back to index

Jeff Hawkins: Thousand Brains Theory of Intelligence | Lex Fridman Podcast #25


small model | large model

link |
00:00:00.000
The following is a conversation with Jeff Hawkins.
link |
00:00:02.360
He's the founder of the Redwood Center
link |
00:00:04.120
for Theoretical Neuroscience in 2002, and NuMenta in 2005.
link |
00:00:08.980
In his 2004 book, titled On Intelligence,
link |
00:00:11.920
and in the research before and after,
link |
00:00:13.840
he and his team have worked to reverse engineer
link |
00:00:16.200
the neural cortex, and propose artificial intelligence
link |
00:00:19.160
architectures, approaches, and ideas
link |
00:00:21.360
that are inspired by the human brain.
link |
00:00:23.640
These ideas include Hierarchical Tupperware Memory,
link |
00:00:25.960
HTM, from 2004, and new work,
link |
00:00:28.920
the Thousand Brains Theory of Intelligence
link |
00:00:30.720
from 2017, 18, and 19.
link |
00:00:33.760
Jeff's ideas have been an inspiration
link |
00:00:36.120
to many who have looked for progress
link |
00:00:38.200
beyond the current machine learning approaches,
link |
00:00:40.480
but they have also received criticism
link |
00:00:42.720
for lacking a body of empirical evidence
link |
00:00:44.680
supporting the models.
link |
00:00:46.240
This is always a challenge when seeking more
link |
00:00:48.440
than small incremental steps forward in AI.
link |
00:00:51.440
Jeff is a brilliant mind, and many of the ideas
link |
00:00:54.120
he has developed and aggregated from neuroscience
link |
00:00:56.500
are worth understanding and thinking about.
link |
00:00:59.120
There are limits to deep learning,
link |
00:01:00.920
as it is currently defined.
link |
00:01:02.920
Forward progress in AI is shrouded in mystery.
link |
00:01:05.760
My hope is that conversations like this
link |
00:01:07.760
can help provide an inspiring spark for new ideas.
link |
00:01:11.440
This is the Artificial Intelligence Podcast.
link |
00:01:14.020
If you enjoy it, subscribe on YouTube, iTunes,
link |
00:01:16.720
or simply connect with me on Twitter
link |
00:01:18.640
at Lex Friedman, spelled F R I D.
link |
00:01:21.520
And now, here's my conversation with Jeff Hawkins.
link |
00:01:26.780
Are you more interested in understanding the human brain
link |
00:01:29.860
or in creating artificial systems
link |
00:01:32.000
that have many of the same qualities
link |
00:01:34.640
but don't necessarily require that you actually understand
link |
00:01:38.560
the underpinning workings of our mind?
link |
00:01:41.480
So there's a clear answer to that question.
link |
00:01:44.000
My primary interest is understanding the human brain.
link |
00:01:46.760
No question about it.
link |
00:01:47.700
But I also firmly believe that we will not be able
link |
00:01:53.280
to create fully intelligent machines
link |
00:01:55.040
until we understand how the human brain works.
link |
00:01:57.280
So I don't see those as separate problems.
link |
00:02:00.120
I think there's limits to what can be done
link |
00:02:01.720
with machine intelligence if you don't understand
link |
00:02:03.540
the principles by which the brain works.
link |
00:02:05.680
And so I actually believe that studying the brain
link |
00:02:07.900
is actually the fastest way to get to machine intelligence.
link |
00:02:11.960
And within that, let me ask the impossible question,
link |
00:02:14.640
how do you, not define, but at least think
link |
00:02:17.160
about what it means to be intelligent?
link |
00:02:19.440
So I didn't try to answer that question first.
link |
00:02:22.280
We said, let's just talk about how the brain works
link |
00:02:24.520
and let's figure out how certain parts of the brain,
link |
00:02:26.720
mostly the neocortex, but some other parts too.
link |
00:02:29.920
The parts of the brain most associated with intelligence.
link |
00:02:32.340
And let's discover the principles by how they work.
link |
00:02:35.840
Because intelligence isn't just like some mechanism
link |
00:02:39.360
and it's not just some capabilities.
link |
00:02:40.680
It's like, okay, we don't even know
link |
00:02:42.520
where to begin on this stuff.
link |
00:02:44.040
And so now that we've made a lot of progress on this,
link |
00:02:49.320
after we've made a lot of progress
link |
00:02:50.480
on how the neocortex works, and we can talk about that,
link |
00:02:53.200
I now have a very good idea what's gonna be required
link |
00:02:55.840
to make intelligent machines.
link |
00:02:57.200
I can tell you today, some of the things
link |
00:02:59.600
are gonna be necessary, I believe,
link |
00:03:02.140
to create intelligent machines.
link |
00:03:03.480
Well, so we'll get there.
link |
00:03:04.600
We'll get to the neocortex and some of the theories
link |
00:03:07.440
of how the whole thing works.
link |
00:03:09.200
And you're saying, as we understand more and more
link |
00:03:11.760
about the neocortex, about our own human mind,
link |
00:03:14.760
we'll be able to start to more specifically define
link |
00:03:17.680
what it means to be intelligent.
link |
00:03:18.680
It's not useful to really talk about that until.
link |
00:03:21.840
I don't know if it's not useful.
link |
00:03:23.560
Look, there's a long history of AI, as you know.
link |
00:03:26.160
And there's been different approaches taken to it.
link |
00:03:28.900
And who knows, maybe they're all useful.
link |
00:03:32.240
So the good old fashioned AI, the expert systems,
link |
00:03:37.280
the current convolutional neural networks,
link |
00:03:38.920
they all have their utility.
link |
00:03:40.380
They all have a value in the world.
link |
00:03:43.780
But I would think almost everyone agree
link |
00:03:45.220
that none of them are really intelligent
link |
00:03:46.620
in a sort of a deep way that humans are.
link |
00:03:49.860
And so it's just the question of how do you get
link |
00:03:53.620
from where those systems were or are today
link |
00:03:56.420
to where a lot of people think we're gonna go.
link |
00:03:59.240
And there's a big, big gap there, a huge gap.
link |
00:04:02.340
And I think the quickest way of bridging that gap
link |
00:04:06.220
is to figure out how the brain does that.
link |
00:04:08.820
And then we can sit back and look and say,
link |
00:04:10.100
oh, which of these principles that the brain works on
link |
00:04:12.980
are necessary and which ones are not?
link |
00:04:15.140
Clearly, we don't have to build this in,
link |
00:04:16.620
and intelligent machines aren't gonna be built
link |
00:04:18.460
out of organic living cells.
link |
00:04:22.720
But there's a lot of stuff that goes on the brain
link |
00:04:24.700
that's gonna be necessary.
link |
00:04:25.900
So let me ask maybe, before we get into the fun details,
link |
00:04:30.260
let me ask maybe a depressing or a difficult question.
link |
00:04:33.060
Do you think it's possible that we will never
link |
00:04:36.220
be able to understand how our brain works,
link |
00:04:38.060
that maybe there's aspects to the human mind,
link |
00:04:41.820
like we ourselves cannot introspectively get to the core,
link |
00:04:46.140
that there's a wall you eventually hit?
link |
00:04:48.100
Yeah, I don't believe that's the case.
link |
00:04:52.020
I have never believed that's the case.
link |
00:04:53.240
There's not been a single thing humans have ever put
link |
00:04:56.620
their minds to that we've said, oh, we reached the wall,
link |
00:04:58.620
we can't go any further.
link |
00:04:59.700
It's just, people keep saying that.
link |
00:05:01.660
People used to believe that about life.
link |
00:05:03.380
Alain Vital, right, there's like,
link |
00:05:05.180
what's the difference between living matter
link |
00:05:06.380
and nonliving matter, something special
link |
00:05:07.980
that we never understand.
link |
00:05:09.100
We no longer think that.
link |
00:05:10.660
So there's no historical evidence that suggests this
link |
00:05:14.220
is the case, and I just never even consider
link |
00:05:16.340
that's a possibility.
link |
00:05:17.620
I would also say, today, we understand so much
link |
00:05:21.860
about the neocortex.
link |
00:05:22.820
We've made tremendous progress in the last few years
link |
00:05:25.480
that I no longer think of it as an open question.
link |
00:05:30.000
The answers are very clear to me.
link |
00:05:32.100
The pieces we don't know are clear to me,
link |
00:05:34.800
but the framework is all there, and it's like,
link |
00:05:36.740
oh, okay, we're gonna be able to do this.
link |
00:05:38.620
This is not a problem anymore, just takes time and effort,
link |
00:05:41.100
but there's no mystery, a big mystery anymore.
link |
00:05:44.060
So then let's get into it for people like myself
link |
00:05:47.800
who are not very well versed in the human brain,
link |
00:05:52.940
except my own.
link |
00:05:54.780
Can you describe to me, at the highest level,
link |
00:05:57.300
what are the different parts of the human brain,
link |
00:05:59.140
and then zooming in on the neocortex,
link |
00:06:02.060
the parts of the neocortex, and so on,
link |
00:06:04.120
a quick overview.
link |
00:06:05.500
Yeah, sure.
link |
00:06:06.620
The human brain, we can divide it roughly into two parts.
link |
00:06:10.780
There's the old parts, lots of pieces,
link |
00:06:14.220
and then there's the new part.
link |
00:06:15.700
The new part is the neocortex.
link |
00:06:18.020
It's new because it didn't exist before mammals.
link |
00:06:20.420
The only mammals have a neocortex,
link |
00:06:22.180
and in humans, in primates, it's very large.
link |
00:06:24.780
In the human brain, the neocortex occupies
link |
00:06:26.900
about 70 to 75% of the volume of the brain.
link |
00:06:30.660
It's huge.
link |
00:06:32.100
And the old parts of the brain are,
link |
00:06:34.860
there's lots of pieces there.
link |
00:06:36.020
There's the spinal cord, and there's the brain stem,
link |
00:06:38.740
and the cerebellum, and the different parts
link |
00:06:40.240
of the basal ganglia, and so on.
link |
00:06:42.020
In the old parts of the brain,
link |
00:06:42.960
you have the autonomic regulation,
link |
00:06:44.800
like breathing and heart rate.
link |
00:06:46.280
You have basic behaviors, so like walking and running
link |
00:06:49.460
are controlled by the old parts of the brain.
link |
00:06:51.380
All the emotional centers of the brain
link |
00:06:53.060
are in the old part of the brain,
link |
00:06:53.940
so when you feel anger or hungry, lust,
link |
00:06:55.380
or things like that, those are all
link |
00:06:56.500
in the old parts of the brain.
link |
00:06:57.940
And we associate with the neocortex
link |
00:07:02.180
all the things we think about as sort of
link |
00:07:04.060
high level perception and cognitive functions,
link |
00:07:08.100
anything from seeing and hearing and touching things
link |
00:07:12.240
to language to mathematics and engineering
link |
00:07:15.140
and science and so on.
link |
00:07:16.940
Those are all associated with the neocortex,
link |
00:07:19.760
and they're certainly correlated.
link |
00:07:21.760
Our abilities in those regards are correlated
link |
00:07:23.980
with the relative size of our neocortex
link |
00:07:25.820
compared to other mammals.
link |
00:07:27.940
So that's like the rough division,
link |
00:07:30.520
and you obviously can't understand
link |
00:07:32.740
the neocortex completely isolated,
link |
00:07:35.160
but you can understand a lot of it
link |
00:07:37.020
with just a few interfaces to the old parts of the brain,
link |
00:07:40.340
and so it gives you a system to study.
link |
00:07:44.980
The other remarkable thing about the neocortex,
link |
00:07:48.020
compared to the old parts of the brain,
link |
00:07:49.880
is the neocortex is extremely uniform.
link |
00:07:52.900
It's not visibly or anatomically,
link |
00:07:57.060
it's very, I always like to say
link |
00:07:59.460
it's like the size of a dinner napkin,
link |
00:08:01.300
about two and a half millimeters thick,
link |
00:08:03.740
and it looks remarkably the same everywhere.
link |
00:08:05.980
Everywhere you look in that two and a half millimeters
link |
00:08:07.900
is this detailed architecture,
link |
00:08:10.060
and it looks remarkably the same everywhere,
link |
00:08:11.580
and that's across species.
link |
00:08:12.620
A mouse versus a cat and a dog and a human.
link |
00:08:15.380
Where if you look at the old parts of the brain,
link |
00:08:17.060
there's lots of little pieces do specific things.
link |
00:08:19.620
So it's like the old parts of our brain evolved,
link |
00:08:22.060
like this is the part that controls heart rate,
link |
00:08:23.660
and this is the part that controls this,
link |
00:08:24.860
and this is this kind of thing,
link |
00:08:25.780
and that's this kind of thing,
link |
00:08:27.180
and these evolved for eons a long, long time,
link |
00:08:30.100
and they have their specific functions,
link |
00:08:31.580
and all of a sudden mammals come along,
link |
00:08:33.220
and they got this thing called the neocortex,
link |
00:08:35.180
and it got large by just replicating the same thing
link |
00:08:38.140
over and over and over again.
link |
00:08:39.420
This is like, wow, this is incredible.
link |
00:08:42.660
So all the evidence we have,
link |
00:08:46.260
and this is an idea that was first articulated
link |
00:08:50.020
in a very cogent and beautiful argument
link |
00:08:52.020
by a guy named Vernon Malcastle in 1978, I think it was,
link |
00:08:56.820
that the neocortex all works on the same principle.
link |
00:09:01.580
So language, hearing, touch, vision, engineering,
link |
00:09:05.260
all these things are basically underlying,
link |
00:09:06.980
are all built on the same computational substrate.
link |
00:09:10.340
They're really all the same problem.
link |
00:09:11.820
So the low level of the building blocks all look similar.
link |
00:09:14.860
Yeah, and they're not even that low level.
link |
00:09:16.300
We're not talking about like neurons.
link |
00:09:17.900
We're talking about this very complex circuit
link |
00:09:19.940
that exists throughout the neocortex.
link |
00:09:21.420
It's remarkably similar.
link |
00:09:23.500
It's like, yes, you see variations of it here and there,
link |
00:09:26.580
more of the cell, less and less, and so on.
link |
00:09:29.620
But what Malcastle argued was, he says,
link |
00:09:32.700
you know, if you take a section of neocortex,
link |
00:09:35.580
why is one a visual area and one is a auditory area?
link |
00:09:38.580
Or why is, and his answer was,
link |
00:09:41.180
it's because one is connected to eyes
link |
00:09:43.180
and one is connected to ears.
link |
00:09:45.380
Literally, you mean just it's most closest
link |
00:09:47.820
in terms of number of connections
link |
00:09:49.020
to the sensor. Literally, literally,
link |
00:09:50.900
if you took the optic nerve and attached it
link |
00:09:53.780
to a different part of the neocortex,
link |
00:09:55.300
that part would become a visual region.
link |
00:09:57.940
This actually, this experiment was actually done
link |
00:10:00.380
by Merkankasur in developing, I think it was lemurs,
link |
00:10:04.980
I can't remember what it was, some animal.
link |
00:10:06.700
And there's a lot of evidence to this.
link |
00:10:08.540
You know, if you take a blind person,
link |
00:10:09.940
a person who's born blind at birth,
link |
00:10:12.180
they're born with a visual neocortex.
link |
00:10:15.420
It doesn't, may not get any input from the eyes
link |
00:10:18.260
because of some congenital defect or something.
link |
00:10:21.260
And that region becomes, does something else.
link |
00:10:24.700
It picks up another task.
link |
00:10:27.020
So, and it's, so it's this very complex thing.
link |
00:10:32.300
It's not like, oh, they're all built on neurons.
link |
00:10:33.740
No, they're all built in this very complex circuit
link |
00:10:36.460
and somehow that circuit underlies everything.
link |
00:10:40.300
And so this is the, it's called
link |
00:10:43.580
the common cortical algorithm, if you will.
link |
00:10:45.900
Some scientists just find it hard to believe
link |
00:10:47.980
and they just, I can't believe that's true,
link |
00:10:50.060
but the evidence is overwhelming in this case.
link |
00:10:52.100
And so a large part of what it means
link |
00:10:54.340
to figure out how the brain creates intelligence
link |
00:10:56.420
and what is intelligence in the brain
link |
00:10:59.860
is to understand what that circuit does.
link |
00:11:02.020
If you can figure out what that circuit does,
link |
00:11:05.020
as amazing as it is, then you can,
link |
00:11:06.940
then you understand what all these
link |
00:11:08.620
other cognitive functions are.
link |
00:11:10.500
So if you were to sort of put neocortex
link |
00:11:13.300
outside of your book on intelligence,
link |
00:11:15.140
you look, if you wrote a giant tome, a textbook
link |
00:11:18.020
on the neocortex, and you look maybe
link |
00:11:21.980
a couple of centuries from now,
link |
00:11:23.740
how much of what we know now would still be accurate
link |
00:11:26.500
two centuries from now?
link |
00:11:27.660
So how close are we in terms of understanding?
link |
00:11:30.820
I have to speak from my own particular experience here.
link |
00:11:32.980
So I run a small research lab here.
link |
00:11:35.860
It's like any other research lab.
link |
00:11:38.020
I'm sort of the principal investigator.
link |
00:11:39.420
There's actually two of us
link |
00:11:40.260
and there's a bunch of other people.
link |
00:11:42.540
And this is what we do.
link |
00:11:43.820
We study the neocortex and we publish our results
link |
00:11:46.060
and so on.
link |
00:11:46.900
So about three years ago,
link |
00:11:49.820
we had a real breakthrough in this field.
link |
00:11:52.460
Just tremendous breakthrough.
link |
00:11:53.300
We've now published, I think, three papers on it.
link |
00:11:56.500
And so I have a pretty good understanding
link |
00:12:00.180
of all the pieces and what we're missing.
link |
00:12:02.300
I would say that almost all the empirical data
link |
00:12:06.260
we've collected about the brain, which is enormous.
link |
00:12:08.540
If you don't know the neuroscience literature,
link |
00:12:10.340
it's just incredibly big.
link |
00:12:13.980
And it's, for the most part, all correct.
link |
00:12:16.860
It's facts and experimental results and measurements
link |
00:12:21.660
and all kinds of stuff.
link |
00:12:22.980
But none of that has been really assimilated
link |
00:12:25.860
into a theoretical framework.
link |
00:12:27.900
It's data without, in the language of Thomas Kuhn,
link |
00:12:32.300
the historian, would be a sort of a pre paradigm science.
link |
00:12:35.340
Lots of data, but no way to fit it together.
link |
00:12:38.180
I think almost all of that's correct.
link |
00:12:39.540
There's just gonna be some mistakes in there.
link |
00:12:42.180
And for the most part,
link |
00:12:43.300
there aren't really good cogent theories about it,
link |
00:12:45.940
how to put it together.
link |
00:12:47.300
It's not like we have two or three competing good theories,
link |
00:12:50.060
which ones are right and which ones are wrong.
link |
00:12:51.540
It's like, nah, people are just scratching their heads.
link |
00:12:54.780
Some people have given up
link |
00:12:55.620
on trying to figure out what the whole thing does.
link |
00:12:57.620
In fact, there's very, very few labs that we do
link |
00:13:01.020
that focus really on theory
link |
00:13:03.340
and all this unassimilated data and trying to explain it.
link |
00:13:06.780
So it's not like we've got it wrong.
link |
00:13:08.940
It's just that we haven't got it at all.
link |
00:13:11.100
So it's really, I would say, pretty early days
link |
00:13:15.020
in terms of understanding the fundamental theory's forces
link |
00:13:19.060
of the way our mind works.
link |
00:13:20.220
I don't think so.
link |
00:13:21.500
I would have said that's true five years ago.
link |
00:13:25.340
So as I said,
link |
00:13:26.980
we had some really big breakthroughs on this recently
link |
00:13:29.300
and we started publishing papers on this.
link |
00:13:30.780
So we'll get to that.
link |
00:13:34.180
But so I don't think it's,
link |
00:13:35.940
I'm an optimist and from where I sit today,
link |
00:13:38.260
most people would disagree with this,
link |
00:13:39.420
but from where I sit today, from what I know,
link |
00:13:43.260
it's not super early days anymore.
link |
00:13:44.940
We are, the way these things go
link |
00:13:46.860
is it's not a linear path, right?
link |
00:13:48.180
You don't just start accumulating
link |
00:13:49.820
and get better and better and better.
link |
00:13:50.820
No, all this stuff you've collected,
link |
00:13:52.900
none of it makes sense.
link |
00:13:53.780
All these different things are just sort of around.
link |
00:13:55.580
And then you're gonna have some breaking points
link |
00:13:57.100
where all of a sudden, oh my God, now we got it right.
link |
00:13:59.420
That's how it goes in science.
link |
00:14:01.100
And I personally feel like we passed that little thing
link |
00:14:04.460
about a couple of years ago,
link |
00:14:06.300
all that big thing a couple of years ago.
link |
00:14:07.580
So we can talk about that.
link |
00:14:09.620
Time will tell if I'm right,
link |
00:14:11.020
but I feel very confident about it.
link |
00:14:12.660
That's why I'm willing to say it on tape like this.
link |
00:14:15.220
At least very optimistic.
link |
00:14:18.060
So let's, before those few years ago,
link |
00:14:20.220
let's take a step back to HTM,
link |
00:14:23.260
the hierarchical temporal memory theory,
link |
00:14:26.020
which you first proposed on intelligence
link |
00:14:27.580
and went through a few different generations.
link |
00:14:29.340
Can you describe what it is,
link |
00:14:31.300
how it evolved through the three generations
link |
00:14:33.740
since you first put it on paper?
link |
00:14:35.460
Yeah, so one of the things that neuroscientists
link |
00:14:39.340
just sort of missed for many, many years,
link |
00:14:42.980
and especially people who were thinking about theory,
link |
00:14:45.820
was the nature of time in the brain.
link |
00:14:49.100
Brains process information through time.
link |
00:14:51.700
The information coming into the brain
link |
00:14:52.900
is constantly changing.
link |
00:14:55.220
The patterns from my speech right now,
link |
00:14:57.620
if you were listening to it at normal speed,
link |
00:15:00.140
would be changing on your ears
link |
00:15:01.500
about every 10 milliseconds or so, you'd have a change.
link |
00:15:04.100
This constant flow, when you look at the world,
link |
00:15:06.740
your eyes are moving constantly,
link |
00:15:08.220
three to five times a second,
link |
00:15:09.700
and the input's completely changing.
link |
00:15:11.380
If I were to touch something like a coffee cup,
link |
00:15:13.500
as I move my fingers, the input changes.
link |
00:15:15.220
So this idea that the brain works on time changing patterns
link |
00:15:19.500
is almost completely, or was almost completely missing
link |
00:15:22.340
from a lot of the basic theories,
link |
00:15:23.620
like fears of vision and so on.
link |
00:15:25.020
It's like, oh no, we're gonna put this image
link |
00:15:26.860
in front of you and flash it and say, what is it?
link |
00:15:29.580
Convolutional neural networks work that way today, right?
link |
00:15:32.180
Classify this picture.
link |
00:15:34.220
But that's not what vision is like.
link |
00:15:35.980
Vision is this sort of crazy time based pattern
link |
00:15:38.740
that's going all over the place,
link |
00:15:40.060
and so is touch and so is hearing.
link |
00:15:41.820
So the first part of hierarchical temporal memory
link |
00:15:43.780
was the temporal part.
link |
00:15:45.060
It's to say, you won't understand the brain,
link |
00:15:48.260
nor will you understand intelligent machines
link |
00:15:50.020
unless you're dealing with time based patterns.
link |
00:15:52.460
The second thing was, the memory component of it was,
link |
00:15:55.460
is to say that we aren't just processing input,
link |
00:16:00.300
we learn a model of the world.
link |
00:16:02.820
And the memory stands for that model.
link |
00:16:05.500
The point of the brain, the part of the neocortex,
link |
00:16:07.340
it learns a model of the world.
link |
00:16:08.500
We have to store things, our experiences,
link |
00:16:11.580
in a form that leads to a model of the world.
link |
00:16:14.220
So we can move around the world,
link |
00:16:15.700
we can pick things up and do things and navigate
link |
00:16:17.380
and know how it's going on.
link |
00:16:18.220
So that's what the memory referred to.
link |
00:16:19.980
And many people just, they were thinking about
link |
00:16:22.100
like certain processes without memory at all.
link |
00:16:25.140
They're just like processing things.
link |
00:16:26.740
And then finally, the hierarchical component
link |
00:16:29.020
was a reflection to that the neocortex,
link |
00:16:32.260
although it's this uniform sheet of cells,
link |
00:16:35.820
different parts of it project to other parts,
link |
00:16:37.580
which project to other parts.
link |
00:16:39.340
And there is a sort of rough hierarchy in terms of that.
link |
00:16:43.060
So the hierarchical temporal memory is just saying,
link |
00:16:45.980
look, we should be thinking about the brain
link |
00:16:47.700
as time based, model memory based,
link |
00:16:52.020
and hierarchical processing.
link |
00:16:54.780
And that was a placeholder for a bunch of components
link |
00:16:58.180
that we would then plug into that.
link |
00:17:00.860
We still believe all those things I just said,
link |
00:17:02.620
but we now know so much more that I'm stopping to use
link |
00:17:06.980
the word hierarchical temporal memory yet
link |
00:17:08.180
because it's insufficient to capture the stuff we know.
link |
00:17:11.340
So again, it's not incorrect, but it's,
link |
00:17:13.660
I now know more and I would rather describe it
link |
00:17:15.820
more accurately.
link |
00:17:16.820
Yeah, so you're basically, we could think of HTM
link |
00:17:20.340
as emphasizing that there's three aspects of intelligence
link |
00:17:24.780
that are important to think about
link |
00:17:25.900
whatever the eventual theory it converges to.
link |
00:17:28.900
So in terms of time, how do you think of nature of time
link |
00:17:32.460
across different time scales?
link |
00:17:33.860
So you mentioned things changing,
link |
00:17:36.820
sensory inputs changing every 10, 20 minutes.
link |
00:17:39.140
What about every few minutes, every few months and years?
link |
00:17:42.100
Well, if you think about a neuroscience problem,
link |
00:17:44.820
the brain problem, neurons themselves can stay active
link |
00:17:49.620
for certain periods of time, parts of the brain
link |
00:17:52.780
where they stay active for minutes.
link |
00:17:54.260
You could hold a certain perception or an activity
link |
00:17:59.460
for a certain period of time,
link |
00:18:01.580
but most of them don't last that long.
link |
00:18:04.820
And so if you think about your thoughts
link |
00:18:07.180
are the activity of neurons,
link |
00:18:09.180
if you're gonna wanna involve something
link |
00:18:10.580
that happened a long time ago,
link |
00:18:11.980
even just this morning, for example,
link |
00:18:14.420
the neurons haven't been active throughout that time.
link |
00:18:16.420
So you have to store that.
link |
00:18:17.860
So if I ask you, what did you have for breakfast today?
link |
00:18:20.860
That is memory, that is you've built into your model
link |
00:18:23.660
the world now, you remember that.
link |
00:18:24.860
And that memory is in the synapses,
link |
00:18:27.780
is basically in the formation of synapses.
link |
00:18:29.980
And so you're sliding into what,
link |
00:18:34.700
you know, it's the different timescales.
link |
00:18:36.660
There's timescales of which we are like understanding
link |
00:18:39.060
my language and moving about and seeing things rapidly
link |
00:18:41.260
and over time, that's the timescales
link |
00:18:42.540
of activities of neurons.
link |
00:18:44.220
But if you wanna get in longer timescales,
link |
00:18:46.140
then it's more memory.
link |
00:18:47.100
And we have to invoke those memories to say,
link |
00:18:49.460
oh yes, well now I can remember what I had for breakfast
link |
00:18:51.740
because I stored that someplace.
link |
00:18:54.180
I may forget it tomorrow, but I'd store it for now.
link |
00:18:58.140
So does memory also need to have,
link |
00:19:02.820
so the hierarchical aspect of reality
link |
00:19:06.180
is not just about concepts, it's also about time?
link |
00:19:08.780
Do you think of it that way?
link |
00:19:10.260
Yeah, time is infused in everything.
link |
00:19:12.820
It's like you really can't separate it out.
link |
00:19:15.540
If I ask you, what is your, you know,
link |
00:19:18.700
how's the brain learn a model of this coffee cup here?
link |
00:19:21.340
I have a coffee cup and I'm at the coffee cup.
link |
00:19:23.220
I say, well, time is not an inherent property
link |
00:19:25.980
of the model I have of this cup,
link |
00:19:28.540
whether it's a visual model or a tactile model.
link |
00:19:31.460
I can sense it through time,
link |
00:19:32.580
but the model itself doesn't really have much time.
link |
00:19:34.900
If I asked you, if I said,
link |
00:19:36.420
well, what is the model of my cell phone?
link |
00:19:38.980
My brain has learned a model of the cell phone.
link |
00:19:40.740
So if you have a smartphone like this,
link |
00:19:43.380
and I said, well, this has time aspects to it.
link |
00:19:45.700
I have expectations when I turn it on,
link |
00:19:48.040
what's gonna happen, what or how long it's gonna take
link |
00:19:50.460
to do certain things, if I bring up an app,
link |
00:19:52.860
what sequences, and so I have,
link |
00:19:54.540
and it's like melodies in the world, you know?
link |
00:19:57.260
Melody has a sense of time.
link |
00:19:58.540
So many things in the world move and act,
link |
00:20:01.220
and there's a sense of time related to them.
link |
00:20:03.740
Some don't, but most things do actually.
link |
00:20:08.260
So it's sort of infused throughout the models of the world.
link |
00:20:12.100
You build a model of the world,
link |
00:20:13.700
you're learning the structure of the objects in the world,
link |
00:20:16.420
and you're also learning how those things change
link |
00:20:18.980
through time.
link |
00:20:20.780
Okay, so it really is just a fourth dimension
link |
00:20:23.900
that's infused deeply, and you have to make sure
link |
00:20:26.760
that your models of intelligence incorporate it.
link |
00:20:30.940
So, like you mentioned, the state of neuroscience
link |
00:20:34.840
is deeply empirical, a lot of data collection.
link |
00:20:37.800
It's, you know, that's where it is.
link |
00:20:41.420
You mentioned Thomas Kuhn, right?
link |
00:20:43.100
Yeah.
link |
00:20:44.580
And then you're proposing a theory of intelligence,
link |
00:20:48.020
and which is really the next step,
link |
00:20:50.460
the really important step to take,
link |
00:20:52.900
but why is HTM, or what we'll talk about soon,
link |
00:21:00.840
the right theory?
link |
00:21:03.700
So is it more in the, is it backed by intuition?
link |
00:21:07.700
Is it backed by evidence?
link |
00:21:09.920
Is it backed by a mixture of both?
link |
00:21:11.980
Is it kind of closer to where string theory is in physics,
link |
00:21:15.580
where there's mathematical components
link |
00:21:18.460
which show that, you know what,
link |
00:21:21.060
it seems that this, it fits together too well
link |
00:21:24.740
for it not to be true, which is where string theory is.
link |
00:21:28.100
Is that where you're kind of seeing?
link |
00:21:29.500
It's a mixture of all those things,
link |
00:21:30.740
although definitely where we are right now
link |
00:21:32.780
is definitely much more on the empirical side
link |
00:21:34.620
than, let's say, string theory.
link |
00:21:37.060
The way this goes about, we're theorists, right?
link |
00:21:39.280
So we look at all this data, and we're trying to come up
link |
00:21:41.580
with some sort of model that explains it, basically,
link |
00:21:44.340
and there's, unlike string theory,
link |
00:21:46.860
there's vast more amounts of empirical data here
link |
00:21:50.220
that I think than most physicists deal with.
link |
00:21:54.660
And so our challenge is to sort through that
link |
00:21:57.540
and figure out what kind of constructs would explain this.
link |
00:22:02.020
And when we have an idea,
link |
00:22:04.940
you come up with a theory of some sort,
link |
00:22:06.400
you have lots of ways of testing it.
link |
00:22:08.740
First of all, there are 100 years of assimilated,
link |
00:22:13.740
assimilated, unassimilated empirical data from neuroscience.
link |
00:22:16.620
So we go back and read papers,
link |
00:22:18.140
and we say, oh, did someone find this already?
link |
00:22:20.680
We can predict X, Y, and Z,
link |
00:22:23.280
and maybe no one's even talked about it
link |
00:22:25.220
since 1972 or something, but we go back and find that,
link |
00:22:28.180
and we say, oh, either it can support the theory
link |
00:22:31.140
or it can invalidate the theory.
link |
00:22:33.420
And we say, okay, we have to start over again.
link |
00:22:34.880
Oh, no, it's supportive, let's keep going with that one.
link |
00:22:38.140
So the way I kind of view it, when we do our work,
link |
00:22:42.260
we look at all this empirical data,
link |
00:22:45.460
and what I call it is a set of constraints.
link |
00:22:47.700
We're not interested in something
link |
00:22:48.700
that's biologically inspired.
link |
00:22:49.900
We're trying to figure out how the actual brain works.
link |
00:22:52.140
So every piece of empirical data
link |
00:22:53.660
is a constraint on a theory.
link |
00:22:55.500
In theory, if you have the correct theory,
link |
00:22:57.020
it needs to explain every pin, right?
link |
00:22:59.420
So we have this huge number of constraints on the problem,
link |
00:23:03.140
which initially makes it very, very difficult.
link |
00:23:05.960
If you don't have many constraints,
link |
00:23:07.220
you can make up stuff all the day.
link |
00:23:08.500
You can say, oh, here's an answer on how you can do this,
link |
00:23:10.200
you can do that, you can do this.
link |
00:23:11.360
But if you consider all biology as a set of constraints,
link |
00:23:13.760
all neuroscience as a set of constraints,
link |
00:23:15.580
and even if you're working in one little part
link |
00:23:17.240
of the neocortex, for example,
link |
00:23:18.380
there are hundreds and hundreds of constraints.
link |
00:23:20.620
These are empirical constraints
link |
00:23:22.460
that it's very, very difficult initially
link |
00:23:24.840
to come up with a theoretical framework for that.
link |
00:23:27.260
But when you do, and it solves all those constraints
link |
00:23:30.100
at once, you have a high confidence
link |
00:23:32.980
that you got something close to correct.
link |
00:23:35.660
It's just mathematically almost impossible not to be.
link |
00:23:39.160
So that's the curse and the advantage of what we have.
link |
00:23:43.900
The curse is we have to solve,
link |
00:23:45.260
we have to meet all these constraints, which is really hard.
link |
00:23:48.960
But when you do meet them,
link |
00:23:50.900
then you have a great confidence
link |
00:23:53.220
that you've discovered something.
link |
00:23:54.940
In addition, then we work with scientific labs.
link |
00:23:58.040
So we'll say, oh, there's something we can't find,
link |
00:24:00.000
we can predict something,
link |
00:24:01.260
but we can't find it anywhere in the literature.
link |
00:24:04.180
So we will then, we have people we've collaborated with,
link |
00:24:06.900
we'll say, sometimes they'll say, you know what?
link |
00:24:09.220
I have some collected data, which I didn't publish,
link |
00:24:11.740
but we can go back and look at it
link |
00:24:13.020
and see if we can find that,
link |
00:24:14.780
which is much easier than designing a new experiment.
link |
00:24:17.020
You know, neuroscience experiments take a long time, years.
link |
00:24:20.340
So, although some people are doing that now too.
link |
00:24:23.300
So, but between all of these things,
link |
00:24:27.740
I think it's a reasonable,
link |
00:24:30.020
actually a very, very good approach.
link |
00:24:31.620
We are blessed with the fact that we can test our theories
link |
00:24:35.020
out the yin yang here because there's so much
link |
00:24:37.100
unassimilar data and we can also falsify our theories
link |
00:24:39.640
very easily, which we do often.
link |
00:24:41.460
So it's kind of reminiscent to whenever that was
link |
00:24:44.380
with Copernicus, you know, when you figure out
link |
00:24:47.300
that the sun's at the center of the solar system
link |
00:24:51.140
as opposed to earth, the pieces just fall into place.
link |
00:24:54.900
Yeah, I think that's the general nature of aha moments
link |
00:24:59.580
is, and it's Copernicus, it could be,
link |
00:25:02.020
you could say the same thing about Darwin,
link |
00:25:05.220
you could say the same thing about, you know,
link |
00:25:07.580
about the double helix,
link |
00:25:09.660
that people have been working on a problem for so long
link |
00:25:12.780
and have all this data and they can't make sense of it,
link |
00:25:14.580
they can't make sense of it.
link |
00:25:15.820
But when the answer comes to you
link |
00:25:17.420
and everything falls into place,
link |
00:25:19.380
it's like, oh my gosh, that's it.
link |
00:25:21.660
That's got to be right.
link |
00:25:23.080
I asked both Jim Watson and Francis Crick about this.
link |
00:25:29.140
I asked them, you know, when you were working on
link |
00:25:31.700
trying to discover the structure of the double helix,
link |
00:25:35.760
and when you came up with the sort of the structure
link |
00:25:39.620
that ended up being correct, but it was sort of a guess,
link |
00:25:44.020
you know, it wasn't really verified yet.
link |
00:25:45.700
I said, did you know that it was right?
link |
00:25:48.460
And they both said, absolutely.
link |
00:25:50.220
So we absolutely knew it was right.
link |
00:25:51.860
And it doesn't matter if other people didn't believe it
link |
00:25:54.740
or not, we knew it was right.
link |
00:25:55.660
They'd get around to thinking it
link |
00:25:56.700
and agree with it eventually anyway.
link |
00:25:59.060
And that's the kind of thing you hear a lot with scientists
link |
00:26:01.300
who really are studying a difficult problem.
link |
00:26:04.220
And I feel that way too about our work.
link |
00:26:07.140
Have you talked to Crick or Watson about the problem
link |
00:26:10.700
you're trying to solve, the, of finding the DNA of the brain?
link |
00:26:15.940
Yeah, in fact, Francis Crick was very interested in this
link |
00:26:19.960
in the latter part of his life.
link |
00:26:21.540
And in fact, I got interested in brains
link |
00:26:23.780
by reading an essay he wrote in 1979
link |
00:26:26.900
called Thinking About the Brain.
link |
00:26:28.800
And that was when I decided I'm gonna leave my profession
link |
00:26:32.620
of computers and engineering and become a neuroscientist.
link |
00:26:35.580
Just reading that one essay from Francis Crick.
link |
00:26:37.660
I got to meet him later in life.
link |
00:26:41.640
I spoke at the Salk Institute and he was in the audience.
link |
00:26:44.660
And then I had a tea with him afterwards.
link |
00:26:48.820
He was interested in a different problem.
link |
00:26:50.620
He was focused on consciousness.
link |
00:26:53.380
The easy problem, right?
link |
00:26:54.260
Well, I think it's the red herring.
link |
00:26:58.640
And so we weren't really overlapping a lot there.
link |
00:27:02.260
Jim Watson, who's still alive,
link |
00:27:05.380
is also interested in this problem.
link |
00:27:07.420
And he was, when he was director
link |
00:27:09.020
of the Cold Spring Harbor Laboratories,
link |
00:27:12.420
he was really sort of behind moving in the direction
link |
00:27:15.140
of neuroscience there.
link |
00:27:16.580
And so he had a personal interest in this field.
link |
00:27:20.220
And I have met with him numerous times.
link |
00:27:23.620
And in fact, the last time was a little bit over a year ago,
link |
00:27:27.680
I gave a talk at Cold Spring Harbor Labs
link |
00:27:30.340
about the progress we were making in our work.
link |
00:27:34.620
And it was a lot of fun because he said,
link |
00:27:39.860
well, you wouldn't be coming here
link |
00:27:41.100
unless you had something important to say.
link |
00:27:42.380
So I'm gonna go attend your talk.
link |
00:27:44.740
So he sat in the very front row.
link |
00:27:46.620
Next to him was the director of the lab, Bruce Stillman.
link |
00:27:50.140
So these guys are in the front row of this auditorium.
link |
00:27:52.540
Nobody else in the auditorium wants to sit in the front row
link |
00:27:54.620
because there's Jim Watson and there's the director.
link |
00:27:56.980
And I gave a talk and then I had dinner with him afterwards.
link |
00:28:03.700
But there's a great picture of my colleague Subitai Amantak
link |
00:28:07.060
where I'm up there sort of like screaming the basics
link |
00:28:09.860
of this new framework we have.
link |
00:28:11.700
And Jim Watson's on the edge of his chair.
link |
00:28:13.780
He's literally on the edge of his chair,
link |
00:28:15.180
like intently staring up at the screen.
link |
00:28:17.820
And when he discovered the structure of DNA,
link |
00:28:21.740
the first public talk he gave
link |
00:28:23.800
was at Cold Spring Harbor Labs.
link |
00:28:25.940
And there's a picture, there's a famous picture
link |
00:28:27.460
of Jim Watson standing at the whiteboard
link |
00:28:29.340
with an overrated thing pointing at something,
link |
00:28:31.540
pointing at the double helix with his pointer.
link |
00:28:33.180
And it actually looks a lot like the picture of me.
link |
00:28:34.980
So there was a sort of funny,
link |
00:28:36.100
there's Arian talking about the brain
link |
00:28:37.460
and there's Jim Watson staring intently at it.
link |
00:28:39.300
And of course there with, whatever, 60 years earlier,
link |
00:28:41.620
he was standing pointing at the double helix.
link |
00:28:44.260
That's one of the great discoveries in all of,
link |
00:28:47.260
whatever, biology, science, all science and DNA.
link |
00:28:49.740
So it's funny that there's echoes of that in your presentation.
link |
00:28:54.540
Do you think, in terms of evolutionary timeline and history,
link |
00:28:58.360
the development of the neocortex was a big leap?
link |
00:29:01.960
Or is it just a small step?
link |
00:29:07.020
So like, if we ran the whole thing over again,
link |
00:29:09.780
from the birth of life on Earth,
link |
00:29:12.660
how likely would we develop the mechanism of the neocortex?
link |
00:29:15.260
Okay, well those are two separate questions.
link |
00:29:17.220
One is, was it a big leap?
link |
00:29:18.660
And one was how likely it is, okay?
link |
00:29:21.380
They're not necessarily related.
link |
00:29:22.880
Maybe correlated.
link |
00:29:23.720
Maybe correlated, maybe not.
link |
00:29:25.100
And we don't really have enough data
link |
00:29:26.100
to make a judgment about that.
link |
00:29:28.100
I would say definitely it was a big leap.
link |
00:29:29.980
And I can tell you why.
link |
00:29:30.980
I don't think it was just another incremental step.
link |
00:29:34.060
I don't get that at the moment.
link |
00:29:35.900
I don't really have any idea how likely it is.
link |
00:29:38.420
If we look at evolution,
link |
00:29:39.860
we have one data point, which is Earth, right?
link |
00:29:42.540
Life formed on Earth billions of years ago,
link |
00:29:45.220
whether it was introduced here or it created it here,
link |
00:29:48.100
or someone introduced it, we don't really know,
link |
00:29:49.560
but it was here early.
link |
00:29:51.220
It took a long, long time to get to multicellular life.
link |
00:29:55.140
And then for multicellular life,
link |
00:29:58.940
it took a long, long time to get the neocortex.
link |
00:30:02.300
And we've only had the neocortex for a few 100,000 years.
link |
00:30:05.460
So that's like nothing, okay?
link |
00:30:08.000
So is it likely?
link |
00:30:09.600
Well, it certainly isn't something
link |
00:30:10.740
that happened right away on Earth.
link |
00:30:13.560
And there were multiple steps to get there.
link |
00:30:15.200
So I would say it's probably not gonna be something
link |
00:30:17.220
that would happen instantaneously
link |
00:30:18.260
on other planets that might have life.
link |
00:30:20.620
It might take several billion years on average.
link |
00:30:23.160
Is it likely?
link |
00:30:24.380
I don't know, but you'd have to survive
link |
00:30:25.740
for several billion years to find out.
link |
00:30:27.900
Probably.
link |
00:30:29.340
Is it a big leap?
link |
00:30:30.260
Yeah, I think it is a qualitative difference
link |
00:30:35.500
in all other evolutionary steps.
link |
00:30:37.860
I can try to describe that if you'd like.
link |
00:30:39.820
Sure, in which way?
link |
00:30:41.980
Yeah, I can tell you how.
link |
00:30:43.940
Pretty much, let's start with a little preface.
link |
00:30:47.740
Many of the things that humans are able to do
link |
00:30:50.500
do not have obvious survival advantages precedent.
link |
00:30:58.620
We could create music, is that,
link |
00:31:00.260
is there a really survival advantage to that?
link |
00:31:02.700
Maybe, maybe not.
link |
00:31:03.900
What about mathematics?
link |
00:31:04.900
Is there a real survival advantage to mathematics?
link |
00:31:07.020
Well, you could stretch it.
link |
00:31:09.340
You can try to figure these things out, right?
link |
00:31:13.140
But most of evolutionary history,
link |
00:31:14.800
everything had immediate survival advantages to it.
link |
00:31:18.700
So, I'll tell you a story, which I like,
link |
00:31:22.020
may or may not be true, but the story goes as follows.
link |
00:31:29.140
Organisms have been evolving for,
link |
00:31:30.860
since the beginning of life here on Earth,
link |
00:31:33.740
and adding this sort of complexity onto that,
link |
00:31:35.700
and this sort of complexity onto that,
link |
00:31:36.860
and the brain itself is evolved this way.
link |
00:31:39.700
In fact, there's old parts, and older parts,
link |
00:31:42.420
and older, older parts of the brain
link |
00:31:43.740
that kind of just keeps calming on new things,
link |
00:31:45.500
and we keep adding capabilities.
link |
00:31:47.260
When we got to the neocortex,
link |
00:31:48.700
initially it had a very clear survival advantage
link |
00:31:52.500
in that it produced better vision,
link |
00:31:54.380
and better hearing, and better touch,
link |
00:31:55.700
and maybe, and so on.
link |
00:31:57.780
But what I think happens is that evolution discovered,
link |
00:32:01.140
it took a mechanism, and this is in our recent theories,
link |
00:32:05.100
but it took a mechanism evolved a long time ago
link |
00:32:08.140
for navigating in the world, for knowing where you are.
link |
00:32:10.380
These are the so called grid cells and place cells
link |
00:32:13.360
of an old part of the brain.
link |
00:32:15.160
And it took that mechanism for building maps of the world,
link |
00:32:20.900
and knowing where you are on those maps,
link |
00:32:22.580
and how to navigate those maps,
link |
00:32:24.140
and turns it into a sort of a slimmed down,
link |
00:32:27.060
idealized version of it.
link |
00:32:29.540
And that idealized version could now apply
link |
00:32:31.600
to building maps of other things.
link |
00:32:32.820
Maps of coffee cups, and maps of phones,
link |
00:32:35.100
maps of mathematics.
link |
00:32:36.460
Concepts almost.
link |
00:32:37.300
Concepts, yes, and not just almost, exactly.
link |
00:32:40.260
And so, and it just started replicating this stuff, right?
link |
00:32:44.140
You just think more, and more, and more.
link |
00:32:45.220
So we went from being sort of dedicated purpose
link |
00:32:48.780
neural hardware to solve certain problems
link |
00:32:51.460
that are important to survival,
link |
00:32:53.200
to a general purpose neural hardware
link |
00:32:55.820
that could be applied to all problems.
link |
00:32:58.100
And now it's escaped the orbit of survival.
link |
00:33:02.600
We are now able to apply it to things
link |
00:33:04.460
which we find enjoyment,
link |
00:33:08.700
but aren't really clearly survival characteristics.
link |
00:33:13.700
And that it seems to only have happened in humans,
link |
00:33:16.740
to the large extent.
link |
00:33:19.260
And so that's what's going on,
link |
00:33:20.980
where we sort of have,
link |
00:33:22.940
we've sort of escaped the gravity of evolutionary pressure,
link |
00:33:26.360
in some sense, in the neocortex.
link |
00:33:28.620
And it now does things which are not,
link |
00:33:31.540
that are really interesting,
link |
00:33:32.780
discovering models of the universe,
link |
00:33:34.340
which may not really help us.
link |
00:33:36.100
Does it matter?
link |
00:33:37.100
How does it help us surviving,
link |
00:33:38.600
knowing that there might be multiverses,
link |
00:33:40.240
or that there might be the age of the universe,
link |
00:33:42.940
or how do various stellar things occur?
link |
00:33:46.140
It doesn't really help us survive at all.
link |
00:33:47.820
But we enjoy it, and that's what happened.
link |
00:33:50.460
Or at least not in the obvious way, perhaps.
link |
00:33:53.300
It is required,
link |
00:33:56.200
if you look at the entire universe in an evolutionary way,
link |
00:33:58.540
it's required for us to do interplanetary travel,
link |
00:34:00.900
and therefore survive past our own sun.
link |
00:34:03.140
But you know, let's not get too.
link |
00:34:04.500
Yeah, but evolution works at one time frame,
link |
00:34:07.220
it's survival, if you think of survival of the phenotype,
link |
00:34:11.340
survival of the individual.
link |
00:34:13.180
What you're talking about there is spans well beyond that.
link |
00:34:16.360
So there's no genetic,
link |
00:34:18.740
I'm not transferring any genetic traits to my children
link |
00:34:23.420
that are gonna help them survive better on Mars.
link |
00:34:26.540
Totally different mechanism, that's right.
link |
00:34:28.260
So let's get into the new, as you've mentioned,
link |
00:34:31.340
this idea of the, I don't know if you have a nice name,
link |
00:34:34.860
thousand.
link |
00:34:35.700
We call it the thousand brain theory of intelligence.
link |
00:34:37.340
I like it.
link |
00:34:38.180
Can you talk about this idea of a spatial view of concepts
link |
00:34:43.620
and so on?
link |
00:34:44.460
Yeah, so can I just describe sort of the,
link |
00:34:46.500
there's an underlying core discovery,
link |
00:34:49.300
which then everything comes from that.
link |
00:34:51.140
That's a very simple, this is really what happened.
link |
00:34:55.660
We were deep into problems about understanding
link |
00:34:58.580
how we build models of stuff in the world
link |
00:35:00.540
and how we make predictions about things.
link |
00:35:03.020
And I was holding a coffee cup just like this in my hand.
link |
00:35:07.220
And my finger was touching the side, my index finger.
link |
00:35:10.540
And then I moved it to the top
link |
00:35:12.700
and I was gonna feel the rim at the top of the cup.
link |
00:35:15.460
And I asked myself a very simple question.
link |
00:35:18.280
I said, well, first of all, I say,
link |
00:35:20.100
I know that my brain predicts what it's gonna feel
link |
00:35:22.260
before it touches it.
link |
00:35:23.300
You can just think about it and imagine it.
link |
00:35:26.040
And so we know that the brain's making predictions
link |
00:35:27.660
all the time.
link |
00:35:28.500
So the question is, what does it take to predict that?
link |
00:35:31.540
And there's a very interesting answer.
link |
00:35:33.620
First of all, it says the brain has to know
link |
00:35:35.400
it's touching a coffee cup.
link |
00:35:36.500
It has to have a model of a coffee cup.
link |
00:35:38.020
It needs to know where the finger currently is
link |
00:35:41.020
on the cup relative to the cup.
link |
00:35:43.260
Because when I make a movement,
link |
00:35:44.420
it needs to know where it's going to be on the cup
link |
00:35:46.340
after the movement is completed relative to the cup.
link |
00:35:50.380
And then it can make a prediction
link |
00:35:51.900
about what it's gonna sense.
link |
00:35:53.340
So this told me that the neocortex,
link |
00:35:54.960
which is making this prediction,
link |
00:35:56.380
needs to know that it's sensing it's touching a cup.
link |
00:35:59.420
And it needs to know the location of my finger
link |
00:36:01.420
relative to that cup in a reference frame of the cup.
link |
00:36:04.380
It doesn't matter where the cup is relative to my body.
link |
00:36:06.300
It doesn't matter its orientation.
link |
00:36:08.260
None of that matters.
link |
00:36:09.160
It's where my finger is relative to the cup,
link |
00:36:10.940
which tells me then that the neocortex
link |
00:36:13.540
has a reference frame that's anchored to the cup.
link |
00:36:17.340
Because otherwise I wouldn't be able to say the location
link |
00:36:19.280
and I wouldn't be able to predict my new location.
link |
00:36:21.500
And then we quickly, very instantly can say,
link |
00:36:24.120
well, every part of my skin could touch this cup.
link |
00:36:26.240
And therefore every part of my skin is making predictions
link |
00:36:28.100
and every part of my skin must have a reference frame
link |
00:36:30.940
that it's using to make predictions.
link |
00:36:33.520
So the big idea is that throughout the neocortex,
link |
00:36:39.500
there are, everything is being stored
link |
00:36:44.940
and referenced in reference frames.
link |
00:36:46.740
You can think of them like XYZ reference frames,
link |
00:36:48.820
but they're not like that.
link |
00:36:50.380
We know a lot about the neural mechanisms for this,
link |
00:36:52.060
but the brain thinks in reference frames.
link |
00:36:54.860
And as an engineer, if you're an engineer,
link |
00:36:56.700
this is not surprising.
link |
00:36:57.740
You'd say, if I wanted to build a CAD model
link |
00:37:00.340
of the coffee cup, well, I would bring it up
link |
00:37:02.120
and some CAD software, and I would assign
link |
00:37:04.100
some reference frame and say this features
link |
00:37:05.460
at this locations and so on.
link |
00:37:06.980
But the fact that this, the idea that this is occurring
link |
00:37:09.700
throughout the neocortex everywhere, it was a novel idea.
link |
00:37:14.360
And then a zillion things fell into place after that,
link |
00:37:19.080
a zillion.
link |
00:37:19.940
So now we think about the neocortex
link |
00:37:21.860
as processing information quite differently
link |
00:37:23.420
than we used to do it.
link |
00:37:24.260
We used to think about the neocortex
link |
00:37:25.540
as processing sensory data and extracting features
link |
00:37:28.700
from that sensory data and then extracting features
link |
00:37:30.860
from the features, very much like a deep learning network
link |
00:37:33.580
does today.
link |
00:37:34.900
But that's not how the brain works at all.
link |
00:37:36.620
The brain works by assigning everything,
link |
00:37:39.300
every input, everything to reference frames.
link |
00:37:41.860
And there are thousands, hundreds of thousands
link |
00:37:44.380
of them active at once in your neocortex.
link |
00:37:47.660
It's a surprising thing to think about,
link |
00:37:49.580
but once you sort of internalize this,
link |
00:37:51.060
you understand that it explains almost every,
link |
00:37:54.380
almost all the mysteries we've had about this structure.
link |
00:37:57.780
So one of the consequences of that
link |
00:38:00.200
is that every small part of the neocortex,
link |
00:38:02.620
say a millimeter square, and there's 150,000 of those.
link |
00:38:06.340
So it's about 150,000 square millimeters.
link |
00:38:08.620
If you take every little square millimeter of the cortex,
link |
00:38:11.380
it's got some input coming into it
link |
00:38:13.260
and it's gonna have reference frames
link |
00:38:14.940
where it's assigned that input to.
link |
00:38:16.800
And each square millimeter can learn
link |
00:38:19.320
complete models of objects.
link |
00:38:20.980
So what do I mean by that?
link |
00:38:22.020
If I'm touching the coffee cup,
link |
00:38:23.300
well, if I just touch it in one place,
link |
00:38:25.580
I can't learn what this coffee cup is
link |
00:38:27.180
because I'm just feeling one part.
link |
00:38:28.980
But if I move it around the cup
link |
00:38:31.060
and touch it at different areas,
link |
00:38:32.540
I can build up a complete model of the cup
link |
00:38:34.060
because I'm now filling in that three dimensional map,
link |
00:38:36.700
which is the coffee cup.
link |
00:38:37.540
I can say, oh, what am I feeling
link |
00:38:38.660
at all these different locations?
link |
00:38:39.900
That's the basic idea, it's more complicated than that.
link |
00:38:43.020
But so through time, and we talked about time earlier,
link |
00:38:46.220
through time, even a single column,
link |
00:38:48.180
which is only looking at, or a single part of the cortex,
link |
00:38:50.300
which is only looking at a small part of the world,
link |
00:38:52.720
can build up a complete model of an object.
link |
00:38:55.060
And so if you think about the part of the brain,
link |
00:38:57.100
which is getting input from all my fingers,
link |
00:38:59.100
so they're spread across the top of your head here.
link |
00:39:01.700
This is the somatosensory cortex.
link |
00:39:04.040
There's columns associated
link |
00:39:05.180
with all the different areas of my skin.
link |
00:39:07.380
And what we believe is happening
link |
00:39:10.100
is that all of them are building models of this cup,
link |
00:39:12.900
every one of them, or things.
link |
00:39:15.340
They're not all building,
link |
00:39:16.620
not every column or every part of the cortex
link |
00:39:18.180
builds models of everything,
link |
00:39:19.500
but they're all building models of something.
link |
00:39:21.700
And so you have, so when I touch this cup with my hand,
link |
00:39:26.700
there are multiple models of the cup being invoked.
link |
00:39:28.980
If I look at it with my eyes,
link |
00:39:30.460
there are, again, many models of the cup being invoked,
link |
00:39:32.540
because each part of the visual system,
link |
00:39:34.300
the brain doesn't process an image.
link |
00:39:35.820
That's a misleading idea.
link |
00:39:38.740
It's just like your fingers touching the cup,
link |
00:39:40.460
so different parts of my retina
link |
00:39:41.300
are looking at different parts of the cup.
link |
00:39:42.980
And thousands and thousands of models of the cup
link |
00:39:45.540
are being invoked at once.
link |
00:39:47.380
And they're all voting with each other,
link |
00:39:48.900
trying to figure out what's going on.
link |
00:39:50.140
So that's why we call it the thousand brains theory
link |
00:39:51.740
of intelligence, because there isn't one model of a cup.
link |
00:39:54.700
There are thousands of models of this cup.
link |
00:39:56.300
There are thousands of models of your cellphone
link |
00:39:57.940
and about cameras and microphones and so on.
link |
00:40:00.860
It's a distributed modeling system,
link |
00:40:02.860
which is very different
link |
00:40:03.700
than the way people have thought about it.
link |
00:40:04.860
And so that's a really compelling and interesting idea.
link |
00:40:07.340
I have two first questions.
link |
00:40:08.700
So one, on the ensemble part of everything coming together,
link |
00:40:12.060
you have these thousand brains.
link |
00:40:14.860
How do you know which one has done the best job
link |
00:40:17.900
of forming the...
link |
00:40:18.740
Great question.
link |
00:40:19.580
Let me try to explain it.
link |
00:40:20.420
There's a problem that's known in neuroscience
link |
00:40:23.500
called the sensor fusion problem.
link |
00:40:25.220
Yes.
link |
00:40:26.060
And so the idea is there's something like,
link |
00:40:27.740
oh, the image comes from the eye.
link |
00:40:29.140
There's a picture on the retina
link |
00:40:30.620
and then it gets projected to the neocortex.
link |
00:40:32.380
Oh, by now it's all spread out all over the place
link |
00:40:35.100
and it's kind of squirrely and distorted
link |
00:40:37.100
and pieces are all over the...
link |
00:40:39.020
It doesn't look like a picture anymore.
link |
00:40:40.900
When does it all come back together again?
link |
00:40:43.660
Or you might say, well, yes,
link |
00:40:45.380
but I also have sounds or touches associated with the cup.
link |
00:40:48.620
So I'm seeing the cup and touching the cup.
link |
00:40:50.660
How do they get combined together again?
link |
00:40:52.620
So it's called the sensor fusion problem.
link |
00:40:54.260
As if all these disparate parts
link |
00:40:55.860
have to be brought together into one model someplace.
link |
00:40:59.020
That's the wrong idea.
link |
00:41:01.140
The right idea is that you've got all these guys voting.
link |
00:41:03.500
There's auditory models of the cup.
link |
00:41:05.420
There's visual models of the cup.
link |
00:41:06.620
There's tactile models of the cup.
link |
00:41:09.860
In the vision system,
link |
00:41:10.700
there might be ones that are more focused on black and white
link |
00:41:12.580
and ones focusing on color.
link |
00:41:13.620
It doesn't really matter.
link |
00:41:14.460
There's just thousands and thousands of models of this cup.
link |
00:41:17.020
And they vote.
link |
00:41:17.900
They don't actually come together in one spot.
link |
00:41:20.620
Just literally think of it this way.
link |
00:41:21.900
Imagine you have these columns
link |
00:41:24.100
that are like about the size of a little piece of spaghetti.
link |
00:41:26.660
Like a two and a half millimeters tall
link |
00:41:28.500
and about a millimeter in wide.
link |
00:41:30.020
They're not physical, but you could think of them that way.
link |
00:41:33.300
And each one's trying to guess what this thing is
link |
00:41:35.300
or touching.
link |
00:41:36.140
Now, they can do a pretty good job
link |
00:41:38.060
if they're allowed to move over time.
link |
00:41:40.060
So I can reach my hand into a black box
link |
00:41:41.620
and move my finger around an object.
link |
00:41:43.540
And if I touch enough spaces, I go, okay,
link |
00:41:45.540
now I know what it is.
link |
00:41:46.980
But often we don't do that.
link |
00:41:48.300
Often I can just reach and grab something with my hand
link |
00:41:49.940
all at once and I get it.
link |
00:41:51.020
Or if I had to look through the world through a straw,
link |
00:41:53.740
so I'm only invoking one little column,
link |
00:41:55.860
I can only see part of something
link |
00:41:56.700
because I have to move the straw around.
link |
00:41:58.140
But if I open my eyes, I see the whole thing at once.
link |
00:42:00.460
So what we think is going on
link |
00:42:01.460
is all these little pieces of spaghetti,
link |
00:42:03.180
if you will, all these little columns in the cortex,
link |
00:42:05.300
are all trying to guess what it is that they're sensing.
link |
00:42:08.580
They'll do a better guess if they have time
link |
00:42:10.740
and can move over time.
link |
00:42:11.700
So if I move my eyes, I move my fingers.
link |
00:42:13.620
But if they don't, they have a poor guess.
link |
00:42:16.580
It's a probabilistic guess of what they might be touching.
link |
00:42:20.060
Now, imagine they can post their probability
link |
00:42:22.940
at the top of a little piece of spaghetti.
link |
00:42:24.580
Each one of them says,
link |
00:42:25.420
I think, and it's not really a probability distribution.
link |
00:42:27.420
It's more like a set of possibilities.
link |
00:42:29.460
In the brain, it doesn't work as a probability distribution.
link |
00:42:31.980
It works as more like what we call a union.
link |
00:42:34.020
So you could say, and one column says,
link |
00:42:35.860
I think it could be a coffee cup,
link |
00:42:37.540
a soda can, or a water bottle.
link |
00:42:39.940
And another column says,
link |
00:42:40.900
I think it could be a coffee cup
link |
00:42:42.300
or a telephone or a camera or whatever, right?
link |
00:42:46.460
And all these guys are saying what they think it might be.
link |
00:42:49.940
And there's these long range connections
link |
00:42:51.620
in certain layers in the cortex.
link |
00:42:53.460
So there's in some layers in some cells types
link |
00:42:56.660
in each column, send the projections across the brain.
link |
00:43:00.060
And that's the voting occurs.
link |
00:43:01.740
And so there's a simple associative memory mechanism.
link |
00:43:04.100
We've described this in a recent paper
link |
00:43:06.140
and we've modeled this that says,
link |
00:43:09.500
they can all quickly settle on the only
link |
00:43:11.940
or the one best answer for all of them.
link |
00:43:14.900
If there is a single best answer,
link |
00:43:16.420
they all vote and say, yep, it's gotta be the coffee cup.
link |
00:43:18.940
And at that point, they all know it's a coffee cup.
link |
00:43:21.060
And at that point, everyone acts as if it's a coffee cup.
link |
00:43:23.380
They're like, yep, we know it's a coffee,
link |
00:43:24.220
even though I've only seen one little piece of this world,
link |
00:43:26.380
I know it's a coffee cup I'm touching
link |
00:43:27.700
or I'm seeing or whatever.
link |
00:43:28.980
And so you can think of all these columns
link |
00:43:30.900
are looking at different parts in different places,
link |
00:43:33.020
different sensory input, different locations,
link |
00:43:35.220
they're all different.
link |
00:43:36.180
But this layer that's doing the voting, it solidifies.
link |
00:43:40.460
It's just like it crystallizes and says,
link |
00:43:42.260
oh, we all know what we're doing.
link |
00:43:44.140
And so you don't bring these models together in one model,
link |
00:43:46.460
you just vote and there's a crystallization of the vote.
link |
00:43:49.140
Great, that's at least a compelling way
link |
00:43:51.780
to think about the way you form a model of the world.
link |
00:43:58.180
Now, you talk about a coffee cup.
link |
00:44:00.420
Do you see this, as far as I understand,
link |
00:44:03.220
you are proposing this as well,
link |
00:44:04.660
that this extends to much more than coffee cups?
link |
00:44:06.900
Yeah.
link |
00:44:07.740
It does.
link |
00:44:09.540
Or at least the physical world,
link |
00:44:10.780
it expands to the world of concepts.
link |
00:44:14.100
Yeah, it does.
link |
00:44:15.020
And well, first, the primary thing is evidence for that
link |
00:44:18.220
is that the regions of the neocortex
link |
00:44:20.700
that are associated with language
link |
00:44:22.340
or high level thought or mathematics
link |
00:44:23.860
or things like that,
link |
00:44:24.700
they look like the regions of the neocortex
link |
00:44:26.180
that process vision, hearing, and touch.
link |
00:44:28.300
They don't look any different.
link |
00:44:29.700
Or they look only marginally different.
link |
00:44:32.820
And so one would say, well, if Vernon Mountcastle,
link |
00:44:36.420
who proposed that all the parts of the neocortex
link |
00:44:38.860
do the same thing, if he's right,
link |
00:44:41.060
then the parts that are doing language
link |
00:44:42.820
or mathematics or physics
link |
00:44:44.540
are working on the same principle.
link |
00:44:45.700
They must be working on the principle of reference frames.
link |
00:44:48.500
So that's a little odd thought.
link |
00:44:51.820
But of course, we had no prior idea
link |
00:44:53.940
how these things happen.
link |
00:44:55.020
So let's go with that.
link |
00:44:57.340
And we, in our recent paper,
link |
00:44:59.900
we talked a little bit about that.
link |
00:45:01.620
I've been working on it more since.
link |
00:45:02.820
I have better ideas about it now.
link |
00:45:05.380
I'm sitting here very confident
link |
00:45:06.980
that that's what's happening.
link |
00:45:08.020
And I can give you some examples
link |
00:45:09.260
that help you think about that.
link |
00:45:11.220
It's not we understand it completely,
link |
00:45:12.500
but I understand it better than I've described it
link |
00:45:14.300
in any paper so far.
link |
00:45:15.660
So, but we did put that idea out there.
link |
00:45:17.700
It says, okay, this is,
link |
00:45:18.940
it's a good place to start, you know?
link |
00:45:22.620
And the evidence would suggest it's how it's happening.
link |
00:45:24.900
And then we can start tackling that problem
link |
00:45:26.660
one piece at a time.
link |
00:45:27.500
Like, what does it mean to do high level thought?
link |
00:45:29.060
What does it mean to do language?
link |
00:45:30.020
How would that fit into a reference frame framework?
link |
00:45:34.220
Yeah, so there's a,
link |
00:45:35.980
I don't know if you could tell me if there's a connection,
link |
00:45:37.580
but there's an app called Anki
link |
00:45:40.180
that helps you remember different concepts.
link |
00:45:42.420
And they talk about like a memory palace
link |
00:45:45.100
that helps you remember completely random concepts
link |
00:45:47.780
by trying to put them in a physical space in your mind
link |
00:45:51.380
and putting them next to each other.
link |
00:45:52.220
It's called the method of loci.
link |
00:45:53.580
Loci, yeah.
link |
00:45:54.700
For some reason, that seems to work really well.
link |
00:45:57.580
Now, that's a very narrow kind of application
link |
00:45:59.420
of just remembering some facts.
link |
00:46:00.580
But that's a very, very telling one.
link |
00:46:03.260
Yes, exactly.
link |
00:46:04.100
So this seems like you're describing a mechanism
link |
00:46:06.740
why this seems to work.
link |
00:46:09.620
So basically the way what we think is going on
link |
00:46:11.820
is all things you know, all concepts, all ideas,
link |
00:46:15.060
words, everything you know are stored in reference frames.
link |
00:46:20.460
And so if you want to remember something,
link |
00:46:24.300
you have to basically navigate through a reference frame
link |
00:46:26.860
the same way a rat navigates through a maze
link |
00:46:28.620
and the same way my finger navigates to this coffee cup.
link |
00:46:31.420
You are moving through some space.
link |
00:46:33.500
And so if you have a random list of things
link |
00:46:35.900
you were asked to remember,
link |
00:46:37.460
by assigning them to a reference frame,
link |
00:46:39.300
you've already know very well to see your house, right?
link |
00:46:42.100
And the idea of the method of loci is you can say,
link |
00:46:43.580
okay, in my lobby, I'm going to put this thing.
link |
00:46:45.820
And then the bedroom, I put this one.
link |
00:46:47.660
I go down the hall, I put this thing.
link |
00:46:48.940
And then you want to recall those facts
link |
00:46:50.820
or recall those things.
link |
00:46:51.660
You just walk mentally, you walk through your house.
link |
00:46:54.100
You're mentally moving through a reference frame
link |
00:46:56.540
that you already had.
link |
00:46:57.660
And that tells you,
link |
00:46:59.260
there's two things that are really important about that.
link |
00:47:00.580
It tells us the brain prefers to store things
link |
00:47:02.740
in reference frames.
link |
00:47:03.940
And that the method of recalling things
link |
00:47:06.820
or thinking, if you will,
link |
00:47:08.220
is to move mentally through those reference frames.
link |
00:47:11.500
You could move physically through some reference frames,
link |
00:47:13.540
like I could physically move through the reference frame
link |
00:47:15.220
of this coffee cup.
link |
00:47:16.300
I can also mentally move through the reference frame
link |
00:47:17.900
of the coffee cup, imagining me touching it.
link |
00:47:19.980
But I can also mentally move my house.
link |
00:47:22.420
And so now we can ask yourself,
link |
00:47:24.660
or are all concepts stored this way?
link |
00:47:26.740
There was some recent research using human subjects
link |
00:47:31.380
in fMRI, and I'm going to apologize for not knowing
link |
00:47:33.540
the name of the scientists who did this.
link |
00:47:36.660
But what they did is they put humans in this fMRI machine,
link |
00:47:41.060
which is one of these imaging machines.
link |
00:47:42.780
And they gave the humans tasks to think about birds.
link |
00:47:46.460
So they had different types of birds,
link |
00:47:47.780
and birds that look big and small,
link |
00:47:49.660
and long necks and long legs, things like that.
link |
00:47:52.220
And what they could tell from the fMRI
link |
00:47:55.260
was a very clever experiment.
link |
00:47:57.580
You get to tell when humans were thinking about the birds,
link |
00:48:00.780
that the birds, the knowledge of birds
link |
00:48:03.580
was arranged in a reference frame,
link |
00:48:05.500
similar to the ones that are used
link |
00:48:07.100
when you navigate in a room.
link |
00:48:08.980
That these are called grid cells,
link |
00:48:10.340
and there are grid cell like patterns of activity
link |
00:48:12.820
in the neocortex when they do this.
link |
00:48:15.380
So it's a very clever experiment.
link |
00:48:18.980
And what it basically says,
link |
00:48:20.180
that even when you're thinking about something abstract,
link |
00:48:22.140
and you're not really thinking about it as a reference frame,
link |
00:48:24.700
it tells us the brain is actually using a reference frame.
link |
00:48:26.980
And it's using the same neural mechanisms.
link |
00:48:28.780
These grid cells are the basic same neural mechanism
link |
00:48:30.780
that we propose that grid cells,
link |
00:48:32.860
which exist in the old part of the brain,
link |
00:48:34.980
the entorhinal cortex, that that mechanism
link |
00:48:37.340
is now similar mechanism is used throughout the neocortex.
link |
00:48:40.060
It's the same nature to preserve this interesting way
link |
00:48:43.180
of creating reference frames.
link |
00:48:44.580
And so now they have empirical evidence
link |
00:48:46.940
that when you think about concepts like birds,
link |
00:48:49.500
that you're using reference frames
link |
00:48:51.220
that are built on grid cells.
link |
00:48:53.180
So that's similar to the method of loci,
link |
00:48:55.180
but in this case, the birds are related.
link |
00:48:56.820
So they create their own reference frame,
link |
00:48:58.620
which is consistent with bird space.
link |
00:49:01.100
And when you think about something, you go through that.
link |
00:49:03.540
You can make the same example,
link |
00:49:04.820
let's take mathematics.
link |
00:49:06.620
Let's say you wanna prove a conjecture.
link |
00:49:09.260
What is a conjecture?
link |
00:49:10.100
A conjecture is a statement you believe to be true,
link |
00:49:13.300
but you haven't proven it.
link |
00:49:15.140
And so it might be an equation.
link |
00:49:16.540
I wanna show that this is equal to that.
link |
00:49:19.140
And you have some places you start with.
link |
00:49:21.180
You say, well, I know this is true,
link |
00:49:22.340
and I know this is true.
link |
00:49:23.420
And I think that maybe to get to the final proof,
link |
00:49:25.900
I need to go through some intermediate results.
link |
00:49:28.700
What I believe is happening is literally these equations
link |
00:49:33.140
or these points are assigned to a reference frame,
link |
00:49:36.380
a mathematical reference frame.
link |
00:49:37.980
And when you do mathematical operations,
link |
00:49:39.820
a simple one might be multiply or divide,
link |
00:49:41.660
but you might be a little plus transform or something else.
link |
00:49:44.060
That is like a movement in the reference frame of the math.
link |
00:49:47.500
And so you're literally trying to discover a path
link |
00:49:50.260
from one location to another location
link |
00:49:52.660
in a space of mathematics.
link |
00:49:56.140
And if you can get to these intermediate results,
link |
00:49:58.220
then you know your map is pretty good,
link |
00:50:00.420
and you know you're using the right operations.
link |
00:50:02.940
Much of what we think about is solving hard problems
link |
00:50:05.940
is designing the correct reference frame for that problem,
link |
00:50:08.820
figuring out how to organize the information
link |
00:50:11.100
and what behaviors I wanna use in that space
link |
00:50:14.300
to get me there.
link |
00:50:16.220
Yeah, so if you dig in an idea of this reference frame,
link |
00:50:19.260
whether it's the math, you start a set of axioms
link |
00:50:21.700
to try to get to proving the conjecture.
link |
00:50:25.140
Can you try to describe, maybe take a step back,
link |
00:50:28.140
how you think of the reference frame in that context?
link |
00:50:30.660
Is it the reference frame that the axioms are happy in?
link |
00:50:36.140
Is it the reference frame that might contain everything?
link |
00:50:38.780
Is it a changing thing as you?
link |
00:50:41.780
You have many, many reference frames.
link |
00:50:43.140
I mean, in fact, the way the theory,
link |
00:50:44.580
the thousand brain theory of intelligence says
link |
00:50:46.140
that every single thing in the world
link |
00:50:47.380
has its own reference frame.
link |
00:50:48.300
So every word has its own reference frames.
link |
00:50:50.860
And we can talk about this.
link |
00:50:52.940
The mathematics work out,
link |
00:50:54.460
this is no problem for neurons to do this.
link |
00:50:55.940
But how many reference frames does a coffee cup have?
link |
00:50:58.740
Well, it's on a table.
link |
00:51:00.140
Let's say you ask how many reference frames
link |
00:51:03.700
could a column in my finger
link |
00:51:06.020
that's touching the coffee cup have?
link |
00:51:07.420
Because there are many, many copy,
link |
00:51:09.060
there are many, many models of the coffee cup.
link |
00:51:10.500
So the coffee, there is no one model of a coffee cup.
link |
00:51:13.020
There are many models of a coffee cup.
link |
00:51:14.220
And you could say, well,
link |
00:51:15.220
how many different things can my finger learn?
link |
00:51:17.260
Is this the question you want to ask?
link |
00:51:19.540
Imagine I say every concept, every idea,
link |
00:51:21.780
everything you've ever know about that you can say,
link |
00:51:23.860
I know that thing has a reference frame
link |
00:51:27.260
associated with it.
link |
00:51:28.180
And what we do when we build composite objects,
link |
00:51:30.180
we assign reference frames
link |
00:51:32.460
to point another reference frame.
link |
00:51:33.940
So my coffee cup has multiple components to it.
link |
00:51:37.060
It's got a limb, it's got a cylinder, it's got a handle.
link |
00:51:40.660
And those things have their own reference frames
link |
00:51:42.820
and they're assigned to a master reference frame,
link |
00:51:45.060
which is called this cup.
link |
00:51:46.380
And now I have this Numenta logo on it.
link |
00:51:48.180
Well, that's something that exists elsewhere in the world.
link |
00:51:50.420
It's its own thing.
link |
00:51:51.260
So it has its own reference frame.
link |
00:51:52.300
So we now have to say,
link |
00:51:53.140
how can I assign the Numenta logo reference frame
link |
00:51:56.740
onto the cylinder or onto the coffee cup?
link |
00:51:59.180
So it's all, we talked about this in the paper
link |
00:52:01.500
that came out in December of this last year.
link |
00:52:06.860
The idea of how you can assign reference frames
link |
00:52:08.780
to reference frames, how neurons could do this.
link |
00:52:10.540
So, well, my question is,
link |
00:52:12.620
even though you mentioned reference frames a lot,
link |
00:52:14.740
I almost feel it's really useful to dig into
link |
00:52:16.940
how you think of what a reference frame is.
link |
00:52:20.140
I mean, it was already helpful for me to understand
link |
00:52:22.020
that you think of reference frames
link |
00:52:23.700
as something there is a lot of.
link |
00:52:26.340
Okay, so let's just say that we're gonna have
link |
00:52:28.780
some neurons in the brain, not many, actually,
link |
00:52:31.060
10,000, 20,000 are gonna create
link |
00:52:32.740
a whole bunch of reference frames.
link |
00:52:34.300
What does it mean?
link |
00:52:35.540
What is a reference frame?
link |
00:52:37.300
First of all, these reference frames are different
link |
00:52:40.060
than the ones you might be used to.
link |
00:52:42.220
We know lots of reference frames.
link |
00:52:43.420
For example, we know the Cartesian coordinates, X, Y, Z,
link |
00:52:46.060
that's a type of reference frame.
link |
00:52:47.580
We know longitude and latitude,
link |
00:52:50.260
that's a different type of reference frame.
link |
00:52:52.780
If I look at a printed map,
link |
00:52:54.540
you might have columns A through M,
link |
00:52:58.460
and rows one through 20,
link |
00:53:00.060
that's a different type of reference frame.
link |
00:53:01.420
It's kind of a Cartesian coordinate reference frame.
link |
00:53:04.660
The interesting thing about the reference frames
link |
00:53:06.580
in the brain, and we know this because these
link |
00:53:08.580
have been established through neuroscience
link |
00:53:10.820
studying the entorhinal cortex.
link |
00:53:12.260
So I'm not speculating here, okay?
link |
00:53:13.580
This is known neuroscience in an old part of the brain.
link |
00:53:16.780
The way these cells create reference frames,
link |
00:53:18.860
they have no origin.
link |
00:53:20.700
So what it's more like, you have a point,
link |
00:53:24.340
a point in some space, and you,
link |
00:53:27.620
given a particular movement,
link |
00:53:29.060
you can then tell what the next point should be.
link |
00:53:32.340
And you can then tell what the next point would be,
link |
00:53:34.100
and so on.
link |
00:53:35.460
You can use this to calculate
link |
00:53:38.700
how to get from one point to another.
link |
00:53:40.340
So how do I get from my house to my home,
link |
00:53:43.180
or how do I get my finger from the side of my cup
link |
00:53:44.940
to the top of the cup?
link |
00:53:46.740
How do I get from the axioms to the conjecture?
link |
00:53:50.540
So it's a different type of reference frame,
link |
00:53:54.420
and I can, if you want, I can describe in more detail,
link |
00:53:57.380
I can paint a picture of how you might want
link |
00:53:59.060
to think about that.
link |
00:53:59.900
It's really helpful to think it's something
link |
00:54:00.980
you can move through, but is there,
link |
00:54:03.740
is it helpful to think of it as spatial in some sense,
link |
00:54:08.700
or is there something that's more?
link |
00:54:09.540
No, it's definitely spatial.
link |
00:54:11.140
It's spatial in a mathematical sense.
link |
00:54:13.820
How many dimensions?
link |
00:54:14.820
Can it be a crazy number of dimensions?
link |
00:54:16.260
Well, that's an interesting question.
link |
00:54:17.460
In the old part of the brain, the entorhinal cortex,
link |
00:54:20.260
they studied rats, and initially it looks like,
link |
00:54:22.940
oh, this is just two dimensional.
link |
00:54:24.220
It's like the rat is in some box in the maze or whatever,
link |
00:54:27.260
and they know where the rat is using
link |
00:54:28.820
these two dimensional reference frames
link |
00:54:30.300
to know where it is in the maze.
link |
00:54:32.380
We said, well, okay, but what about bats?
link |
00:54:35.540
That's a mammal, and they fly in three dimensional space.
link |
00:54:38.740
How do they do that?
link |
00:54:39.580
They seem to know where they are, right?
link |
00:54:41.700
So this is a current area of active research,
link |
00:54:44.300
and it seems like somehow the neurons
link |
00:54:46.380
in the entorhinal cortex can learn three dimensional space.
link |
00:54:50.300
We just, two members of our team,
link |
00:54:52.700
along with Elif Fett from MIT,
link |
00:54:55.940
just released a paper this literally last week.
link |
00:54:59.580
It's on bioRxiv, where they show that you can,
link |
00:55:03.620
if you, the way these things work,
link |
00:55:05.460
and I won't get, unless you want to,
link |
00:55:06.700
I won't get into the detail,
link |
00:55:08.100
but grid cells can represent any n dimensional space.
link |
00:55:12.540
It's not inherently limited.
link |
00:55:15.340
You can think of it this way.
link |
00:55:16.620
If you had two dimensional, the way it works
link |
00:55:18.620
is you had a bunch of two dimensional slices.
link |
00:55:20.780
That's the way these things work.
link |
00:55:21.940
There's a whole bunch of two dimensional models,
link |
00:55:24.260
and you can just, you can slice up
link |
00:55:26.140
any n dimensional space with two dimensional projections.
link |
00:55:29.300
So, and you could have one dimensional models.
link |
00:55:31.660
So there's nothing inherent about the mathematics
link |
00:55:34.420
about the way the neurons do this,
link |
00:55:35.780
which constrain the dimensionality of the space,
link |
00:55:39.460
which I think was important.
link |
00:55:41.460
So obviously I have a three dimensional map of this cup.
link |
00:55:44.060
Maybe it's even more than that, I don't know.
link |
00:55:46.340
But it's clearly a three dimensional map of the cup.
link |
00:55:48.340
I don't just have a projection of the cup.
link |
00:55:50.900
But when I think about birds,
link |
00:55:52.020
or when I think about mathematics,
link |
00:55:53.180
perhaps it's more than three dimensions.
link |
00:55:55.260
Who knows?
link |
00:55:56.260
So in terms of each individual column
link |
00:56:00.100
building up more and more information over time,
link |
00:56:04.020
do you think that mechanism is well understood?
link |
00:56:06.380
In your mind, you've proposed a lot of architectures there.
link |
00:56:09.860
Is that a key piece, or is it,
link |
00:56:11.820
is the big piece, the thousand brain theory of intelligence,
link |
00:56:16.220
the ensemble of it all?
link |
00:56:17.500
Well, I think they're both big.
link |
00:56:18.460
I mean, clearly the concept, as a theorist,
link |
00:56:20.940
the concept is most exciting, right?
link |
00:56:23.060
The high level concept.
link |
00:56:23.900
The high level concept.
link |
00:56:24.740
This is a totally new way of thinking
link |
00:56:26.140
about how the neocortex works.
link |
00:56:27.220
So that is appealing.
link |
00:56:28.660
It has all these ramifications.
link |
00:56:30.700
And with that, as a framework for how the brain works,
link |
00:56:33.780
you can make all kinds of predictions
link |
00:56:34.980
and solve all kinds of problems.
link |
00:56:36.220
Now we're trying to work through
link |
00:56:37.260
many of these details right now.
link |
00:56:38.460
Okay, how do the neurons actually do this?
link |
00:56:40.540
Well, it turns out, if you think about grid cells
link |
00:56:42.500
and place cells in the old parts of the brain,
link |
00:56:44.740
there's a lot that's known about them,
link |
00:56:45.980
but there's still some mysteries.
link |
00:56:47.020
There's a lot of debate about exactly the details,
link |
00:56:49.060
how these work and what are the signs.
link |
00:56:50.740
And we have that still, that same level of detail,
link |
00:56:52.860
that same level of concern.
link |
00:56:54.140
What we spend here most of our time doing
link |
00:56:56.820
is trying to make a very good list
link |
00:57:00.060
of the things we don't understand yet.
link |
00:57:02.660
That's the key part here.
link |
00:57:04.020
What are the constraints?
link |
00:57:05.260
It's not like, oh, this thing seems to work, we're done.
link |
00:57:07.020
No, it's like, okay, it kind of works,
link |
00:57:08.820
but these are other things we know it has to do
link |
00:57:10.700
and it's not doing those yet.
link |
00:57:12.860
I would say we're well on the way here.
link |
00:57:15.060
We're not done yet.
link |
00:57:17.100
There's a lot of trickiness to this system,
link |
00:57:20.020
but the basic principles about how different layers
link |
00:57:23.180
in the neocortex are doing much of this, we understand.
link |
00:57:27.340
But there's some fundamental parts
link |
00:57:28.620
that we don't understand as well.
link |
00:57:30.020
So what would you say is one of the harder open problems
link |
00:57:34.100
or one of the ones that have been bothering you,
link |
00:57:37.220
keeping you up at night the most?
link |
00:57:38.460
Oh, well, right now, this is a detailed thing
link |
00:57:40.620
that wouldn't apply to most people, okay?
link |
00:57:42.980
Sure.
link |
00:57:43.820
But you want me to answer that question?
link |
00:57:44.660
Yeah, please.
link |
00:57:46.180
We've talked about as if, oh,
link |
00:57:48.380
to predict what you're going to sense on this coffee cup,
link |
00:57:50.660
I need to know where my finger is gonna be
link |
00:57:52.300
on the coffee cup.
link |
00:57:53.580
That is true, but it's insufficient.
link |
00:57:56.340
Think about my finger touches the edge of the coffee cup.
link |
00:57:58.460
My finger can touch it at different orientations.
link |
00:58:01.660
I can rotate my finger around here and that doesn't change.
link |
00:58:06.340
I can make that prediction and somehow,
link |
00:58:08.780
so it's not just the location.
link |
00:58:10.100
There's an orientation component of this as well.
link |
00:58:13.300
This is known in the old parts of the brain too.
link |
00:58:15.140
There's things called head direction cells,
link |
00:58:16.620
which way the rat is facing.
link |
00:58:18.020
It's the same kind of basic idea.
link |
00:58:20.460
So if my finger were a rat, you know, in three dimensions,
link |
00:58:23.620
I have a three dimensional orientation
link |
00:58:25.740
and I have a three dimensional location.
link |
00:58:27.220
If I was a rat, I would have a,
link |
00:58:28.620
you might think of it as a two dimensional location,
link |
00:58:30.620
a two dimensional orientation,
link |
00:58:31.460
a one dimensional orientation,
link |
00:58:32.540
like just which way is it facing?
link |
00:58:35.100
So how the two components work together,
link |
00:58:38.260
how it is that I combine orientation,
link |
00:58:41.500
the orientation of my sensor,
link |
00:58:43.940
as well as the location is a tricky problem.
link |
00:58:49.660
And I think I've made progress on it.
link |
00:58:52.740
So at a bigger version of that,
link |
00:58:55.140
so perspective is super interesting, but super specific.
link |
00:58:58.460
Yeah, I warned you.
link |
00:59:00.060
No, no, no, that's really good,
link |
00:59:01.260
but there's a more general version of that.
link |
00:59:03.740
Do you think context matters,
link |
00:59:06.940
the fact that we're in a building in North America,
link |
00:59:10.700
that we, in the day and age where we have mugs?
link |
00:59:15.940
I mean, there's all this extra information
link |
00:59:19.180
that you bring to the table about everything else
link |
00:59:22.060
in the room that's outside of just the coffee cup.
link |
00:59:24.700
How does it get connected, do you think?
link |
00:59:27.340
Yeah, and that is another really interesting question.
link |
00:59:30.300
I'm gonna throw that under the rubric
link |
00:59:32.140
or the name of attentional problems.
link |
00:59:35.100
First of all, we have this model,
link |
00:59:36.180
I have many, many models.
link |
00:59:38.020
And also the question, does it matter?
link |
00:59:40.140
Well, it matters for certain things, of course it does.
link |
00:59:42.620
Maybe what we think of that as a coffee cup
link |
00:59:44.980
in another part of the world
link |
00:59:45.900
is viewed as something completely different.
link |
00:59:47.660
Or maybe our logo, which is very benign
link |
00:59:50.420
in this part of the world,
link |
00:59:51.340
it means something very different
link |
00:59:52.540
in another part of the world.
link |
00:59:53.780
So those things do matter.
link |
00:59:57.380
I think the way to think about it is the following,
link |
01:00:00.380
one way to think about it,
link |
01:00:01.740
is we have all these models of the world, okay?
link |
01:00:04.740
And we model everything.
link |
01:00:06.140
And as I said earlier, I kind of snuck it in there,
link |
01:00:08.860
our models are actually, we build composite structure.
link |
01:00:12.500
So every object is composed of other objects,
link |
01:00:15.260
which are composed of other objects,
link |
01:00:16.420
and they become members of other objects.
link |
01:00:18.700
So this room has chairs and a table and a room
link |
01:00:20.700
and walls and so on.
link |
01:00:21.620
Now we can just arrange these things in a certain way
link |
01:00:24.300
and go, oh, that's the nomenclature conference room.
link |
01:00:26.580
So, and what we do is when we go around the world
link |
01:00:31.260
and we experience the world,
link |
01:00:33.620
by walking into a room, for example,
link |
01:00:35.740
the first thing I do is I can say,
link |
01:00:36.780
oh, I'm in this room, do I recognize the room?
link |
01:00:38.660
Then I can say, oh, look, there's a table here.
link |
01:00:41.900
And by attending to the table,
link |
01:00:43.460
I'm then assigning this table in the context of the room.
link |
01:00:45.620
Then I can say, oh, on the table, there's a coffee cup.
link |
01:00:48.100
Oh, and on the table, there's a logo.
link |
01:00:49.740
And in the logo, there's the word Nementa.
link |
01:00:51.260
Oh, and look in the logo, there's the letter E.
link |
01:00:53.420
Oh, and look, it has an unusual serif.
link |
01:00:55.740
And it doesn't actually, but I pretended to serif.
link |
01:00:59.660
So the point is your attention is kind of drilling
link |
01:01:03.860
deep in and out of these nested structures.
link |
01:01:07.460
And I can pop back up and I can pop back down.
link |
01:01:09.340
I can pop back up and I can pop back down.
link |
01:01:10.900
So when I attend to the coffee cup,
link |
01:01:13.220
I haven't lost the context of everything else,
link |
01:01:15.660
but it's sort of, there's this sort of nested structure.
link |
01:01:18.900
So the attention filters the reference frame information
link |
01:01:22.980
for that particular period of time?
link |
01:01:24.420
Yes, it basically, moment to moment,
link |
01:01:26.620
you attend the sub components,
link |
01:01:28.420
and then you can attend the sub components
link |
01:01:29.740
to sub components.
link |
01:01:30.580
And you can move up and down.
link |
01:01:31.420
You can move up and down.
link |
01:01:32.340
We do that all the time.
link |
01:01:33.180
You're not even, now that I'm aware of it,
link |
01:01:35.580
I'm very conscious of it.
link |
01:01:36.700
But until, but most people don't even think about this.
link |
01:01:39.980
You just walk in a room and you don't say,
link |
01:01:41.700
oh, I looked at the chair and I looked at the board
link |
01:01:43.500
and looked at that word on the board
link |
01:01:44.620
and I looked over here, what's going on, right?
link |
01:01:47.100
So what percent of your day are you deeply aware of this?
link |
01:01:50.020
And what part can you actually relax and just be Jeff?
link |
01:01:52.860
Me personally, like my personal day?
link |
01:01:54.460
Yeah.
link |
01:01:55.540
Unfortunately, I'm afflicted with too much of the former.
link |
01:02:01.340
Well, unfortunately or unfortunately.
link |
01:02:02.820
Yeah.
link |
01:02:03.660
You don't think it's useful?
link |
01:02:04.580
Oh, it is useful, totally useful.
link |
01:02:06.820
I think about this stuff almost all the time.
link |
01:02:09.180
And one of my primary ways of thinking
link |
01:02:12.540
is when I'm in sleep at night,
link |
01:02:13.860
I always wake up in the middle of the night.
link |
01:02:15.860
And then I stay awake for at least an hour
link |
01:02:17.860
with my eyes shut in sort of a half sleep state
link |
01:02:20.700
thinking about these things.
link |
01:02:21.660
I come up with answers to problems very often
link |
01:02:23.700
in that sort of half sleeping state.
link |
01:02:25.660
I think about it on my bike ride, I think about it on walks.
link |
01:02:27.460
I'm just constantly thinking about this.
link |
01:02:28.780
I have to almost schedule time
link |
01:02:32.420
to not think about this stuff
link |
01:02:34.100
because it's very, it's mentally taxing.
link |
01:02:37.820
Are you, when you're thinking about this stuff,
link |
01:02:39.780
are you thinking introspectively,
link |
01:02:41.220
like almost taking a step outside of yourself
link |
01:02:43.700
and trying to figure out what is your mind doing right now?
link |
01:02:45.660
I do that all the time, but that's not all I do.
link |
01:02:49.060
I'm constantly observing myself.
link |
01:02:50.780
So as soon as I started thinking about grid cells,
link |
01:02:53.060
for example, and getting into that,
link |
01:02:55.260
I started saying, oh, well, grid cells
link |
01:02:56.780
can have my place of sense in the world.
link |
01:02:58.380
That's where you know where you are.
link |
01:02:59.660
And it's interesting, we always have a sense
link |
01:03:01.380
of where we are unless we're lost.
link |
01:03:03.020
And so I started at night when I got up
link |
01:03:04.740
to go to the bathroom, I would start trying to do it
link |
01:03:06.980
completely with my eyes closed all the time.
link |
01:03:08.500
And I would test my sense of grid cells.
link |
01:03:10.060
I would walk five feet and say, okay, I think I'm here.
link |
01:03:13.700
Am I really there?
link |
01:03:14.540
What's my error?
link |
01:03:15.460
And then I would calculate my error again
link |
01:03:16.780
and see how the errors could accumulate.
link |
01:03:17.940
So even something as simple as getting up
link |
01:03:19.460
in the middle of the night to go to the bathroom,
link |
01:03:20.420
I'm testing these theories out.
link |
01:03:22.620
It's kind of fun.
link |
01:03:23.460
I mean, the coffee cup is an example of that too.
link |
01:03:25.580
So I find that these sort of everyday introspections
link |
01:03:30.380
are actually quite helpful.
link |
01:03:32.820
It doesn't mean you can ignore the science.
link |
01:03:34.860
I mean, I spend hours every day
link |
01:03:37.060
reading ridiculously complex papers.
link |
01:03:40.180
That's not nearly as much fun,
link |
01:03:41.740
but you have to sort of build up those constraints
link |
01:03:44.580
and the knowledge about the field and who's doing what
link |
01:03:46.860
and what exactly they think is happening here.
link |
01:03:48.860
And then you can sit back and say,
link |
01:03:50.060
okay, let's try to piece this all together.
link |
01:03:53.380
Let's come up with some, I'm very,
link |
01:03:56.020
in this group here, people, they know they do,
link |
01:03:58.460
I do this all the time.
link |
01:03:59.300
I come in with these introspective ideas and say,
link |
01:04:01.220
well, have you ever thought about this?
link |
01:04:02.380
Now watch, well, let's all do this together.
link |
01:04:04.700
And it's helpful.
link |
01:04:05.940
It's not, as long as you don't,
link |
01:04:09.580
all you did was that, then you're just making up stuff.
link |
01:04:12.340
But if you're constraining it by the reality
link |
01:04:14.780
of the neuroscience, then it's really helpful.
link |
01:04:17.820
So let's talk a little bit about deep learning
link |
01:04:20.180
and the successes in the applied space of neural networks,
link |
01:04:26.860
ideas of training model on data
link |
01:04:29.020
and these simple computational units,
link |
01:04:31.420
artificial neurons that with backpropagation,
link |
01:04:36.580
statistical ways of being able to generalize
link |
01:04:40.460
from the training set onto data
link |
01:04:42.780
that's similar to that training set.
link |
01:04:44.300
So where do you think are the limitations
link |
01:04:47.420
of those approaches?
link |
01:04:48.460
What do you think are its strengths
link |
01:04:50.380
relative to your major efforts
link |
01:04:52.180
of constructing a theory of human intelligence?
link |
01:04:56.020
Well, I'm not an expert in this field.
link |
01:04:57.820
I'm somewhat knowledgeable.
link |
01:04:59.140
So, but I'm not.
link |
01:04:59.980
Some of it is in just your intuition.
link |
01:05:01.620
What are your?
link |
01:05:02.460
Well, I have a little bit more than intuition,
link |
01:05:03.860
but I just want to say like,
link |
01:05:05.420
you know, one of the things that you asked me,
link |
01:05:07.660
do I spend all my time thinking about neuroscience?
link |
01:05:09.220
I do.
link |
01:05:10.060
That's to the exclusion of thinking about things
link |
01:05:11.340
like convolutional neural networks.
link |
01:05:13.660
But I try to stay current.
link |
01:05:15.260
So look, I think it's great, the progress they've made.
link |
01:05:17.860
It's fantastic.
link |
01:05:18.780
And as I mentioned earlier,
link |
01:05:19.860
it's very highly useful for many things.
link |
01:05:22.940
The models that we have today are actually derived
link |
01:05:26.140
from a lot of neuroscience principles.
link |
01:05:28.220
There are distributed processing systems
link |
01:05:30.020
and distributed memory systems,
link |
01:05:31.260
and that's how the brain works.
link |
01:05:33.260
They use things that we might call them neurons,
link |
01:05:35.900
but they're really not neurons at all.
link |
01:05:37.020
So we can just, they're not really neurons.
link |
01:05:39.220
So they're distributed processing systems.
link |
01:05:41.220
And that nature of hierarchy,
link |
01:05:44.700
that came also from neuroscience.
link |
01:05:47.140
And so there's a lot of things,
link |
01:05:48.220
the learning rules, basically,
link |
01:05:49.780
not back prop, but other, you know,
link |
01:05:51.140
sort of heavy on top of that.
link |
01:05:52.540
I'd be curious to say they're not neurons at all.
link |
01:05:55.020
Can you describe in which way?
link |
01:05:56.180
I mean, some of it is obvious,
link |
01:05:57.700
but I'd be curious if you have specific ways
link |
01:06:00.380
in which you think are the biggest differences.
link |
01:06:02.820
Yeah, we had a paper in 2016 called
link |
01:06:04.940
Why Neurons Have Thousands of Synapses.
link |
01:06:06.940
And if you read that paper,
link |
01:06:09.460
you'll know what I'm talking about here.
link |
01:06:11.420
A real neuron in the brain is a complex thing.
link |
01:06:14.460
And let's just start with the synapses on it,
link |
01:06:17.180
which is a connection between neurons.
link |
01:06:19.020
Real neurons can have everywhere
link |
01:06:20.700
from five to 30,000 synapses on them.
link |
01:06:25.460
The ones near the cell body,
link |
01:06:27.220
the ones that are close to the soma of the cell body,
link |
01:06:30.420
those are like the ones that people model
link |
01:06:32.100
in artificial neurons.
link |
01:06:33.740
There is a few hundred of those.
link |
01:06:35.060
Maybe they can affect the cell.
link |
01:06:37.100
They can make the cell become active.
link |
01:06:39.700
95% of the synapses can't do that.
link |
01:06:43.540
They're too far away.
link |
01:06:44.580
So if you activate one of those synapses,
link |
01:06:45.980
it just doesn't affect the cell body enough
link |
01:06:47.860
to make any difference.
link |
01:06:48.860
Any one of them individually.
link |
01:06:50.100
Any one of them individually,
link |
01:06:50.940
or even if you do a mass of them.
link |
01:06:54.060
What real neurons do is the following.
link |
01:06:57.420
If you activate or you get 10 to 20 of them
link |
01:07:03.500
active at the same time,
link |
01:07:04.460
meaning they're all receiving an input at the same time,
link |
01:07:06.660
and those 10 to 20 synapses or 40 synapses
link |
01:07:09.100
within a very short distance on the dendrite,
link |
01:07:11.340
like 40 microns, a very small area.
link |
01:07:13.300
So if you activate a bunch of these
link |
01:07:14.580
right next to each other at some distant place,
link |
01:07:17.580
what happens is it creates
link |
01:07:19.300
what's called the dendritic spike.
link |
01:07:21.300
And the dendritic spike travels through the dendrites
link |
01:07:24.540
and can reach the soma or the cell body.
link |
01:07:27.820
Now, when it gets there, it changes the voltage,
link |
01:07:31.260
which is sort of like gonna make the cell fire,
link |
01:07:33.580
but never enough to make the cell fire.
link |
01:07:36.060
It's sort of what we call, it says we depolarize the cell,
link |
01:07:38.500
you raise the voltage a little bit,
link |
01:07:39.580
but not enough to do anything.
link |
01:07:41.620
It's like, well, what good is that?
link |
01:07:42.580
And then it goes back down again.
link |
01:07:44.460
So we propose a theory,
link |
01:07:47.780
which I'm very confident in basics are,
link |
01:07:50.500
is that what's happening there is
link |
01:07:52.780
those 95% of the synapses are recognizing
link |
01:07:55.860
dozens to hundreds of unique patterns.
link |
01:07:58.460
They can write about 10, 20 synapses at a time,
link |
01:08:02.060
and they're acting like predictions.
link |
01:08:04.460
So the neuron actually is a predictive engine on its own.
link |
01:08:07.620
It can fire when it gets enough,
link |
01:08:09.700
what they call proximal input
link |
01:08:10.900
from those ones near the cell fire,
link |
01:08:11.980
but it can get ready to fire from dozens to hundreds
link |
01:08:15.460
of patterns that it recognizes from the other guys.
link |
01:08:18.100
And the advantage of this to the neuron
link |
01:08:21.260
is that when it actually does produce a spike
link |
01:08:23.500
in action potential,
link |
01:08:24.780
it does so slightly sooner than it would have otherwise.
link |
01:08:27.700
And so what could is slightly sooner?
link |
01:08:29.740
Well, the slightly sooner part is it,
link |
01:08:31.820
all the excitatory neurons in the brain
link |
01:08:34.940
are surrounded by these inhibitory neurons,
link |
01:08:36.660
and they're very fast, the inhibitory neurons,
link |
01:08:38.980
these basket cells.
link |
01:08:40.420
And if I get my spike out
link |
01:08:42.580
a little bit sooner than someone else,
link |
01:08:44.220
I inhibit all my neighbors around me, right?
link |
01:08:47.020
And what you end up with is a different representation.
link |
01:08:49.740
You end up with a reputation that matches your prediction.
link |
01:08:52.060
It's a sparser representation,
link |
01:08:53.780
meaning fewer neurons are active,
link |
01:08:55.740
but it's much more specific.
link |
01:08:57.860
And so we showed how networks of these neurons
link |
01:09:00.300
can do very sophisticated temporal prediction, basically.
link |
01:09:04.180
So this, summarize this,
link |
01:09:07.020
real neurons in the brain are time based prediction engines,
link |
01:09:10.980
and there's no concept of this at all
link |
01:09:14.660
in artificial, what we call point neurons.
link |
01:09:18.100
I don't think you can build a brain without them.
link |
01:09:20.060
I don't think you can build intelligence without them,
link |
01:09:21.340
because it's where a large part of the time comes from.
link |
01:09:26.020
These are predictive models, and the time is,
link |
01:09:29.060
there's a prior and a prediction and an action,
link |
01:09:32.220
and it's inherent through every neuron in the neocortex.
link |
01:09:34.940
So I would say that point neurons sort of model
link |
01:09:37.740
a piece of that, and not very well at that either.
link |
01:09:40.620
But like for example, synapses are very unreliable,
link |
01:09:46.060
and you cannot assign any precision to them.
link |
01:09:49.900
So even one digit of precision is not possible.
link |
01:09:52.460
So the way real neurons work is they don't add these,
link |
01:09:55.540
they don't change these weights accurately
link |
01:09:57.420
like artificial neural networks do.
link |
01:09:59.340
They basically form new synapses,
link |
01:10:01.020
and so what you're trying to always do is
link |
01:10:03.780
detect the presence of some 10 to 20
link |
01:10:06.540
active synapses at the same time,
link |
01:10:08.780
as opposed, and they're almost binary.
link |
01:10:11.300
It's like, because you can't really represent
link |
01:10:12.820
anything much finer than that.
link |
01:10:14.620
So these are the kind of,
link |
01:10:16.220
and I think that's actually another essential component,
link |
01:10:18.060
because the brain works on sparse patterns,
link |
01:10:20.940
and all that mechanism is based on sparse patterns,
link |
01:10:24.180
and I don't actually think you could build real brains
link |
01:10:26.620
or machine intelligence without
link |
01:10:29.100
incorporating some of those ideas.
link |
01:10:30.900
It's hard to even think about the complexity
link |
01:10:32.660
that emerges from the fact that
link |
01:10:34.420
the timing of the firing matters in the brain,
link |
01:10:37.140
the fact that you form new synapses,
link |
01:10:40.980
and I mean, everything you just mentioned
link |
01:10:44.020
in the past couple minutes.
link |
01:10:44.940
Trust me, if you spend time on it,
link |
01:10:46.540
you can get your mind around it.
link |
01:10:47.940
It's not like, it's no longer a mystery to me.
link |
01:10:49.860
No, but sorry, as a function, in a mathematical way,
link |
01:10:53.820
can you start getting an intuition about
link |
01:10:56.940
what gets it excited, what not,
link |
01:10:58.540
and what kind of representation?
link |
01:10:59.380
Yeah, it's not as easy as,
link |
01:11:02.580
there's many other types of neural networks
link |
01:11:04.660
that are more amenable to pure analysis,
link |
01:11:09.220
especially very simple networks.
link |
01:11:10.780
Oh, I have four neurons, and they're doing this.
link |
01:11:12.580
Can we describe to them mathematically
link |
01:11:14.500
what they're doing type of thing?
link |
01:11:16.300
Even the complexity of convolutional neural networks today,
link |
01:11:19.340
it's sort of a mystery.
link |
01:11:20.300
They can't really describe the whole system.
link |
01:11:22.500
And so it's different.
link |
01:11:24.780
My colleague Subitai Ahmad, he did a nice paper on this.
link |
01:11:31.500
You can get all this stuff on our website
link |
01:11:32.740
if you're interested,
link |
01:11:34.100
talking about sort of the mathematical properties
link |
01:11:36.180
of sparse representations.
link |
01:11:37.660
And so what we can do is we can show mathematically,
link |
01:11:40.620
for example, why 10 to 20 synapses to recognize a pattern
link |
01:11:44.940
is the correct number, is the right number you'd wanna use.
link |
01:11:47.740
And by the way, that matches biology.
link |
01:11:49.980
We can show mathematically some of these concepts
link |
01:11:53.900
about the show why the brain is so robust
link |
01:11:58.620
to noise and error and fallout and so on.
link |
01:12:01.020
We can show that mathematically
link |
01:12:02.260
as well as empirically in simulations.
link |
01:12:05.020
But the system can't be analyzed completely.
link |
01:12:07.860
Any complex system can't, and so that's out of the realm.
link |
01:12:11.980
But there is mathematical benefits and intuitions
link |
01:12:17.660
that can be derived from mathematics.
link |
01:12:19.460
And we try to do that as well.
link |
01:12:20.860
Most of our papers have a section about that.
link |
01:12:23.300
So I think it's refreshing and useful for me
link |
01:12:25.900
to be talking to you about deep neural networks,
link |
01:12:29.060
because your intuition basically says
link |
01:12:30.900
that we can't achieve anything like intelligence
link |
01:12:34.540
with artificial neural networks.
link |
01:12:35.940
Well, not in the current form.
link |
01:12:36.940
Not in the current form.
link |
01:12:37.780
I'm sure we can do it in the ultimate form, sure.
link |
01:12:40.180
So let me dig into it
link |
01:12:41.260
and see what your thoughts are there a little bit.
link |
01:12:43.300
So I'm not sure if you read this little blog post
link |
01:12:45.980
called Bitter Lesson by Rich Sutton recently.
link |
01:12:49.460
He's a reinforcement learning pioneer.
link |
01:12:51.660
I'm not sure if you're familiar with him.
link |
01:12:53.260
His basic idea is that all the stuff we've done in AI
link |
01:12:56.780
in the past 70 years, he's one of the old school guys.
link |
01:13:02.980
The biggest lesson learned is that all the tricky things
link |
01:13:06.860
we've done, they benefit in the short term,
link |
01:13:10.420
but in the long term, what wins out
link |
01:13:12.100
is a simple general method that just relies on Moore's law,
link |
01:13:16.700
on computation getting faster and faster.
link |
01:13:19.820
This is what he's saying.
link |
01:13:21.260
This is what has worked up to now.
link |
01:13:23.220
This is what has worked up to now.
link |
01:13:25.380
If you're trying to build a system,
link |
01:13:29.060
if we're talking about,
link |
01:13:30.060
he's not concerned about intelligence.
link |
01:13:31.420
He's concerned about a system that works
link |
01:13:34.420
in terms of making predictions
link |
01:13:36.500
on applied narrow AI problems, right?
link |
01:13:38.780
That's what this discussion is about.
link |
01:13:40.620
That you just try to go as general as possible
link |
01:13:44.220
and wait years or decades for the computation
link |
01:13:48.500
to make it actually.
link |
01:13:50.220
Is he saying that as a criticism
link |
01:13:51.700
or is he saying this is a prescription
link |
01:13:53.260
of what we ought to be doing?
link |
01:13:54.340
Well, it's very difficult.
link |
01:13:55.860
He's saying this is what has worked
link |
01:13:57.980
and yes, a prescription, but it's a difficult prescription
link |
01:14:00.340
because it says all the fun things
link |
01:14:02.380
you guys are trying to do, we are trying to do.
link |
01:14:05.820
He's part of the community.
link |
01:14:07.340
He's saying it's only going to be short term gains.
link |
01:14:10.780
So this all leads up to a question, I guess,
link |
01:14:13.780
on artificial neural networks
link |
01:14:15.580
and maybe our own biological neural networks
link |
01:14:19.060
is do you think if we just scale things up significantly,
link |
01:14:23.780
so take these dumb artificial neurons,
link |
01:14:27.180
the point neurons, I like that term.
link |
01:14:30.420
If we just have a lot more of them,
link |
01:14:33.260
do you think some of the elements
link |
01:14:34.540
that we see in the brain may start emerging?
link |
01:14:38.060
No, I don't think so.
link |
01:14:39.540
We can do bigger problems of the same type.
link |
01:14:43.420
I mean, it's been pointed out by many people
link |
01:14:45.260
that today's convolutional neural networks
link |
01:14:46.860
aren't really much different
link |
01:14:47.860
than the ones we had quite a while ago.
link |
01:14:50.580
They're bigger and train more
link |
01:14:51.820
and we have more labeled data and so on.
link |
01:14:56.300
But I don't think you can get to the kind of things
link |
01:14:58.580
I know the brain can do and that we think about
link |
01:15:01.380
as intelligence by just scaling it up.
link |
01:15:03.700
So that may be, it's a good description
link |
01:15:06.580
of what's happened in the past,
link |
01:15:07.660
what's happened recently with the reemergence
link |
01:15:09.940
of artificial neural networks.
link |
01:15:12.500
It may be a good prescription
link |
01:15:14.380
for what's gonna happen in the short term.
link |
01:15:17.580
But I don't think that's the path.
link |
01:15:19.180
I've said that earlier.
link |
01:15:20.860
There's an alternate path.
link |
01:15:21.700
I should mention to you, by the way,
link |
01:15:22.900
that we've made sufficient progress
link |
01:15:25.900
on the whole cortical theory in the last few years
link |
01:15:28.900
that last year we decided to start actively pursuing
link |
01:15:35.660
how do we get these ideas embedded into machine learning?
link |
01:15:40.100
Well, that's, again, being led by my colleague,
link |
01:15:41.860
Subed Tariman, and he's more of a machine learning guy.
link |
01:15:45.140
I'm more of a neuroscience guy.
link |
01:15:46.740
So this is now, I wouldn't say our focus,
link |
01:15:51.180
but it is now an equal focus here
link |
01:15:54.140
because we need to proselytize what we've learned
link |
01:15:58.220
and we need to show how it's beneficial
link |
01:16:01.460
to the machine learning layer.
link |
01:16:03.740
So we're putting, we have a plan in place right now.
link |
01:16:05.580
In fact, we just did our first paper on this.
link |
01:16:07.700
I can tell you about that.
link |
01:16:09.700
But one of the reasons I wanna talk to you
link |
01:16:11.380
is because I'm trying to get more people
link |
01:16:14.100
in the machine learning community to say,
link |
01:16:15.980
I need to learn about this stuff.
link |
01:16:17.140
And maybe we should just think about this a bit more
link |
01:16:19.380
about what we've learned about the brain
link |
01:16:20.860
and what are those team at Nimenta, what have they done?
link |
01:16:23.860
Is that useful for us?
link |
01:16:25.220
Yeah, so is there elements of all the cortical theory
link |
01:16:28.500
that things we've been talking about
link |
01:16:29.820
that may be useful in the short term?
link |
01:16:31.900
Yes, in the short term, yes.
link |
01:16:33.420
This is the, sorry to interrupt,
link |
01:16:34.780
but the open question is,
link |
01:16:37.740
it certainly feels from my perspective
link |
01:16:39.260
that in the long term,
link |
01:16:41.060
some of the ideas we've been talking about
link |
01:16:42.820
will be extremely useful.
link |
01:16:44.260
The question is whether in the short term.
link |
01:16:46.020
Well, this is always what I would call
link |
01:16:48.340
the entrepreneur's dilemma.
link |
01:16:50.620
So you have this long term vision,
link |
01:16:53.060
oh, we're gonna all be driving electric cars
link |
01:16:55.300
or we're all gonna have computers
link |
01:16:56.780
or we're all gonna, whatever.
link |
01:16:59.020
And you're at some point in time and you say,
link |
01:17:01.860
I can see that long term vision,
link |
01:17:02.980
I'm sure it's gonna happen.
link |
01:17:03.820
How do I get there without killing myself?
link |
01:17:05.780
Without going out of business, right?
link |
01:17:07.380
That's the challenge.
link |
01:17:08.740
That's the dilemma.
link |
01:17:09.580
That's the really difficult thing to do.
link |
01:17:11.100
So we're facing that right now.
link |
01:17:13.100
So ideally what you'd wanna do
link |
01:17:14.660
is find some steps along the way
link |
01:17:16.100
that you can get there incrementally.
link |
01:17:17.420
You don't have to like throw it all out
link |
01:17:19.180
and start over again.
link |
01:17:20.460
The first thing that we've done
link |
01:17:22.340
is we focus on the sparse representations.
link |
01:17:25.380
So just in case you don't know what that means
link |
01:17:28.420
or some of the listeners don't know what that means,
link |
01:17:31.220
in the brain, if I have like 10,000 neurons,
link |
01:17:34.100
what you would see is maybe 2% of them active at a time.
link |
01:17:36.980
You don't see 50%, you don't see 30%,
link |
01:17:39.540
you might see 2%.
link |
01:17:41.220
And it's always like that.
link |
01:17:42.660
For any set of sensory inputs?
link |
01:17:44.380
It doesn't matter if anything,
link |
01:17:45.340
doesn't matter any part of the brain.
link |
01:17:47.380
But which neurons differs?
link |
01:17:51.100
Which neurons are active?
link |
01:17:52.620
Yeah, so let's say I take 10,000 neurons
link |
01:17:55.380
that are representing something.
link |
01:17:56.300
They're sitting there in a little block together.
link |
01:17:57.940
It's a teeny little block of neurons, 10,000 neurons.
link |
01:18:00.060
And they're representing a location,
link |
01:18:01.620
they're representing a cup,
link |
01:18:02.500
they're representing the input from my sensors.
link |
01:18:04.060
I don't know, it doesn't matter.
link |
01:18:05.380
It's representing something.
link |
01:18:07.020
The way the representations occur,
link |
01:18:09.140
it's always a sparse representation.
link |
01:18:10.620
Meaning it's a population code.
link |
01:18:11.860
So which 200 cells are active tells me what's going on.
link |
01:18:14.980
It's not, individual cells aren't that important at all.
link |
01:18:18.060
It's the population code that matters.
link |
01:18:20.260
And when you have sparse population codes,
link |
01:18:23.140
then all kinds of beautiful properties come out of them.
link |
01:18:26.300
So the brain uses sparse population codes.
link |
01:18:28.100
We've written and described these benefits
link |
01:18:30.780
in some of our papers.
link |
01:18:32.420
So they give this tremendous robustness to the systems.
link |
01:18:37.660
Brains are incredibly robust.
link |
01:18:39.180
Neurons are dying all the time and spasming
link |
01:18:41.140
and synapses are falling apart all the time.
link |
01:18:43.940
And it keeps working.
link |
01:18:45.340
So what Sibutai and Louise, one of our other engineers here
link |
01:18:51.220
have done, have shown they're introducing sparseness
link |
01:18:55.740
into convolutional neural networks.
link |
01:18:56.860
Now other people are thinking along these lines,
link |
01:18:58.140
but we're going about it in a more principled way, I think.
link |
01:19:00.980
And we're showing that if you enforce sparseness
link |
01:19:04.100
throughout these convolutional neural networks
link |
01:19:07.340
in both the act, which sort of,
link |
01:19:09.660
which neurons are active and the connections between them,
link |
01:19:12.780
that you get some very desirable properties.
link |
01:19:15.660
So one of the current hot topics in deep learning right now
link |
01:19:18.860
are these adversarial examples.
link |
01:19:20.900
So, you know, you give me any deep learning network
link |
01:19:23.500
and I can give you a picture that looks perfect
link |
01:19:26.060
and you're going to call it, you know,
link |
01:19:27.100
you're going to say the monkey is, you know, an airplane.
link |
01:19:30.300
So that's a problem.
link |
01:19:32.540
And DARPA just announced some big thing.
link |
01:19:34.140
They're trying to, you know, have some contest for this.
link |
01:19:36.580
But if you enforce sparse representations here,
link |
01:19:40.180
many of these problems go away.
link |
01:19:41.500
They're much more robust and they're not easy to fool.
link |
01:19:44.940
So we've already shown some of those results,
link |
01:19:48.340
just literally in January or February,
link |
01:19:51.140
just like last month we did that.
link |
01:19:53.740
And you can, I think it's on bioRxiv right now,
link |
01:19:57.340
or on iRxiv, you can read about it.
link |
01:19:59.540
But, so that's like a baby step, okay?
link |
01:20:03.100
That's taking something from the brain.
link |
01:20:04.340
We know about sparseness.
link |
01:20:05.620
We know why it's important.
link |
01:20:06.500
We know what it gives the brain.
link |
01:20:08.060
So let's try to enforce that onto this.
link |
01:20:09.500
What's your intuition why sparsity leads to robustness?
link |
01:20:12.420
Because it feels like it would be less robust.
link |
01:20:15.060
Why would you feel the rest robust to you?
link |
01:20:17.260
So it just feels like if the fewer neurons are involved,
link |
01:20:24.380
the more fragile the representation.
link |
01:20:26.660
But I didn't say there was lots of few neurons.
link |
01:20:28.260
I said, let's say 200.
link |
01:20:29.860
That's a lot.
link |
01:20:31.020
There's still a lot, it's just.
link |
01:20:32.620
So here's an intuition for it.
link |
01:20:35.260
This is a bit technical, so for engineers,
link |
01:20:39.860
machine learning people, this will be easy,
link |
01:20:41.260
but all the listeners, maybe not.
link |
01:20:44.300
If you're trying to classify something,
link |
01:20:45.740
you're trying to divide some very high dimensional space
link |
01:20:48.380
into different pieces, A and B.
link |
01:20:50.380
And you're trying to create some point where you say,
link |
01:20:52.820
all these points in this high dimensional space are A,
link |
01:20:54.780
and all these points in this high dimensional space are B.
link |
01:20:57.580
And if you have points that are close to that line,
link |
01:21:01.980
it's not very robust.
link |
01:21:02.900
It works for all the points you know about,
link |
01:21:04.940
but it's not very robust,
link |
01:21:07.100
because you can just move a little bit
link |
01:21:08.260
and you've crossed over the line.
link |
01:21:10.300
When you have sparse representations,
link |
01:21:12.700
imagine I pick, I'm gonna pick 200 cells active
link |
01:21:16.060
out of 10,000, okay?
link |
01:21:19.260
So I have 200 cells active.
link |
01:21:20.340
Now let's say I pick randomly another,
link |
01:21:22.220
a different representation, 200.
link |
01:21:24.420
The overlap between those is gonna be very small,
link |
01:21:26.740
just a few.
link |
01:21:28.060
I can pick millions of samples randomly of 200 neurons,
link |
01:21:32.740
and not one of them will overlap more than just a few.
link |
01:21:36.980
So one way to think about it is,
link |
01:21:39.140
if I wanna fool one of these representations
link |
01:21:41.460
to look like one of those other representations,
link |
01:21:43.460
I can't move just one cell, or two cells,
link |
01:21:45.660
or three cells, or four cells.
link |
01:21:46.780
I have to move 100 cells.
link |
01:21:49.140
And that makes them robust.
link |
01:21:52.700
In terms of further, so you mentioned sparsity.
link |
01:21:56.180
What would be the next thing?
link |
01:21:57.260
Yeah.
link |
01:21:58.100
Okay, so we have, we picked one.
link |
01:22:00.460
We don't know if it's gonna work well yet.
link |
01:22:02.380
So again, we're trying to come up with incremental ways
link |
01:22:04.540
to moving from brain theory to add pieces
link |
01:22:07.860
to machine learning, current machine learning world,
link |
01:22:10.140
and one step at a time.
link |
01:22:12.260
So the next thing we're gonna try to do
link |
01:22:13.740
is sort of incorporate some of the ideas
link |
01:22:15.820
of the thousand brains theory,
link |
01:22:19.100
that you have many, many models that are voting.
link |
01:22:22.580
Now that idea is not new.
link |
01:22:23.700
There's a mixture of models that's been around
link |
01:22:25.300
for a long time.
link |
01:22:27.160
But the way the brain does it is a little different.
link |
01:22:29.740
And the way it votes is different.
link |
01:22:33.620
And the kind of way it represents uncertainty
link |
01:22:36.220
is different.
link |
01:22:37.180
So we're just starting this work,
link |
01:22:39.980
but we're gonna try to see if we can sort of incorporate
link |
01:22:42.280
some of the principles of voting,
link |
01:22:43.760
or principles of the thousand brain theory.
link |
01:22:45.940
Like lots of simple models that talk to each other
link |
01:22:49.420
in a certain way.
link |
01:22:53.940
And can we build more machines, systems that learn faster
link |
01:22:57.700
and also, well mostly are multimodal
link |
01:23:03.220
and robust to multimodal type of issues.
link |
01:23:07.500
So one of the challenges there
link |
01:23:09.580
is the machine learning computer vision community
link |
01:23:13.100
has certain sets of benchmarks,
link |
01:23:15.600
sets of tests based on which they compete.
link |
01:23:18.180
And I would argue, especially from your perspective,
link |
01:23:22.060
that those benchmarks aren't that useful
link |
01:23:24.660
for testing the aspects that the brain is good at,
link |
01:23:28.860
or intelligence.
link |
01:23:29.940
They're not really testing intelligence.
link |
01:23:31.300
They're very fine.
link |
01:23:32.980
And it's been extremely useful
link |
01:23:34.780
for developing specific mathematical models,
link |
01:23:37.420
but it's not useful in the long term
link |
01:23:40.420
for creating intelligence.
link |
01:23:41.680
So you think you also have a role in proposing
link |
01:23:44.660
better tests?
link |
01:23:47.020
Yeah, this is a very,
link |
01:23:48.460
you've identified a very serious problem.
link |
01:23:51.440
First of all, the tests that they have
link |
01:23:53.340
are the tests that they want.
link |
01:23:54.580
Not the tests of the other things
link |
01:23:55.860
that we're trying to do, right?
link |
01:23:58.740
You know, what are the, so on.
link |
01:24:01.700
The second thing is sometimes these,
link |
01:24:04.220
to be competitive in these tests,
link |
01:24:06.620
you have to have huge data sets and huge computing power.
link |
01:24:10.820
And so, you know, and we don't have that here.
link |
01:24:13.420
We don't have it as well as other big teams
link |
01:24:15.500
that big companies do.
link |
01:24:18.700
So there's numerous issues there.
link |
01:24:20.900
You know, we come out, you know,
link |
01:24:22.420
where our approach to this is all based on,
link |
01:24:24.260
in some sense, you might argue, elegance.
link |
01:24:26.100
We're coming at it from like a theoretical base
link |
01:24:27.780
that we think, oh my God, this is so clearly elegant.
link |
01:24:29.980
This is how brains work.
link |
01:24:30.820
This is what intelligence is.
link |
01:24:31.860
But the machine learning world has gotten in this phase
link |
01:24:33.940
where they think it doesn't matter.
link |
01:24:35.500
Doesn't matter what you think,
link |
01:24:36.600
as long as you do, you know, 0.1% better on this benchmark,
link |
01:24:39.440
that's what, that's all that matters.
link |
01:24:40.780
And that's a problem.
link |
01:24:43.860
You know, we have to figure out how to get around that.
link |
01:24:46.060
That's a challenge for us.
link |
01:24:47.300
That's one of the challenges that we have to deal with.
link |
01:24:50.500
So I agree, you've identified a big issue.
link |
01:24:52.820
It's difficult for those reasons.
link |
01:24:55.900
But you know, part of the reasons I'm talking to you here
link |
01:24:59.580
today is I hope I'm gonna get some machine learning people
link |
01:25:01.620
to say, I'm gonna read those papers.
link |
01:25:03.260
Those might be some interesting ideas.
link |
01:25:04.500
I'm tired of doing this 0.1% improvement stuff, you know?
link |
01:25:08.460
Well, that's why I'm here as well,
link |
01:25:10.340
because I think machine learning now as a community
link |
01:25:13.020
is at a place where the next step needs to be orthogonal
link |
01:25:18.500
to what has received success in the past.
link |
01:25:21.300
Well, you see other leaders saying this,
link |
01:25:23.100
machine learning leaders, you know,
link |
01:25:25.500
Jeff Hinton with his capsules idea.
link |
01:25:27.940
Many people have gotten up to say, you know,
link |
01:25:29.300
we're gonna hit road map, maybe we should look at the brain,
link |
01:25:32.100
you know, things like that.
link |
01:25:33.460
So hopefully that thinking will occur organically.
link |
01:25:38.100
And then we're in a nice position for people to come
link |
01:25:40.740
and look at our work and say,
link |
01:25:41.740
well, what can we learn from these guys?
link |
01:25:43.180
Yeah, MIT is launching a billion dollar computing college
link |
01:25:47.500
that's centered around this idea, so.
link |
01:25:49.220
Is it on this idea of what?
link |
01:25:50.980
Well, the idea that, you know,
link |
01:25:52.700
the humanities, psychology, and neuroscience
link |
01:25:54.980
have to work all together to get to build the S.
link |
01:25:58.860
Yeah, I mean, Stanford just did
link |
01:26:00.340
this Human Centered AI Center.
link |
01:26:02.500
I'm a little disappointed in these initiatives
link |
01:26:04.420
because, you know, they're focusing
link |
01:26:08.340
on sort of the human side of it,
link |
01:26:09.940
and it could very easily slip into
link |
01:26:12.140
how humans interact with intelligent machines,
link |
01:26:16.060
which is nothing wrong with that,
link |
01:26:17.620
but that's not, that is orthogonal
link |
01:26:19.420
to what we're trying to do.
link |
01:26:20.380
We're trying to say, like,
link |
01:26:21.340
what is the essence of intelligence?
link |
01:26:22.860
I don't care.
link |
01:26:23.700
In fact, I wanna build intelligent machines
link |
01:26:25.500
that aren't emotional, that don't smile at you,
link |
01:26:28.620
that, you know, that aren't trying to tuck you in at night.
link |
01:26:31.820
Yeah, there is that pattern that you,
link |
01:26:34.020
when you talk about understanding humans
link |
01:26:36.500
is important for understanding intelligence,
link |
01:26:38.380
that you start slipping into topics of ethics
link |
01:26:41.140
or, yeah, like you said,
link |
01:26:43.700
the interactive elements as opposed to,
link |
01:26:45.700
no, no, no, we have to zoom in on the brain,
link |
01:26:47.380
study what the human brain, the baby, the...
link |
01:26:51.460
Let's study what a brain does.
link |
01:26:52.900
Does.
link |
01:26:53.740
And then we can decide which parts of that
link |
01:26:54.780
we wanna recreate in some system,
link |
01:26:57.740
but until you have that theory about what the brain does,
link |
01:26:59.900
what's the point, you know, it's just,
link |
01:27:01.300
you're gonna be wasting time, I think.
link |
01:27:02.740
Right, just to break it down
link |
01:27:04.060
on the artificial neural network side,
link |
01:27:05.620
maybe you could speak to this
link |
01:27:06.740
on the biological neural network side,
link |
01:27:09.180
the process of learning versus the process of inference.
link |
01:27:13.300
Maybe you can explain to me,
link |
01:27:15.620
is there a difference between,
link |
01:27:18.460
you know, in artificial neural networks,
link |
01:27:19.860
there's a difference between the learning stage
link |
01:27:21.500
and the inference stage.
link |
01:27:22.940
Do you see the brain as something different?
link |
01:27:24.980
One of the big distinctions that people often say,
link |
01:27:29.020
I don't know how correct it is,
link |
01:27:30.660
is artificial neural networks need a lot of data.
link |
01:27:32.940
They're very inefficient learning.
link |
01:27:34.820
Do you see that as a correct distinction
link |
01:27:37.340
from the biology of the human brain,
link |
01:27:40.300
that the human brain is very efficient,
link |
01:27:41.980
or is that just something we deceive ourselves?
link |
01:27:44.220
No, it is efficient, obviously.
link |
01:27:45.420
We can learn new things almost instantly.
link |
01:27:47.580
And so what elements do you think are useful?
link |
01:27:50.020
Yeah, I can talk about that.
link |
01:27:50.860
You brought up two issues there.
link |
01:27:52.300
So remember I talked early about the constraints
link |
01:27:54.820
we always feel, well, one of those constraints
link |
01:27:57.260
is the fact that brains are continually learning.
link |
01:28:00.940
That's not something we said, oh, we can add that later.
link |
01:28:03.780
That's something that was upfront,
link |
01:28:05.780
had to be there from the start,
link |
01:28:08.900
made our problems harder.
link |
01:28:11.260
But we showed, going back to the 2016 paper
link |
01:28:14.420
on sequence memory, we showed how that happens,
link |
01:28:16.780
how the brains infer and learn at the same time.
link |
01:28:19.940
And our models do that.
link |
01:28:21.740
And they're not two separate phases,
link |
01:28:24.060
or two separate sets of time.
link |
01:28:26.340
I think that's a big, big problem in AI,
link |
01:28:29.780
at least for many applications, not for all.
link |
01:28:33.420
So I can talk about that.
link |
01:28:34.380
There are some, it gets detailed,
link |
01:28:37.180
there are some parts of the neocortex in the brain
link |
01:28:39.660
where actually what's going on,
link |
01:28:41.740
there's these cycles of activity in the brain.
link |
01:28:46.860
And there's very strong evidence
link |
01:28:49.260
that you're doing more of inference
link |
01:28:51.260
on one part of the phase,
link |
01:28:52.300
and more of learning on the other part of the phase.
link |
01:28:54.100
So the brain can actually sort of separate
link |
01:28:55.500
different populations of cells
link |
01:28:56.660
or going back and forth like this.
link |
01:28:58.340
But in general, I would say that's an important problem.
link |
01:29:01.540
We have all of our networks that we've come up with do both.
link |
01:29:05.620
And they're continuous learning networks.
link |
01:29:08.220
And you mentioned benchmarks earlier.
link |
01:29:10.980
Well, there are no benchmarks about that.
link |
01:29:12.500
So we have to, we get in our little soapbox,
link |
01:29:17.180
and hey, by the way, this is important,
link |
01:29:19.220
and here's a mechanism for doing that.
link |
01:29:20.580
But until you can prove it to someone
link |
01:29:23.900
in some commercial system or something, it's a little harder.
link |
01:29:26.700
So yeah, one of the things I had to linger on that
link |
01:29:28.980
is in some ways to learn the concept of a coffee cup,
link |
01:29:33.780
you only need this one coffee cup
link |
01:29:35.900
and maybe some time alone in a room with it.
link |
01:29:37.980
Well, the first thing is,
link |
01:29:39.940
imagine I reach my hand into a black box
link |
01:29:41.820
and I'm reaching, I'm trying to touch something.
link |
01:29:43.700
I don't know upfront if it's something I already know
link |
01:29:46.220
or if it's a new thing.
link |
01:29:47.860
And I have to, I'm doing both at the same time.
link |
01:29:50.460
I don't say, oh, let's see if it's a new thing.
link |
01:29:53.260
Oh, let's see if it's an old thing.
link |
01:29:54.740
I don't do that.
link |
01:29:55.580
As I go, my brain says, oh, it's new or it's not new.
link |
01:29:59.420
And if it's new, I start learning what it is.
link |
01:30:02.300
And by the way, it starts learning from the get go,
link |
01:30:04.820
even if it's gonna recognize it.
link |
01:30:06.020
So they're not separate problems.
link |
01:30:08.900
And so that's the thing there.
link |
01:30:10.060
The other thing you mentioned was the fast learning.
link |
01:30:13.540
So I was just talking about continuous learning,
link |
01:30:15.580
but there's also fast learning.
link |
01:30:16.660
Literally, I can show you this coffee cup
link |
01:30:18.780
and I say, here's a new coffee cup.
link |
01:30:20.060
It's got the logo on it.
link |
01:30:21.340
Take a look at it, done, you're done.
link |
01:30:23.860
You can predict what it's gonna look like,
link |
01:30:25.380
you know, in different positions.
link |
01:30:27.460
So I can talk about that too.
link |
01:30:29.540
In the brain, the way learning occurs,
link |
01:30:34.220
I mentioned this earlier, but I'll mention it again.
link |
01:30:35.700
The way learning occurs,
link |
01:30:36.820
imagine I am a section of a dendrite of a neuron,
link |
01:30:40.140
and I'm gonna learn something new.
link |
01:30:43.740
Doesn't matter what it is.
link |
01:30:44.580
I'm just gonna learn something new.
link |
01:30:46.180
I need to recognize a new pattern.
link |
01:30:48.900
So what I'm gonna do is I'm gonna form new synapses.
link |
01:30:52.540
New synapses, we're gonna rewire the brain
link |
01:30:55.140
onto that section of the dendrite.
link |
01:30:57.900
Once I've done that, everything else that neuron has learned
link |
01:31:01.020
is not affected by it.
link |
01:31:02.580
That's because it's isolated
link |
01:31:04.340
to that small section of the dendrite.
link |
01:31:06.380
They're not all being added together, like a point neuron.
link |
01:31:09.580
So if I learn something new on this segment here,
link |
01:31:11.740
it doesn't change any of the learning
link |
01:31:13.180
that occur anywhere else in that neuron.
link |
01:31:14.860
So I can add something without affecting previous learning.
link |
01:31:18.420
And I can do it quickly.
link |
01:31:20.940
Now let's talk, we can talk about the quickness,
link |
01:31:22.300
how it's done in real neurons.
link |
01:31:24.020
You might say, well, doesn't it take time to form synapses?
link |
01:31:26.740
Yes, it can take maybe an hour to form a new synapse.
link |
01:31:30.900
We can form memories quicker than that,
link |
01:31:32.500
and I can explain that how it happens too, if you want.
link |
01:31:35.860
But it's getting a bit neurosciencey.
link |
01:31:39.460
That's great, but is there an understanding
link |
01:31:41.380
of these mechanisms at every level?
link |
01:31:43.100
Yeah.
link |
01:31:43.940
So from the short term memories and the forming.
link |
01:31:48.620
So this idea of synaptogenesis, the growth of new synapses,
link |
01:31:51.580
that's well described, it's well understood.
link |
01:31:54.100
And that's an essential part of learning.
link |
01:31:55.820
That is learning.
link |
01:31:56.780
That is learning.
link |
01:31:58.180
Okay.
link |
01:32:01.980
Going back many, many years,
link |
01:32:03.860
people, you know, it was, what's his name,
link |
01:32:06.340
the psychologist who proposed, Hebb, Donald Hebb.
link |
01:32:09.580
He proposed that learning was the modification
link |
01:32:12.020
of the strength of a connection between two neurons.
link |
01:32:15.460
People interpreted that as the modification
link |
01:32:18.180
of the strength of a synapse.
link |
01:32:19.660
He didn't say that.
link |
01:32:20.980
He just said there's a modification
link |
01:32:22.340
between the effect of one neuron and another.
link |
01:32:24.540
So synaptogenesis is totally consistent
link |
01:32:26.500
with what Donald Hebb said.
link |
01:32:28.180
But anyway, there's these mechanisms,
link |
01:32:29.860
the growth of new synapses.
link |
01:32:30.860
You can go online, you can watch a video
link |
01:32:32.260
of a synapse growing in real time.
link |
01:32:33.900
It's literally, you can see this little thing going boop.
link |
01:32:37.140
It's pretty impressive.
link |
01:32:38.420
So those mechanisms are known.
link |
01:32:39.740
Now there's another thing that we've speculated
link |
01:32:42.340
and we've written about,
link |
01:32:43.540
which is consistent with known neuroscience,
link |
01:32:45.780
but it's less proven.
link |
01:32:48.340
And this is the idea, how do I form a memory
link |
01:32:50.580
really, really quickly?
link |
01:32:51.620
Like instantaneous.
link |
01:32:52.820
If it takes an hour to grow a synapse,
link |
01:32:54.580
like that's not instantaneous.
link |
01:32:56.820
So there are types of synapses called silent synapses.
link |
01:33:01.700
They look like a synapse, but they don't do anything.
link |
01:33:04.060
They're just sitting there.
link |
01:33:04.900
It's like if an action potential comes in,
link |
01:33:07.900
it doesn't release any neurotransmitter.
link |
01:33:10.140
Some parts of the brain have more of these than others.
link |
01:33:12.500
For example, the hippocampus has a lot of them,
link |
01:33:14.020
which is where we associate most short term memory with.
link |
01:33:18.540
So what we speculated, again, in that 2016 paper,
link |
01:33:22.100
we proposed that the way we form very quick memories,
link |
01:33:26.420
very short term memories, or quick memories,
link |
01:33:28.940
is that we convert silent synapses into active synapses.
link |
01:33:33.860
It's like saying a synapse has a zero weight
link |
01:33:36.060
and a one weight,
link |
01:33:37.860
but the longterm memory has to be formed by synaptogenesis.
link |
01:33:41.460
So you can remember something really quickly
link |
01:33:43.300
by just flipping a bunch of these guys from silent to active.
link |
01:33:46.220
It's not from 0.1 to 0.15.
link |
01:33:49.140
It's like, it doesn't do anything
link |
01:33:50.700
till it releases transmitter.
link |
01:33:52.260
And if I do that over a bunch of these,
link |
01:33:53.500
I've got a very quick short term memory.
link |
01:33:56.860
So I guess the lesson behind this
link |
01:33:58.500
is that most neural networks today are fully connected.
link |
01:34:01.860
Every neuron connects every other neuron
link |
01:34:03.380
from layer to layer.
link |
01:34:04.580
That's not correct in the brain.
link |
01:34:06.060
We don't want that.
link |
01:34:06.980
We actually don't want that.
link |
01:34:08.340
It's bad.
link |
01:34:09.260
You want a very sparse connectivity
link |
01:34:10.700
so that any neuron connects to some subset of the neurons
link |
01:34:14.500
in the other layer.
link |
01:34:15.340
And it does so on a dendrite by dendrite segment basis.
link |
01:34:18.980
So it's a very some parcelated out type of thing.
link |
01:34:21.580
And that then learning is not adjusting all these weights,
link |
01:34:25.380
but learning is just saying,
link |
01:34:26.340
okay, connect to these 10 cells here right now.
link |
01:34:30.180
In that process, you know, with artificial neural networks,
link |
01:34:32.980
it's a very simple process of backpropagation
link |
01:34:36.060
that adjusts the weights.
link |
01:34:37.180
The process of synaptogenesis.
link |
01:34:40.100
Synaptogenesis.
link |
01:34:40.940
Synaptogenesis.
link |
01:34:42.300
It's even easier.
link |
01:34:43.140
It's even easier.
link |
01:34:43.980
It's even easier.
link |
01:34:44.820
Backpropagation requires something
link |
01:34:47.260
that really can't happen in brains.
link |
01:34:48.700
This backpropagation of this error signal,
link |
01:34:51.220
that really can't happen.
link |
01:34:52.060
People are trying to make it happen in brains,
link |
01:34:53.500
but it doesn't happen in brains.
link |
01:34:54.740
This is pure Hebbian learning.
link |
01:34:56.780
Well, synaptogenesis is pure Hebbian learning.
link |
01:34:58.660
It's basically saying,
link |
01:35:00.140
there's a population of cells over here
link |
01:35:01.540
that are active right now.
link |
01:35:03.020
And there's a population of cells over here
link |
01:35:04.340
active right now.
link |
01:35:05.380
How do I form connections between those active cells?
link |
01:35:07.980
And it's literally saying this guy became active.
link |
01:35:11.100
These 100 neurons here became active
link |
01:35:13.260
before this neuron became active.
link |
01:35:15.080
So form connections to those ones.
link |
01:35:17.140
That's it.
link |
01:35:17.960
There's no propagation of error, nothing.
link |
01:35:19.940
All the networks we do,
link |
01:35:20.980
all the models we have work on almost completely on
link |
01:35:25.700
Hebbian learning,
link |
01:35:26.540
but on dendritic segments
link |
01:35:30.260
and multiple synapses at the same time.
link |
01:35:33.060
So now let's sort of turn the question
link |
01:35:34.540
that you already answered,
link |
01:35:35.820
and maybe you can answer it again.
link |
01:35:38.820
If you look at the history of artificial intelligence,
link |
01:35:41.260
where do you think we stand?
link |
01:35:43.540
How far are we from solving intelligence?
link |
01:35:45.780
You said you were very optimistic.
link |
01:35:47.700
Can you elaborate on that?
link |
01:35:48.900
Yeah, it's always the crazy question to ask
link |
01:35:53.500
because no one can predict the future.
link |
01:35:55.100
Absolutely.
link |
01:35:55.940
So I'll tell you a story.
link |
01:35:58.180
I used to run a different neuroscience institute
link |
01:36:01.400
called the Redwood Neuroscience Institute,
link |
01:36:02.620
and we would hold these symposiums
link |
01:36:04.740
and we'd get like 35 scientists
link |
01:36:06.340
from around the world to come together.
link |
01:36:08.060
And I used to ask them all the same question.
link |
01:36:10.380
I would say, well, how long do you think it'll be
link |
01:36:11.740
before we understand how the neocortex works?
link |
01:36:14.540
And everyone went around the room
link |
01:36:15.540
and they had introduced the name
link |
01:36:16.560
and they have to answer that question.
link |
01:36:18.240
So I got, the typical answer was 50 to 100 years.
link |
01:36:22.940
Some people would say 500 years.
link |
01:36:24.780
Some people said never.
link |
01:36:25.860
I said, why are you a neuroscientist?
link |
01:36:27.820
It's never gonna, it's a good pay.
link |
01:36:32.780
It's interesting.
link |
01:36:34.380
So, you know, but it doesn't work like that.
link |
01:36:36.300
As I mentioned earlier, these are not,
link |
01:36:38.740
these are step functions.
link |
01:36:39.620
Things happen and then bingo, they happen.
link |
01:36:41.780
You can't predict that.
link |
01:36:43.620
I feel I've already passed a step function.
link |
01:36:45.620
So if I can do my job correctly over the next five years,
link |
01:36:50.740
then, meaning I can proselytize these ideas.
link |
01:36:53.540
I can convince other people they're right.
link |
01:36:56.140
We can show that other people,
link |
01:36:58.740
machine learning people should pay attention
link |
01:37:00.260
to these ideas.
link |
01:37:01.420
Then we're definitely in an under 20 year timeframe.
link |
01:37:04.580
If I can do those things, if I'm not successful in that,
link |
01:37:07.780
and this is the last time anyone talks to me
link |
01:37:09.780
and no one reads our papers and you know,
link |
01:37:12.180
and I'm wrong or something like that,
link |
01:37:13.980
then I don't know.
link |
01:37:15.940
But it's not 50 years.
link |
01:37:21.820
Think about electric cars.
link |
01:37:22.940
How quickly are they gonna populate the world?
link |
01:37:24.940
It probably takes about a 20 year span.
link |
01:37:27.900
It'll be something like that.
link |
01:37:28.820
But I think if I can do what I said, we're starting it.
link |
01:37:31.740
And of course there could be other,
link |
01:37:34.220
you said step functions.
link |
01:37:35.400
It could be everybody gives up on your ideas for 20 years
link |
01:37:40.100
and then all of a sudden somebody picks it up again.
link |
01:37:42.180
Wait, that guy was onto something.
link |
01:37:43.620
Yeah, so that would be a failure on my part, right?
link |
01:37:47.540
Think about Charles Babbage.
link |
01:37:49.820
Charles Babbage, he's the guy who invented the computer
link |
01:37:52.220
back in the 18 something, 1800s.
link |
01:37:55.820
And everyone forgot about it until 100 years later.
link |
01:37:59.460
And say, hey, this guy figured this stuff out
link |
01:38:00.900
a long time ago.
link |
01:38:02.380
But he was ahead of his time.
link |
01:38:03.940
I don't think, as I said,
link |
01:38:06.460
I recognize this is part of any entrepreneur's challenge.
link |
01:38:09.780
I use entrepreneur broadly in this case.
link |
01:38:11.500
I'm not meaning like I'm building a business
link |
01:38:12.980
or trying to sell something.
link |
01:38:13.820
I mean, I'm trying to sell ideas.
link |
01:38:15.900
And this is the challenge as to how you get people
link |
01:38:19.380
to pay attention to you, how do you get them
link |
01:38:21.540
to give you positive or negative feedback,
link |
01:38:24.700
how do you get the people to act differently
link |
01:38:25.960
based on your ideas.
link |
01:38:27.220
So we'll see how well we do on that.
link |
01:38:30.180
So you know that there's a lot of hype
link |
01:38:32.300
behind artificial intelligence currently.
link |
01:38:34.640
Do you, as you look to spread the ideas
link |
01:38:39.540
that are of neocortical theory, the things you're working on,
link |
01:38:43.300
do you think there's some possibility
link |
01:38:45.100
we'll hit an AI winter once again?
link |
01:38:47.300
Yeah, it's certainly a possibility.
link |
01:38:48.940
No question about it.
link |
01:38:49.780
Is that something you worry about?
link |
01:38:50.600
Yeah, well, I guess, do I worry about it?
link |
01:38:54.340
I haven't decided yet if that's good or bad for my mission.
link |
01:38:57.540
That's true, that's very true.
link |
01:38:59.660
Because it's almost like you need the winter
link |
01:39:02.940
to refresh the palette.
link |
01:39:04.300
Yeah, it's like, I want, here's what you wanna have it is.
link |
01:39:07.860
You want, like to the extent that everyone is so thrilled
link |
01:39:12.180
about the current state of machine learning and AI
link |
01:39:15.460
and they don't imagine they need anything else,
link |
01:39:18.100
it makes my job harder.
link |
01:39:19.740
If everything crashed completely
link |
01:39:22.580
and every student left the field
link |
01:39:24.260
and there was no money for anybody to do anything
link |
01:39:26.200
and it became an embarrassment
link |
01:39:27.460
to talk about machine intelligence and AI,
link |
01:39:29.020
that wouldn't be good for us either.
link |
01:39:30.740
You want sort of the soft landing approach, right?
link |
01:39:33.400
You want enough people, the senior people in AI
link |
01:39:36.620
and machine learning to say, you know,
link |
01:39:37.860
we need other approaches.
link |
01:39:38.900
We really need other approaches.
link |
01:39:40.460
Damn, we need other approaches.
link |
01:39:42.020
Maybe we should look to the brain.
link |
01:39:43.100
Okay, let's look to the brain.
link |
01:39:44.220
Who's got some brain ideas?
link |
01:39:45.380
Okay, let's start a little project on the side here
link |
01:39:47.900
trying to do brain idea related stuff.
link |
01:39:49.700
That's the ideal outcome we would want.
link |
01:39:51.820
So I don't want a total winter
link |
01:39:53.980
and yet I don't want it to be sunny all the time either.
link |
01:39:57.680
So what do you think it takes to build a system
link |
01:40:00.300
with human level intelligence
link |
01:40:03.020
where once demonstrated you would be very impressed?
link |
01:40:06.820
So does it have to have a body?
link |
01:40:08.700
Does it have to have the C word we used before,
link |
01:40:12.780
consciousness as an entirety in a holistic sense?
link |
01:40:19.140
First of all, I don't think the goal
link |
01:40:20.500
is to create a machine that is human level intelligence.
link |
01:40:23.740
I think it's a false goal.
link |
01:40:24.980
Back to Turing, I think it was a false statement.
link |
01:40:27.380
We want to understand what intelligence is
link |
01:40:29.060
and then we can build intelligent machines
link |
01:40:30.780
of all different scales, all different capabilities.
link |
01:40:34.260
A dog is intelligent.
link |
01:40:35.300
I don't need, that'd be pretty good to have a dog.
link |
01:40:38.460
But what about something that doesn't look
link |
01:40:39.580
like an animal at all, in different spaces?
link |
01:40:41.660
So my thinking about this is that
link |
01:40:44.300
we want to define what intelligence is,
link |
01:40:46.060
agree upon what makes an intelligent system.
link |
01:40:48.840
We can then say, okay, we're now gonna build systems
link |
01:40:51.100
that work on those principles or some subset of them
link |
01:40:54.340
and we can apply them to all different types of problems.
link |
01:40:57.380
And the kind, the idea, it's not computing.
link |
01:41:00.860
We don't ask, if I take a little one chip computer,
link |
01:41:05.340
I don't say, well, that's not a computer
link |
01:41:06.660
because it's not as powerful as this big server over here.
link |
01:41:09.660
No, no, because we know that what the principles
link |
01:41:11.260
of computing are and I can apply those principles
link |
01:41:12.940
to a small problem or into a big problem.
link |
01:41:14.860
And same, intelligence needs to get there.
link |
01:41:16.520
We have to say, these are the principles.
link |
01:41:17.620
I can make a small one, a big one.
link |
01:41:19.020
I can make them distributed.
link |
01:41:19.940
I can put them on different sensors.
link |
01:41:21.620
They don't have to be human like at all.
link |
01:41:23.220
Now, you did bring up a very interesting question
link |
01:41:24.740
about embodiment.
link |
01:41:25.620
Does it have to have a body?
link |
01:41:27.500
It has to have some concept of movement.
link |
01:41:30.660
It has to be able to move through these reference frames
link |
01:41:33.260
I talked about earlier.
link |
01:41:34.460
Whether it's physically moving,
link |
01:41:35.820
like I need, if I'm gonna have an AI
link |
01:41:37.420
that understands coffee cups,
link |
01:41:38.780
it's gonna have to pick up the coffee cup
link |
01:41:40.500
and touch it and look at it with its eyes and hands
link |
01:41:43.180
or something equivalent to that.
link |
01:41:45.380
If I have a mathematical AI,
link |
01:41:48.100
maybe it needs to move through mathematical spaces.
link |
01:41:51.340
I could have a virtual AI that lives in the internet
link |
01:41:55.240
and its movements are traversing links
link |
01:41:58.980
and digging into files,
link |
01:42:00.260
but it's got a location that it's traveling
link |
01:42:03.100
through some space.
link |
01:42:04.940
You can't have an AI that just take some flash thing input.
link |
01:42:09.060
We call it flash inference.
link |
01:42:10.620
Here's a pattern, done.
link |
01:42:12.860
No, it's movement pattern, movement pattern,
link |
01:42:15.740
movement pattern, attention, digging, building structure,
link |
01:42:19.020
figuring out the model of the world.
link |
01:42:20.420
So some sort of embodiment,
link |
01:42:22.740
whether it's physical or not, has to be part of it.
link |
01:42:25.780
So self awareness and the way to be able to answer
link |
01:42:28.020
where am I?
link |
01:42:28.860
Well, you're bringing up self,
link |
01:42:29.680
that's a different topic, self awareness.
link |
01:42:31.460
No, the very narrow definition of self,
link |
01:42:33.700
meaning knowing a sense of self enough to know
link |
01:42:37.740
where am I in the space where it's actually.
link |
01:42:39.980
Yeah, basically the system needs to know its location
link |
01:42:43.500
or each component of the system needs to know
link |
01:42:46.020
where it is in the world at that point in time.
link |
01:42:48.620
So self awareness and consciousness.
link |
01:42:51.660
Do you think one, from the perspective of neuroscience
link |
01:42:55.620
and neurocortex, these are interesting topics,
link |
01:42:58.180
solvable topics.
link |
01:42:59.780
Do you have any ideas of why the heck it is
link |
01:43:02.180
that we have a subjective experience at all?
link |
01:43:04.420
Yeah, I have a lot of thoughts on that.
link |
01:43:05.260
And is it useful or is it just a side effect of us?
link |
01:43:08.460
It's interesting to think about.
link |
01:43:10.140
I don't think it's useful as a means to figure out
link |
01:43:13.460
how to build intelligent machines.
link |
01:43:16.360
It's something that systems do
link |
01:43:20.180
and we can talk about what it is that are like,
link |
01:43:22.780
well, if I build a system like this,
link |
01:43:23.980
then it would be self aware.
link |
01:43:25.300
Or if I build it like this, it wouldn't be self aware.
link |
01:43:28.340
So that's a choice I can have.
link |
01:43:30.040
It's not like, oh my God, it's self aware.
link |
01:43:32.300
I can't turn, I heard an interview recently
link |
01:43:35.800
with this philosopher from Yale,
link |
01:43:37.120
I can't remember his name, I apologize for that.
link |
01:43:39.020
But he was talking about,
link |
01:43:39.860
well, if these computers are self aware,
link |
01:43:41.420
then it would be a crime to unplug them.
link |
01:43:42.900
And I'm like, oh, come on, that's not,
link |
01:43:45.060
I unplug myself every night, I go to sleep.
link |
01:43:47.260
Is that a crime?
link |
01:43:48.260
I plug myself in again in the morning and there I am.
link |
01:43:51.340
So people get kind of bent out of shape about this.
link |
01:43:56.020
I have very definite, very detailed understanding
link |
01:43:59.500
or opinions about what it means to be conscious
link |
01:44:02.260
and what it means to be self aware.
link |
01:44:04.580
I don't think it's that interesting a problem.
link |
01:44:06.780
You've talked to Christoph Koch.
link |
01:44:08.740
He thinks that's the only problem.
link |
01:44:10.900
I didn't actually listen to your interview with him,
link |
01:44:12.380
but I know him and I know that's the thing he cares about.
link |
01:44:15.820
He also thinks intelligence and consciousness are disjoint.
link |
01:44:18.260
So I mean, it's not, you don't have to have one or the other.
link |
01:44:21.020
So he is.
link |
01:44:21.860
I disagree with that.
link |
01:44:22.740
I just totally disagree with that.
link |
01:44:24.600
So where's your thoughts and consciousness,
link |
01:44:26.300
where does it emerge from?
link |
01:44:27.660
Because it is.
link |
01:44:28.500
So then we have to break it down to the two parts, okay?
link |
01:44:30.860
Because consciousness isn't one thing.
link |
01:44:32.140
That's part of the problem with that term
link |
01:44:33.660
is it means different things to different people
link |
01:44:35.460
and there's different components of it.
link |
01:44:37.600
There is a concept of self awareness, okay?
link |
01:44:40.820
That can be very easily explained.
link |
01:44:43.100
You have a model of your own body.
link |
01:44:46.060
The neocortex models things in the world
link |
01:44:48.140
and it also models your own body.
link |
01:44:50.500
And then it has a memory.
link |
01:44:53.340
It can remember what you've done, okay?
link |
01:44:55.860
So it can remember what you did this morning,
link |
01:44:57.540
can remember what you had for breakfast and so on.
link |
01:44:59.640
And so I can say to you, okay, Lex,
link |
01:45:03.080
were you conscious this morning when you had your bagel?
link |
01:45:06.900
And you'd say, yes, I was conscious.
link |
01:45:08.820
Now what if I could take your brain
link |
01:45:10.300
and revert all the synapses back
link |
01:45:12.020
to the state they were this morning?
link |
01:45:14.180
And then I said to you, Lex,
link |
01:45:15.900
were you conscious when you ate the bagel?
link |
01:45:17.220
And you said, no, I wasn't conscious.
link |
01:45:18.540
I said, here's a video of eating the bagel.
link |
01:45:19.740
And you said, I wasn't there.
link |
01:45:22.420
That's not possible
link |
01:45:23.380
because I must've been unconscious at that time.
link |
01:45:25.660
So we can just make this one to one correlation
link |
01:45:27.460
between memory of your body's trajectory through the world
link |
01:45:31.000
over some period of time,
link |
01:45:32.100
a memory and the ability to recall that memory
link |
01:45:34.260
is what you would call conscious.
link |
01:45:35.900
I was conscious of that, it's a self awareness.
link |
01:45:38.940
And any system that can recall,
link |
01:45:41.340
memorize what it's done recently
link |
01:45:43.540
and bring that back and invoke it again
link |
01:45:46.340
would say, yeah, I'm aware.
link |
01:45:48.220
I remember what I did.
link |
01:45:49.380
All right, I got it.
link |
01:45:51.340
That's an easy one.
link |
01:45:52.420
Although some people think that's a hard one.
link |
01:45:54.780
The more challenging part of consciousness
link |
01:45:57.380
is this one that's sometimes used
link |
01:45:59.060
going by the word of qualia,
link |
01:46:01.300
which is, why does an object seem red?
link |
01:46:04.860
Or what is pain?
link |
01:46:06.860
And why does pain feel like something?
link |
01:46:08.740
Why do I feel redness?
link |
01:46:10.380
Or why do I feel painness?
link |
01:46:12.660
And then I could say, well,
link |
01:46:13.500
why does sight seems different than hearing?
link |
01:46:15.620
It's the same problem.
link |
01:46:16.460
It's really, these are all just neurons.
link |
01:46:18.580
And so how is it that,
link |
01:46:20.300
why does looking at you feel different than hearing you?
link |
01:46:24.140
It feels different, but there's just neurons in my head.
link |
01:46:26.080
They're all doing the same thing.
link |
01:46:27.820
So that's an interesting question.
link |
01:46:29.820
The best treatise I've read about this
link |
01:46:31.540
is by a guy named Oregon.
link |
01:46:33.580
He wrote a book called,
link |
01:46:35.740
Why Red Doesn't Sound Like a Bell.
link |
01:46:37.480
It's a little, it's not a trade book, easy to read,
link |
01:46:42.040
but it, and it's an interesting question.
link |
01:46:46.040
Take something like color.
link |
01:46:47.880
Color really doesn't exist in the world.
link |
01:46:49.360
It's not a property of the world.
link |
01:46:51.160
Property of the world that exists is light frequency.
link |
01:46:54.240
And that gets turned into,
link |
01:46:55.640
we have certain cells in the retina
link |
01:46:57.960
that respond to different frequencies
link |
01:46:59.320
different than others.
link |
01:47:00.240
And so when they enter the brain,
link |
01:47:01.440
you just have a bunch of axons
link |
01:47:02.440
that are firing at different rates.
link |
01:47:04.500
And from that, we perceive color.
link |
01:47:06.680
But there is no color in the brain.
link |
01:47:07.960
I mean, there's no color coming in on those synapses.
link |
01:47:10.840
It's just a correlation between some axons
link |
01:47:14.380
and some property of frequency.
link |
01:47:17.360
And that isn't even color itself.
link |
01:47:18.880
Frequency doesn't have a color.
link |
01:47:20.140
It's just what it is.
link |
01:47:22.940
So then the question is,
link |
01:47:24.120
well, why does it even appear to have a color at all?
link |
01:47:27.880
Just as you're describing it,
link |
01:47:29.080
there seems to be a connection to those ideas
link |
01:47:31.000
of reference frames.
link |
01:47:32.560
I mean, it just feels like consciousness
link |
01:47:37.040
having the subject,
link |
01:47:38.400
assigning the feeling of red to the actual color
link |
01:47:42.600
or to the wavelength is useful for intelligence.
link |
01:47:47.920
Yeah, I think that's a good way of putting it.
link |
01:47:49.600
It's useful as a predictive mechanism
link |
01:47:51.600
or useful as a generalization idea.
link |
01:47:53.840
It's a way of grouping things together to say,
link |
01:47:55.660
it's useful to have a model like this.
link |
01:47:57.560
So think about the well known syndrome
link |
01:48:02.640
that people who've lost a limb experience
link |
01:48:04.800
called phantom limbs.
link |
01:48:06.960
And what they claim is they can have their arm is removed,
link |
01:48:12.120
but they feel their arm.
link |
01:48:13.280
That not only feel it, they know it's there.
link |
01:48:15.960
It's there, I know it's there.
link |
01:48:17.740
They'll swear to you that it's there.
link |
01:48:19.000
And then they can feel pain in their arm
link |
01:48:20.360
and they'll feel pain in their finger.
link |
01:48:21.840
And if they move their non existent arm behind their back,
link |
01:48:25.280
then they feel the pain behind their back.
link |
01:48:27.320
So this whole idea that your arm exists
link |
01:48:30.120
is a model of your brain.
link |
01:48:31.360
It may or may not really exist.
link |
01:48:34.360
And just like, but it's useful to have a model of something
link |
01:48:38.520
that sort of correlates to things in the world.
link |
01:48:40.360
So you can make predictions about what would happen
link |
01:48:41.960
when those things occur.
link |
01:48:43.520
It's a little bit of a fuzzy,
link |
01:48:44.640
but I think you're getting quite towards the answer there.
link |
01:48:46.480
It's useful for the model to express things certain ways
link |
01:48:51.280
that we can then map them into these reference frames
link |
01:48:53.640
and make predictions about them.
link |
01:48:55.780
I need to spend more time on this topic.
link |
01:48:57.680
It doesn't bother me.
link |
01:48:58.880
Do you really need to spend more time?
link |
01:49:00.360
Yeah, I know.
link |
01:49:01.840
It does feel special that we have subjective experience,
link |
01:49:04.720
but I'm yet to know why.
link |
01:49:07.320
I'm just personally curious.
link |
01:49:09.040
It's not necessary for the work we're doing here.
link |
01:49:11.400
I don't think I need to solve that problem
link |
01:49:13.080
to build intelligent machines at all, not at all.
link |
01:49:15.560
But there is sort of the silly notion
link |
01:49:17.800
that you described briefly
link |
01:49:20.440
that doesn't seem so silly to us humans is,
link |
01:49:23.280
if you're successful building intelligent machines,
link |
01:49:27.080
it feels wrong to then turn them off.
link |
01:49:30.240
Because if you're able to build a lot of them,
link |
01:49:33.240
it feels wrong to then be able to turn off the...
link |
01:49:38.760
Well, why?
link |
01:49:39.600
Let's break that down a bit.
link |
01:49:41.840
As humans, why do we fear death?
link |
01:49:43.920
There's two reasons we fear death.
link |
01:49:47.060
Well, first of all, I'll say,
link |
01:49:47.900
when you're dead, it doesn't matter at all.
link |
01:49:48.960
Who cares?
link |
01:49:49.800
You're dead.
link |
01:49:50.640
So why do we fear death?
link |
01:49:51.840
We fear death for two reasons.
link |
01:49:53.480
One is because we are programmed genetically to fear death.
link |
01:49:57.760
That's a survival and pop beginning of the genes thing.
link |
01:50:02.940
And we also are programmed to feel sad
link |
01:50:05.120
when people we know die.
link |
01:50:06.880
We don't feel sad for someone we don't know dies.
link |
01:50:08.560
There's people dying right now,
link |
01:50:09.600
they're only just gonna say,
link |
01:50:10.420
I don't feel bad about them,
link |
01:50:11.260
because I don't know them.
link |
01:50:12.080
But if I knew them, I'd feel really bad.
link |
01:50:13.420
So again, these are old brain,
link |
01:50:16.840
genetically embedded things that we fear death.
link |
01:50:19.880
It's outside of those uncomfortable feelings.
link |
01:50:24.280
There's nothing else to worry about.
link |
01:50:25.840
Well, wait, hold on a second.
link |
01:50:27.360
Do you know the denial of death by Becker?
link |
01:50:30.440
No.
link |
01:50:31.360
There's a thought that death is,
link |
01:50:36.760
our whole conception of our world model
link |
01:50:41.280
kind of assumes immortality.
link |
01:50:43.800
And then death is this terror that underlies it all.
link |
01:50:47.040
So like...
link |
01:50:47.880
Some people's world model, not mine.
link |
01:50:50.400
But, okay, so what Becker would say
link |
01:50:52.760
is that you're just living in an illusion.
link |
01:50:54.520
You've constructed an illusion for yourself
link |
01:50:56.200
because it's such a terrible terror,
link |
01:50:59.000
the fact that this...
link |
01:51:00.160
What's the illusion?
link |
01:51:01.160
The illusion that death doesn't matter.
link |
01:51:02.640
You're still not coming to grips with...
link |
01:51:04.800
The illusion of what?
link |
01:51:05.620
That death is...
link |
01:51:07.120
Going to happen.
link |
01:51:08.700
Oh, like it's not gonna happen?
link |
01:51:10.440
You're actually operating.
link |
01:51:11.880
You haven't, even though you said you've accepted it,
link |
01:51:14.280
you haven't really accepted the notion that you're gonna die
link |
01:51:16.120
is what you say.
link |
01:51:16.960
So it sounds like you disagree with that notion.
link |
01:51:21.440
Yeah, yeah, totally.
link |
01:51:22.400
I literally, every night I go to bed, it's like dying.
link |
01:51:28.040
Like little deaths.
link |
01:51:28.880
It's little deaths.
link |
01:51:29.720
And if I didn't wake up, it wouldn't matter to me.
link |
01:51:32.960
Only if I knew that was gonna happen would it be bothersome.
link |
01:51:35.160
If I didn't know it was gonna happen, how would I know?
link |
01:51:37.600
Then I would worry about my wife.
link |
01:51:39.520
So imagine I was a loner and I lived in Alaska
link |
01:51:43.040
and I lived out there and there was no animals.
link |
01:51:45.420
Nobody knew I existed.
link |
01:51:46.480
I was just eating these roots all the time.
link |
01:51:48.720
And nobody knew I was there.
link |
01:51:51.120
And one day I didn't wake up.
link |
01:51:54.680
What pain in the world would there exist?
link |
01:51:57.040
Well, so most people that think about this problem
link |
01:51:59.800
would say that you're just deeply enlightened
link |
01:52:01.960
or are completely delusional.
link |
01:52:04.120
One of the two.
link |
01:52:05.920
But I would say that's a very enlightened way
link |
01:52:10.720
to see the world.
link |
01:52:13.120
That's the rational one as well.
link |
01:52:14.760
It's rational, that's right.
link |
01:52:15.760
But the fact is we don't,
link |
01:52:19.040
I mean, we really don't have an understanding
link |
01:52:22.360
of why the heck it is we're born and why we die
link |
01:52:24.920
and what happens after we die.
link |
01:52:25.960
Well, maybe there isn't a reason, maybe there is.
link |
01:52:27.880
So I'm interested in those big problems too, right?
link |
01:52:30.120
You interviewed Max Tegmark,
link |
01:52:32.560
and there's people like that, right?
link |
01:52:33.600
I'm interested in those big problems as well.
link |
01:52:35.240
And in fact, when I was young,
link |
01:52:38.240
I made a list of the biggest problems I could think of.
link |
01:52:41.200
First, why does anything exist?
link |
01:52:43.360
Second, why do we have the laws of physics that we have?
link |
01:52:46.600
Third, is life inevitable?
link |
01:52:49.200
And why is it here?
link |
01:52:50.120
Fourth, is intelligence inevitable?
link |
01:52:52.240
And why is it here?
link |
01:52:53.080
I stopped there because I figured
link |
01:52:55.000
if you can make a truly intelligent system,
link |
01:52:57.840
that will be the quickest way
link |
01:52:59.240
to answer the first three questions.
link |
01:53:01.040
I'm serious.
link |
01:53:04.440
And so I said, my mission, you asked me earlier,
link |
01:53:07.960
my first mission is to understand the brain,
link |
01:53:09.440
but I felt that is the shortest way
link |
01:53:10.760
to get to true machine intelligence.
link |
01:53:12.160
And I wanna get to true machine intelligence
link |
01:53:13.680
because even if it doesn't occur in my lifetime,
link |
01:53:15.920
other people will benefit from it
link |
01:53:17.480
because I think it'll occur in my lifetime,
link |
01:53:19.200
but 20 years, you never know.
link |
01:53:23.640
But that will be the quickest way for us to,
link |
01:53:26.080
we can make super mathematicians,
link |
01:53:27.800
we can make super space explorers,
link |
01:53:29.520
we can make super physicist brains that do these things
link |
01:53:34.240
and that can run experiments that we can't run.
link |
01:53:37.440
We don't have the abilities to manipulate things and so on,
link |
01:53:40.360
but we can build intelligent machines that do all those things
link |
01:53:42.800
with the ultimate goal of finding out the answers
link |
01:53:46.560
to the other questions.
link |
01:53:48.800
Let me ask you another depressing and difficult question,
link |
01:53:51.480
which is once we achieve that goal of creating,
link |
01:53:57.880
no, of understanding intelligence,
link |
01:54:01.200
do you think we would be happier,
link |
01:54:02.960
more fulfilled as a species?
link |
01:54:04.760
The understanding intelligence
link |
01:54:05.720
or understanding the answers to the big questions?
link |
01:54:07.920
Understanding intelligence.
link |
01:54:08.960
Oh, totally, totally.
link |
01:54:11.800
It would be far more fun place to live.
link |
01:54:13.960
You think so?
link |
01:54:14.800
Oh yeah, why not?
link |
01:54:15.680
I mean, just put aside this terminator nonsense
link |
01:54:19.720
and just think about, you can think about,
link |
01:54:24.320
we can talk about the risks of AI if you want.
link |
01:54:26.840
I'd love to, so let's talk about.
link |
01:54:28.240
But I think the world would be far better knowing things.
link |
01:54:30.640
We're always better than know things.
link |
01:54:32.080
Do you think it's better, is it a better place to live in
link |
01:54:35.080
that I know that our planet is one of many
link |
01:54:37.440
in the solar system and the solar system's one of many
link |
01:54:39.240
in the galaxy?
link |
01:54:40.080
I think it's a more, I dread, I sometimes think like,
link |
01:54:43.400
God, what would it be like to live 300 years ago?
link |
01:54:45.360
I'd be looking up at the sky, I can't understand anything.
link |
01:54:47.440
Oh my God, I'd be like going to bed every night going,
link |
01:54:49.240
what's going on here?
link |
01:54:50.160
Well, I mean, in some sense I agree with you,
link |
01:54:52.480
but I'm not exactly sure.
link |
01:54:54.800
So I'm also a scientist, so I share your views,
link |
01:54:57.240
but I'm not, we're like rolling down the hill together.
link |
01:55:02.640
What's down the hill?
link |
01:55:03.480
I feel like we're climbing a hill.
link |
01:55:05.280
Whatever.
link |
01:55:06.120
We're getting closer to enlightenment
link |
01:55:07.640
and you're going down the hill.
link |
01:55:10.200
We're climbing, we're getting pulled up a hill
link |
01:55:12.240
by our curiosity.
link |
01:55:13.840
Our curiosity is, we're pulling ourselves up the hill
link |
01:55:16.120
by our curiosity.
link |
01:55:16.960
Yeah, Sisyphus was doing the same thing with the rock.
link |
01:55:19.160
Yeah, yeah, yeah, yeah.
link |
01:55:20.880
But okay, our happiness aside, do you have concerns
link |
01:55:24.280
about, you talk about Sam Harris, Elon Musk,
link |
01:55:29.040
of existential threats of intelligent systems?
link |
01:55:31.880
No, I'm not worried about existential threats at all.
link |
01:55:33.800
There are some things we really do need to worry about.
link |
01:55:36.400
Even today's AI, we have things we have to worry about.
link |
01:55:38.440
We have to worry about privacy
link |
01:55:39.560
and about how it impacts false beliefs in the world.
link |
01:55:42.800
And we have real problems and things to worry about
link |
01:55:47.000
with today's AI.
link |
01:55:48.280
And that will continue as we create more intelligent systems.
link |
01:55:51.480
There's no question, the whole issue
link |
01:55:53.080
about making intelligent armaments and weapons
link |
01:55:57.080
is something that really we have to think about carefully.
link |
01:55:59.920
I don't think of those as existential threats.
link |
01:56:01.880
I think those are the kind of threats we always face
link |
01:56:04.320
and we'll have to face them here
link |
01:56:05.840
and we'll have to deal with them.
link |
01:56:10.400
We could talk about what people think
link |
01:56:12.040
are the existential threats,
link |
01:56:13.880
but when I hear people talking about them,
link |
01:56:16.200
they all sound hollow to me.
link |
01:56:17.760
They're based on ideas, they're based on people
link |
01:56:20.000
who really have no idea what intelligence is.
link |
01:56:22.160
And if they knew what intelligence was,
link |
01:56:24.920
they wouldn't say those things.
link |
01:56:26.640
So those are not experts in the field.
link |
01:56:28.600
Yeah, so there's two, right?
link |
01:56:32.040
So one is like super intelligence.
link |
01:56:33.720
So a system that becomes far, far superior
link |
01:56:37.720
in reasoning ability than us humans.
link |
01:56:43.160
How is that an existential threat?
link |
01:56:46.200
Then, so there's a lot of ways in which it could be.
link |
01:56:49.120
One way is us humans are actually irrational, inefficient
link |
01:56:54.040
and get in the way of, not happiness,
link |
01:57:00.520
but whatever the objective function is
link |
01:57:02.120
of maximizing that objective function.
link |
01:57:04.320
Super intelligent.
link |
01:57:05.160
The paperclip problem and things like that.
link |
01:57:06.720
So the paperclip problem but with the super intelligent.
link |
01:57:09.440
Yeah, yeah, yeah, yeah.
link |
01:57:10.480
So we already face this threat in some sense.
link |
01:57:15.680
They're called bacteria.
link |
01:57:17.320
These are organisms in the world
link |
01:57:18.960
that would like to turn everything into bacteria.
link |
01:57:21.400
And they're constantly morphing,
link |
01:57:23.040
they're constantly changing to evade our protections.
link |
01:57:26.360
And in the past, they have killed huge swaths
link |
01:57:30.680
of populations of humans on this planet.
link |
01:57:33.400
So if you wanna worry about something
link |
01:57:34.560
that's gonna multiply endlessly, we have it.
link |
01:57:38.360
And I'm far more worried in that regard.
link |
01:57:40.600
I'm far more worried that some scientists in the laboratory
link |
01:57:43.280
will create a super virus or a super bacteria
link |
01:57:45.440
that we cannot control.
link |
01:57:47.120
That is a more of an existential threat.
link |
01:57:49.640
Putting an intelligence thing on top of it
link |
01:57:52.160
actually seems to make it less existential to me.
link |
01:57:54.240
It's like, it limits its power.
link |
01:57:56.640
It limits where it can go.
link |
01:57:57.720
It limits the number of things it can do in many ways.
link |
01:57:59.820
A bacteria is something you can't even see.
link |
01:58:03.080
So that's only one of those problems.
link |
01:58:04.240
Yes, exactly.
link |
01:58:05.080
So the other one, just in your intuition about intelligence,
link |
01:58:09.600
when you think about intelligence of us humans,
link |
01:58:12.480
do you think of that as something,
link |
01:58:14.880
if you look at intelligence on a spectrum
link |
01:58:16.960
from zero to us humans,
link |
01:58:18.900
do you think you can scale that to something far,
link |
01:58:22.080
far superior to all the mechanisms we've been talking about?
link |
01:58:24.760
I wanna make another point here, Lex, before I get there.
link |
01:58:28.360
Intelligence is the neocortex.
link |
01:58:30.920
It is not the entire brain.
link |
01:58:34.080
The goal is not to make a human.
link |
01:58:36.200
The goal is not to make an emotional system.
link |
01:58:38.400
The goal is not to make a system
link |
01:58:39.560
that wants to have sex and reproduce.
link |
01:58:41.440
Why would I build that?
link |
01:58:42.880
If I wanna have a system that wants to reproduce
link |
01:58:44.560
and have sex, make bacteria, make computer viruses.
link |
01:58:47.260
Those are bad things, don't do that.
link |
01:58:49.800
Those are really bad, don't do those things.
link |
01:58:52.360
Regulate those.
link |
01:58:53.560
But if I just say I want an intelligent system,
link |
01:58:56.120
why does it have to have any of the human like emotions?
link |
01:58:58.560
Why does it even care if it lives?
link |
01:59:00.400
Why does it even care if it has food?
link |
01:59:02.560
It doesn't care about those things.
link |
01:59:03.840
It's just, you know, it's just in a trance
link |
01:59:06.320
thinking about mathematics or it's out there
link |
01:59:08.280
just trying to build the space for it on Mars.
link |
01:59:14.000
That's a choice we make.
link |
01:59:15.320
Don't make human like things,
link |
01:59:17.160
don't make replicating things,
link |
01:59:18.480
don't make things that have emotions,
link |
01:59:19.840
just stick to the neocortex.
link |
01:59:21.000
So that's a view actually that I share
link |
01:59:23.120
but not everybody shares in the sense that
link |
01:59:25.400
you have faith and optimism about us as engineers of systems,
link |
01:59:29.840
humans as builders of systems to not put in stupid, not.
link |
01:59:34.840
So this is why I mentioned the bacteria one.
link |
01:59:37.640
Because you might say, well, some person's gonna do that.
link |
01:59:40.760
Well, some person today could create a bacteria
link |
01:59:42.920
that's resistant to all the known antibacterial agents.
link |
01:59:46.920
So we already have that threat.
link |
01:59:49.200
We already know this is going on.
link |
01:59:51.320
It's not a new threat.
link |
01:59:52.760
So just accept that and then we have to deal with it, right?
link |
01:59:56.280
Yeah, so my point is nothing to do with intelligence.
link |
01:59:59.600
Intelligence is a separate component
link |
02:00:01.920
that you might apply to a system
link |
02:00:03.560
that wants to reproduce and do stupid things.
link |
02:00:06.040
Let's not do that.
link |
02:00:07.240
Yeah, in fact, it is a mystery
link |
02:00:08.400
why people haven't done that yet.
link |
02:00:10.520
My dad is a physicist, believes that the reason,
link |
02:00:14.320
he says, for example, nuclear weapons haven't proliferated
link |
02:00:18.080
amongst evil people.
link |
02:00:19.040
So one belief that I share is that
link |
02:00:21.680
there's not that many evil people in the world
link |
02:00:25.160
that would use, whether it's bacteria or nuclear weapons
link |
02:00:31.280
or maybe the future AI systems to do bad.
link |
02:00:35.000
So the fraction is small.
link |
02:00:36.200
And the second is that it's actually really hard,
link |
02:00:38.400
technically, so the intersection between evil
link |
02:00:42.480
and competent is small in terms of, and that's the.
link |
02:00:45.160
And by the way, to really annihilate humanity,
link |
02:00:47.000
you'd have to have sort of the nuclear winter phenomenon,
link |
02:00:50.800
which is not one person shooting or even 10 bombs.
link |
02:00:54.080
You'd have to have some automated system
link |
02:00:56.440
that detonates a million bombs
link |
02:00:58.520
or whatever many thousands we have.
link |
02:01:00.400
So extreme evil combined with extreme competence.
link |
02:01:03.080
And to start with building some stupid system
link |
02:01:05.080
that would automatically, Dr. Strangelove type of thing,
link |
02:01:08.000
you know, I mean, look, we could have
link |
02:01:11.920
some nuclear bomb go off in some major city in the world.
link |
02:01:14.600
I think that's actually quite likely, even in my lifetime.
link |
02:01:17.120
I don't think that's an unlikely thing.
link |
02:01:18.480
And it'd be a tragedy.
link |
02:01:20.600
But it won't be an existential threat.
link |
02:01:23.160
And it's the same as, you know, the virus of 1917,
link |
02:01:26.560
whatever it was, you know, the influenza.
link |
02:01:30.000
These bad things can happen and the plague and so on.
link |
02:01:33.880
We can't always prevent them.
link |
02:01:35.360
We always try, but we can't.
link |
02:01:37.040
But they're not existential threats
link |
02:01:38.240
until we combine all those crazy things together.
link |
02:01:41.200
So on the spectrum of intelligence from zero to human,
link |
02:01:45.440
do you have a sense of whether it's possible
link |
02:01:47.960
to create several orders of magnitude
link |
02:01:51.560
or at least double that of human intelligence?
link |
02:01:54.680
Talking about neuro context.
link |
02:01:55.920
I think it's the wrong thing to say double the intelligence.
link |
02:01:59.000
Break it down into different components.
link |
02:02:01.600
Can I make something that's a million times fast
link |
02:02:03.640
than a human brain?
link |
02:02:04.520
Yes, I can do that.
link |
02:02:06.280
Could I make something that is,
link |
02:02:09.160
has a lot more storage than the human brain?
link |
02:02:10.960
Yes, I could do that.
link |
02:02:11.880
More common, more copies of common.
link |
02:02:13.600
Can I make something that attaches
link |
02:02:14.720
to different sensors than human brain?
link |
02:02:16.160
Yes, I can do that.
link |
02:02:17.280
Could I make something that's distributed?
link |
02:02:19.280
So these people, yeah, we talked early
link |
02:02:21.160
about the departure of the neocortex voting.
link |
02:02:23.120
They don't have to be co located.
link |
02:02:24.240
Like, you know, they can be all around the place.
link |
02:02:25.680
I could do that too.
link |
02:02:29.080
Those are the levers I have, but is it more intelligent?
link |
02:02:32.440
Well, it depends what I train it on.
link |
02:02:33.760
What is it doing?
link |
02:02:34.800
If it's.
link |
02:02:35.640
Well, so here's the thing.
link |
02:02:36.720
So let's say larger neocortex
link |
02:02:39.400
and or whatever size that allows for higher
link |
02:02:44.720
and higher hierarchies to form,
link |
02:02:47.920
we're talking about reference frames and concepts.
link |
02:02:50.160
Could I have something that's a super physicist
link |
02:02:51.920
or a super mathematician?
link |
02:02:52.920
Yes.
link |
02:02:53.760
And the question is, once you have a super physicist,
link |
02:02:56.680
will they be able to understand something?
link |
02:03:00.440
Do you have a sense that it will be orders of math,
link |
02:03:02.200
like us compared to ants?
link |
02:03:03.040
Could we ever understand it?
link |
02:03:04.560
Yeah.
link |
02:03:06.080
Most people cannot understand general relativity.
link |
02:03:11.920
It's a really hard thing to get.
link |
02:03:13.280
I mean, yeah, you can paint it in a fuzzy picture,
link |
02:03:15.800
stretchy space, you know?
link |
02:03:17.560
But the field equations to do that
link |
02:03:19.920
and the deep intuitions are really, really hard.
link |
02:03:23.080
And I've tried, I'm unable to do it.
link |
02:03:25.960
Like it's easy to get special relativity,
link |
02:03:28.800
but general relativity, man, that's too much.
link |
02:03:32.360
And so we already live with this to some extent.
link |
02:03:34.960
The vast majority of people can't understand actually
link |
02:03:37.320
what the vast majority of other people actually know.
link |
02:03:40.280
We're just, either we don't have the effort to,
link |
02:03:41.960
or we can't, or we don't have time,
link |
02:03:43.280
or just not smart enough, whatever.
link |
02:03:46.920
But we have ways of communicating.
link |
02:03:48.560
Einstein has spoken in a way that I can understand.
link |
02:03:51.600
He's given me analogies that are useful.
link |
02:03:54.600
I can use those analogies from my own work
link |
02:03:56.880
and think about concepts that are similar.
link |
02:04:01.040
It's not stupid.
link |
02:04:02.200
It's not like he's existing some other plane
link |
02:04:04.040
and there's no connection with my plane in the world here.
link |
02:04:06.680
So that will occur.
link |
02:04:07.840
It already has occurred.
link |
02:04:09.280
That's what my point of this story is.
link |
02:04:10.760
It already has occurred.
link |
02:04:11.720
We live it every day.
link |
02:04:14.360
One could argue that when we create machine intelligence
link |
02:04:17.040
that think a million times faster than us
link |
02:04:18.720
that it'll be so far we can't make the connections.
link |
02:04:20.920
But you know, at the moment,
link |
02:04:23.480
everything that seems really, really hard
link |
02:04:25.640
to figure out in the world,
link |
02:04:26.680
when you actually figure it out, it's not that hard.
link |
02:04:29.000
You know, almost everyone can understand the multiverses.
link |
02:04:32.160
Almost everyone can understand quantum physics.
link |
02:04:34.040
Almost everyone can understand these basic things,
link |
02:04:36.120
even though hardly any people could figure those things out.
link |
02:04:39.000
Yeah, but really understand.
link |
02:04:41.320
But you don't need to really.
link |
02:04:42.360
Only a few people really understand.
link |
02:04:43.800
You need to only understand the projections,
link |
02:04:47.880
the sprinkles of the useful insights from that.
link |
02:04:50.120
That was my example of Einstein, right?
link |
02:04:51.760
His general theory of relativity is one thing
link |
02:04:53.800
that very, very, very few people can get.
link |
02:04:56.000
And what if we just said those other few people
link |
02:04:58.240
are also artificial intelligences?
link |
02:05:00.600
How bad is that?
link |
02:05:01.440
In some sense they are, right?
link |
02:05:02.720
Yeah, they say already.
link |
02:05:04.280
I mean, Einstein wasn't a really normal person.
link |
02:05:06.280
He had a lot of weird quirks.
link |
02:05:07.560
And so did the other people who worked with him.
link |
02:05:09.440
So, you know, maybe they already were sort of
link |
02:05:11.360
this astral plane of intelligence that,
link |
02:05:14.200
we live with it already.
link |
02:05:15.240
It's not a problem.
link |
02:05:17.000
It's still useful and, you know.
link |
02:05:20.160
So do you think we are the only intelligent life
link |
02:05:22.960
out there in the universe?
link |
02:05:24.880
I would say that intelligent life
link |
02:05:27.880
has and will exist elsewhere in the universe.
link |
02:05:29.760
I'll say that.
link |
02:05:31.480
There was a question about
link |
02:05:32.600
contemporaneous intelligence life,
link |
02:05:34.080
which is hard to even answer
link |
02:05:35.600
when we think about relativity and the nature of space time.
link |
02:05:39.480
Can't say what exactly is this time
link |
02:05:41.120
someplace else in the world.
link |
02:05:43.160
But I think it's, you know,
link |
02:05:44.600
I do worry a lot about the filter idea,
link |
02:05:48.440
which is that perhaps intelligent species
link |
02:05:52.240
don't last very long.
link |
02:05:54.040
And so we haven't been around very long.
link |
02:05:55.720
And as a technological species,
link |
02:05:57.200
we've been around for almost nothing, you know.
link |
02:05:59.800
What, 200 years, something like that.
link |
02:06:02.720
And we don't have any data,
link |
02:06:04.160
a good data point on whether it's likely
link |
02:06:06.040
that we'll survive or not.
link |
02:06:08.480
So do I think that there have been intelligent life
link |
02:06:10.960
elsewhere in the universe?
link |
02:06:11.800
Almost certainly, of course.
link |
02:06:13.400
In the past, in the future, yes.
link |
02:06:16.440
Does it survive for a long time?
link |
02:06:17.880
I don't know.
link |
02:06:18.840
This is another reason I'm excited about our work,
link |
02:06:21.120
is our work meaning the general world of AI.
link |
02:06:24.240
I think we can build intelligent machines
link |
02:06:28.640
that outlast us.
link |
02:06:32.040
You know, they don't have to be tied to Earth.
link |
02:06:34.080
They don't have to, you know,
link |
02:06:35.800
I'm not saying they're recreating, you know,
link |
02:06:38.200
aliens, I'm just saying,
link |
02:06:40.680
if I asked myself,
link |
02:06:41.920
and this might be a good point to end on here.
link |
02:06:44.280
If I asked myself, you know,
link |
02:06:45.120
what's special about our species?
link |
02:06:47.240
We're not particularly interesting physically.
link |
02:06:49.040
We don't fly, we're not good swimmers,
link |
02:06:51.480
we're not very fast, we're not very strong, you know.
link |
02:06:54.000
It's our brain, that's the only thing.
link |
02:06:55.480
And we are the only species on this planet
link |
02:06:57.440
that's built the model of the world
link |
02:06:58.760
that extends beyond what we can actually sense.
link |
02:07:01.160
We're the only people who know about
link |
02:07:03.000
the far side of the moon, and the other universes,
link |
02:07:05.160
and I mean, other galaxies, and other stars,
link |
02:07:07.280
and about what happens in the atom.
link |
02:07:09.520
There's no, that knowledge doesn't exist anywhere else.
link |
02:07:12.440
It's only in our heads.
link |
02:07:13.800
Cats don't do it, dogs don't do it,
link |
02:07:15.000
monkeys don't do it, it's just on.
link |
02:07:16.360
And that is what we've created that's unique.
link |
02:07:18.320
Not our genes, it's knowledge.
link |
02:07:20.360
And if I asked me, what is the legacy of humanity?
link |
02:07:23.160
What should our legacy be?
link |
02:07:25.120
It should be knowledge.
link |
02:07:25.960
We should preserve our knowledge
link |
02:07:27.560
in a way that it can exist beyond us.
link |
02:07:30.080
And I think the best way of doing that,
link |
02:07:32.040
in fact you have to do it,
link |
02:07:33.080
is it has to go along with intelligent machines
link |
02:07:34.880
that understand that knowledge.
link |
02:07:37.800
It's a very broad idea, but we should be thinking,
link |
02:07:41.920
I call it a state planning for humanity.
link |
02:07:43.800
We should be thinking about what we wanna leave behind
link |
02:07:46.560
when as a species we're no longer here.
link |
02:07:49.320
And that'll happen sometime.
link |
02:07:51.080
Sooner or later it's gonna happen.
link |
02:07:52.480
And understanding intelligence and creating intelligence
link |
02:07:56.080
gives us a better chance to prolong.
link |
02:07:58.400
It does give us a better chance to prolong life, yes.
link |
02:08:01.120
It gives us a chance to live on other planets.
link |
02:08:03.200
But even beyond that, I mean our solar system
link |
02:08:06.080
will disappear one day, just given enough time.
link |
02:08:08.680
So I don't know, I doubt we'll ever be able to travel
link |
02:08:12.880
to other things, but we could tell the stars,
link |
02:08:15.480
but we could send intelligent machines to do that.
link |
02:08:17.800
So you have an optimistic, a hopeful view of our knowledge
link |
02:08:23.160
of the echoes of human civilization
link |
02:08:26.040
living through the intelligent systems we create?
link |
02:08:29.240
Oh, totally.
link |
02:08:30.080
Well, I think the intelligent systems we create
link |
02:08:31.400
are in some sense the vessel for bringing them beyond Earth
link |
02:08:36.200
or making them last beyond humans themselves.
link |
02:08:39.840
How do you feel about that?
link |
02:08:41.280
That they won't be human, quote unquote?
link |
02:08:43.640
Who cares?
link |
02:08:45.080
Human, what is human?
link |
02:08:46.120
Our species are changing all the time.
link |
02:08:48.640
Human today is not the same as human just 50 years ago.
link |
02:08:52.600
What is human?
link |
02:08:53.440
Do we care about our genetics?
link |
02:08:54.520
Why is that important?
link |
02:08:56.160
As I point out, our genetics are no more interesting
link |
02:08:58.320
than a bacterium's genetics.
link |
02:08:59.440
It's no more interesting than a monkey's genetics.
link |
02:09:01.720
What we have, what's unique and what's valuable
link |
02:09:04.560
is our knowledge, what we've learned about the world.
link |
02:09:07.400
And that is the rare thing.
link |
02:09:09.640
That's the thing we wanna preserve.
link |
02:09:11.480
It's, who cares about our genes?
link |
02:09:13.640
That's not.
link |
02:09:14.480
It's the knowledge.
link |
02:09:16.280
It's the knowledge.
link |
02:09:17.120
That's a really good place to end.
link |
02:09:19.080
Thank you so much for talking to me.
link |
02:09:20.120
No, it was fun.