quickly.
And the signature of this learning,
again, happens to be in the alpha rhythm,
and suggests that this oscillation has something
important to do with how the brain learns and updates
predictions about sensory signals.
But we do not go around the world looking at Gabor patches
or rapidly changing things like this.
We go around the world looking at people and objects.
And that's what our visual world is composed of.
So can any of these ideas say anything
about our everyday visual experience?
And I think that's a very important challenge
in neuroscience to cross.
Get out of the lab and think about real world experiences.
So we've been using virtual reality over the last few years
to try to get at some of these ideas.
This is an Oculus Rift, which is now available to buy, I think.
And we've been using these to address
some of these real world aspects of visual perception.
And one of these real world aspects
is called perceptual presence.
And this is the observation that,
in normal perception, objects really seem to be there,
as opposed to being images of objects.
And this is, of course, what Magritte
plays with in his famous painting,
The Treachery of Images.
For instance, this is an object.
I think it's there, and in some sense
I can perceive the back of it, even though I
can't see the back of it, even though the back of it's
not giving me any sensory data, I perceive it
as an object with a back.
How does one explain that?
Well one idea you can come up with within this Bayesian brain
framework is that, the brain is not only
predicting the possible causes of the sensory signals getting
right here, right now.
But it's also predicting how sensory signals
would change were I to make particular actions.
Were I to pick this object up and move it around,
or just move my eyes from one place to another.
There's a long paper.
I wrote about that which I--
please don't read it.
[LAUGHTER]
- But that's the basic idea.
And how do you test an idea like that?
So we've been using some innovative virtual reality
methods, or augmented reality methods,
with my post-doc, Keisuke Suzuki.
And what we do is, we have virtual objects,
and these virtual objects, they either behave
as a normal object would.
They're all weird, unfamiliar objects.
But they can either behave as a normal object would behave,
so you can learn to predict what would happen.
This one is weird.
It always shows you the same face,
a little bit like having the moon on a plate
in front of you.
And then there are other conditions
where objects respond to your movements,
but they do so in unreliable and strange ways.
So the question is, what does the brain
learn about these objects, and how do we experience them?
Do we experience them as objects in different ways
when they behave differently?
And we're still doing those experiments.
Another way we can use VR, is to investigate
what happens in visual hallucinations of the kind
experienced in psychosis, and in certain other more
pharmacologically induced conditions.
What we're doing here is, we're coupling
immersive virtual reality, imagine
you've got a headset strapped to your head
again, with clever image processing that models
the effects of overactive priors on visual perception
to generate a highly immersive experience.
This is Sussex campus, actually, but now it
seems quite different than it did at lunchtime today.
I'll tell you that.
What we've done, we've recorded this panoramic video
which we can feedback through VR headset.
And we've processed this video through one
of these Google deep dream algorithms you might have seen,
that can generate sort of bowls of pasta
that looks like animals.
And this might seem like a lot of fun.
It is fun, but there is a serious purpose here,
because it allows us to model the effects, model very
unusual forms of perception and how
they might play out in different ways in the visual hierarchy.
And understanding how visual hallucinations might happen,
and how the wider effects they have on the mind, I think,
is a very important part of studying visual perception.
So that brings us to the second take-home message,
which is that what we consciously see
is the brain's best guess of the causes of its sensory input.
Normal perception is a fantasy that is constrained by reality.
Now before I move on to the last section,
I want to pay tribute to an unlikely character in a talk
about neuroscience, which is Ernst Gombrich.
Ernst Gombrich was one of the foremost historians
of art of the 20th century.
And it turns out that Gombrich's approach to understanding art
had a lot in common with ideas and the Bayesian brain.
And more specifically, with the idea
that perception is largely an act
of imagination, or construction, on the part of the perceiver.
And this is most apparent in his concept of the beholder's
share, which emphasises that the viewer brings
an awful lot to the table in the act of experiencing an artwork.
So he had this to say in his 1960 book, Art and Illusion,
"the artist gives the beholder 'more to do', he draws them
into the magic circle of creation and allows him
to experience something of the thrill of making which had once
been the privilege of the artist."
I think for me this is very powerful
when looking at, especially, things like Impressionist art.
And here, one way to think about this
is, that the artist has reverse engineered
the whole perceptual process, so that what's
there are not the objects, the end points of perception,
but rather the raw materials; the patterns of light that
engage our perceptual machinery in doing its work.
And for me this might be why paintings like this
are particularly powerful.
Now the final dimension of consciousness
I want to talk about is conscious self.
The fundamental experience of being someone.
Being someone like you.
There are many aspects to our experience
of being a conscious self.
There is the bodily self, the experience
of being and identifying with a particular body.
A bit of the world goes around with you in the world
all the time.
There's the perspectival self, the experience
of seeing the world, or experiencing the world,
from a particular first person perspective,
usually somewhere in the body, but not always.
There is the volitional self, the experience
of intending to do things, and of making things
happen in the world of agency.
And these ideas are, of course, often associated
with concepts of will.
Then there's the narrative self.
This is where-- only until now, we
don't have to worry about the concept of I,
but when we get to the narrative self, there is now and I.
There is a continuity of self experience from hour to hour,
from day to day, from month to month,
and from year to year, that you associate a name with,
and a particular set of autobiographical memories.
And finally, there's a social self.
The way I experience being me is partly
dependent on the way I perceive you as perceiving me.
I'm just going to talk, in the minutes remain,
about the bodily self.
This is something we're working on quite a lot in Sussex.
The experience of identifying with, and owning,
a particular body.
And the basic idea I want to convey, is again, very simple.
It's just that we should think of our experience of body
ownership in the same way that we
think about our experience of other things, as well.
That is, it's the brain's best guess of the causes
of body-related signals.
And the brain is always making this inference.
It's making its inference about what in the world
is part of the body, and what is not part of the body.
But it has access, in this case, to other sorts
of sensory signals, not just visual signals,
or tactile signals, but also proprioceptive signals.
The orange arrows here.
These inform the brain about the body's configuration
and position in space.
And then also, and often overlooked,
are interoceptive signals.
These are signals that originate from within the body, that
tell the brain about the physiological state
or condition of the inside of the internal physiological
milieu.
And you can think the idea is that our experience
of embodied self-hood is the brain's best
guess of the causes of all the signals put together.
Yeah, that's just to emphasise interoception.
An important part of this idea is
that interoception, the sense of the body from within,
should work along the same principles, the same Bayesian
principles that we've been thinking about,
vision and audition, previously.
That is, our experience of the inside of our bodies
is the brain's best guess of the causes of the signals that come
from the inside of our bodies.
So we can think of, again, top-down predictions
carrying predictions about what the bodily state is like,
and bottom-M up prediction errors that
report the differences between what's going on
and what the brain expects.
So what is our experience of the inside of our bodies?
Well, way back at the beginning of psychology,
William James and Carl Langer proposed
that emotions, emotional experience,
was really about the brain's perception of changes
in its physiological state, rather than perception
of the outside world.
So, in this classic example, seeing a bear
does not in itself generate the experience of fear.
Rather seeing the bear sets in train,
a load of physiological changes preparing for fight and flight
responses.
And it's the perception of those bodily changes
in the context with the bear being around that leads
to our experience of fear.
So the Bayesian perspective just generalises that idea,
and says that emotional experience
is the brain's best guess of the causes
of interoceptive signals.
And this fits very nicely with a lot of evidence.
And this is just one study done by a Finnish group.
And all they did here was, they had people report
where on their bodies they felt various emotions to take place.
And so you feel anxiety in one part of your body.
You feel fear in another, and so on.
So our experience of emotion does
seem to be intrinsically embodied.
Now another part of our experience of being a body
is the body as a physical object in the world.
And this might seem quite easy to take for granted,
since our physical body is just always there.
It goes around with us, it changes over the years,
in unfortunate ways.
But it's always there.
But it would be a mistake to take our experience of body
ownership for granted.
And there are some classic experiments
that demonstrate how malleable our experience of body
ownership is.
This is the famous rubber hand experiment.
Probably some of you have seen this.
What happens here is that a volunteer has their hand hidden
under a table, and the fake hand is put on top of the table,
and then both hands are simultaneously
stroked with a paintbrush.
And it turns out that just seeing a hand-like object where
a hand might be, feeling touch, and then seeing
the object being touched, is enough evidence
that the brain's best guess becomes that fake hand is,
in fact, part of my body.
Sort of part of my body.
This is what it looks like in practise.
Here you can see the fake hands, focusing on it.
There's the real hands, not focusing
on it, simultaneous stroking-- and there are
various ways you can test it.
[AUDIENCE LAUGHTER]
- I found in doing this, it works even better on children,
by the way, if you do that.
So that's interesting, because that's
using visual and tactile signals to convince
the brain that this object is part of its body.
In my lab, we've been interested in whether these
signals that come from inside the body also play a role.
So we set up a virtual reality version of this rubber hand
illusion, where people wear these goggles,
and they see a virtual fake hand.
And we also record their heartbeats.
And now what we can do is, we can make the virtual hand
flash either in time or out of time with their heart beat.
And we asked the question, do people
experience this virtual hand as more
belonging to them when it's flashing
in time, rather than out of time, with their heartbeat?
And the answer is that it does.
And this is just some data, basically that,
bigger than that, which means that, indeed, they experience
the hand as more their own.
The way we measure that actually,
is that first we can ask them.
That's the easiest way.
Then we can also ask them to point
to where they think their hand really is,
and we can see how far they drift
from where their hand really is to where the virtual hand is.
And that provides a more objective way
of measuring the strength of the effect.
Here's what it looks like in practise.
Again if you can see this, that's the real hand.
That's somebody's virtual hand.
Again, imagine you're wearing a headset
so you'll see this in 3-D. And you
can just about see it pulsing to read and back, I hope.
And you can also do some other things
with these virtual reality rubber hands
that you couldn't do with real rubber hands.
For instance, you can map movements of the real hand
to the virtual hand, so you can start to ask questions
about the extent to which the virtual hand moves
as I predict it to move.
How much does that affect the extent to which I
feel it to be part of my body?
You can make it change colour.
So you can have somebody embody a skin colour associated with
a cultural out-group, and see if they become less racist
as a result.
And then my favourite is where you can change, actually,
the size of the body.
And that's coming up here.
So here what we do is, we can have the hand telescope
up and down in size.
And again, this might seem like fun, and it is fun,
but there is a serious purpose.
There are various conditions.
There's in fact, a condition called
Alice in Wonderland Syndrome, where people report
that all parts of their body are, indeed, telescoping
up and down in size.
And in a more subtle way, there are lots of body dysmorphias,
of subtle misperceptions of body shape, which
might be associated with eating disorders.
And so these sorts of techniques allow us to approach,
in a very fine grained way, how people might
mis-perceive their own bodies.
That brings me to the third take-home message about self.
And with apologies to Descartes, the take home message
is that, I predict myself, therefore I am.
In the last nine minutes, before the bell rings,
I want to go full circle and return to this Cartesian idea
of the beast machine.
To try to convince you that our experience of being
a conscious self is intimately tied up with our beast machine
nature.
And to do this, I need to mention one final aspects
of perceptual inference, which has a lot
to do with Karl Friston, who's done a lot of work
in the Bayesian brain UCL here in London.
And if we think of the brain as being in the business
of minimising prediction errors, this can be done either
by updating our perceptual predictions,
which is what I've been talking about so far.
And this what Helmholtz said.
Or we can minimise prediction errors by making actions.
We can change what we predict, or we
can make an action so that our predictions come true.
You can change with sensory input,
or you can change what you believe
about your sensory input.
One point of doing this is, that you can make actions,
then, to find out more about the world that surrounds you.
And this is what Helmholtz has in mind
when he says that each movement we make,
by which we alter the appearance of objects,
should be thought of as an experiment designed
to test whether we've understood correctly
the invariant relations of the phenomena before us.
Which Gregory, much later, said something similar
when he talked about perception as hypothesis testing.
The point of this is, that we make
eye movements, and other kinds of movements,
to understand what the world is like.
That, in fact, there was is tomato there, for instance.
But there's another way to think about active inference,
which is that, when we minimise prediction error,
what we're actually doing is controlling a sensory variable.
We're preventing it from changing,
because we're making our prediction about what
it is come true.
And this is the use of active inference,
to control or regulate something,
rather than to understand what the causes of that something
are.
And this brings a very different tradition
to mind, which is 20th century cybernetics.
And this is Ross Ashby, who was a pioneer
of this way of thinking.
And he, with Roger Conan right at the end of his life,
wrote a paper.
The title of the paper, was "Every good regulator
of a system must be a model of that system".
The idea here is, if you want to regulate something very
precisely, then you need a good model
of what effects that system.
Now you could apply this idea to the external world, as well.
When you try to catch a cricket ball,
you are actually trying to control the level
of the angle above the horizon.
But it applies more naturally, I think,
to the internal state of our body.
So really, what matters about my internal physiological
condition, I don't really need to know
exactly what it's like inside my body, and care about it.
But I need to control it, and my brain needs to regulate it.
So this way of thinking about active inference
applies more naturally to interoception.
Think about it this way that.
Having good predictive models are always useful,
but we can have a pendulum that swings, on the one hand,
towards control.
We can use these predictive models for control,
and that's more applied to the state of our internal body.
Or we can swing the other way, and think about perception,
understanding.
You could think of the instrumental, epistemic ways
of thinking about the role of action and perception.
And this brings to mind--
I mentioned Karl Friston.
He's come up with this thing called the free energy
principle.
And I can only nod to the vast body
of work he's done here on this.
With the slogan, which is that organisms, over the long run,
maintain themselves in states in which they expect to be in,
in virtue of having good predictive models
about their own internal condition.
So this takes us right back to Descartes,
but in a very different way.
As I said right at the beginning of this lecture, for Descartes
our physiological reality was rather
irrelevant to our minds, our rationality, our consciousness.
This is a quote from a 1968 paper on his beast machine
argument, "without minds to direct their bodily movements,
animals must be regarded as unthinking, unfeeling machines
that move like clockwork."
Now I think if you try to think how
this idea of our predictive models
controlling our internal physiological states,
and the resulting experiences that perceptual content that
might give rise to, you can make the opposite case.
And the opposite case would be that conscious self-hood
emerges because of, and not in spite of, the fact
that we are beast machines.
And I think this is a deeply embodied view of consciousness
and self, and it speaks to this fundamental link in continuity
between what it is to be alive, what it is to have a mind,
and what it is to be a conscious self.
So I repeat, the third take-home message should
make even more sense now.
That I predict myself, therefore I am.
And I am a conscious self because I'm
a bag of self-fulfilling predictions
about my own physiological persistence over time.
Now why does any of this matter?
It's a lot of interesting ideas, but why should we
be interested in studying consciousness?
Well it's a very interesting thing,
I hope I've convinced you.
But there are lots of practical reasons
to be interested as well.
There are between 20 and 60,000 patients in the UK
alone, who are in disorders of consciousness.
You are in the vegetative state, or in coma,
or in some other severely abnormal state
of consciousness.
Having better measures of conscious level,
as I described at the beginning, is
going to really change again, and how
we treat people like this.
And of course, in psychiatry.
Psychiatric disorders are increasing that prevalence
across all modern societies, and it's
estimated one in six of us, at any one time,
are suffering from a psychiatric condition.
And understanding the mechanisms that underlie conscious
content and conscious self, because a lot
of psychiatric disorders include disturbances
of the way we experience our body,
even though that might not be the most obvious symptom,
can help us understand the mechanisms
involved in psychiatric disorders,
not just the symptoms.
There are also some more general reasons
for studying consciousness, which bring up
some ethical questions.
When does consciousness emerge in development
on newborn babies conscious?
Or does consciousness start even in the womb?
Maybe different dimensions of consciousness
emerge at different times.
Are other animals conscious?
Well I think it can make a very good case for mammals
and primates, but what about the octopus?
The octopus has more neurons in its arms
than in its central brain.
They're very smart creatures.
Here, you have to ask the question not only, what
is it like to be an octopus, but what is
it like to be an octopus arm?
And finally, with the rise of artificial intelligence,
we should begin to ask questions about what would it
take for a machine to have some kind of subjective experience.
I don't think we're anywhere near that yet,
but we should consider what science can tell us
about its possibility, because that would raise some very,
very tricky ethical questions.
But, fundamentally, consciousness
remains fascinating for me for the same reason
that it's motivated people throughout the ages.
I mean, Hippocrates, the founder of modern medicine,
put it one way.
He said, from the brain and from the brain alone arise
our sorrows, our joys.
And he also had his first view of psychiatry
that madness comes from its moistness.
And then Francis Crick, in the 1990s, who
I mentioned in the beginning.
He gave birth, if you like, to the modern neuroscience
of consciousness.
He said much the same thing in his astonishing hypothesis.
But there is this mystery and wonder still,
about how the biological machinery inside our heads
gives rise to the rich inner universe that
makes life worth living.
And despite this mystery, modern science is making progress.
I hope I've given you a flavour, even though we don't understand
how consciousness happens, we can begin to understand its
mechanisms.
So we should not be afraid of naturalising consciousness.
It's not a bad thing to understand its basis
in the material world.
As so often in science, with greater understanding
comes a larger sense of wonder, and a greater realisation
that we are part of, and not apart from, the rest of nature.
[AUDIENCE APPLAUSE]