Dileep George: Brain-Inspired AI
嘉宾：Dileep George
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
本篇目录15 章节
§01Transcript · 00:00–09:34§02Transcript · 09:38–18:44§03Transcript · 18:46–28:13§04Transcript · 28:15–37:32§05Transcript · 37:35–45:49§06Transcript · 45:53–54:05§07Transcript · 54:07–1:02:55§08Transcript · 1:02:58–1:12:49§09Transcript · 1:12:51–1:22:33§10Transcript · 1:22:38–1:32:09§11Transcript · 1:32:13–1:41:40§12Transcript · 1:41:43–1:50:54§13Transcript · 1:50:56–2:00:40§14Transcript · 2:00:44–2:09:49§15Transcript · 2:09:52–2:09:59§01Transcript · 00:00–09:34
The following is a conversation with Dalipe George,
a researcher at the intersection of neuroscience
and artificial intelligence,
co-founder of Vicarious with Scott Phoenix,
and formerly co-founder of Numenta with Jeff Hawkins,
who's been on this podcast, and Donna Dubinsky.
From his early work on hierarchical temporal memory
to recursive cortical networks to today,
Dalipe's always sought to engineer intelligence
that is closely inspired by the human brain.
As a side note, I think we understand very little
about the fundamental principles
underlying the function of the human brain,
but the little we do know gives hints
that may be more useful for engineering intelligence
than any idea in mathematics, computer science, physics,
and scientific fields outside of biology.
And so the brain is a kind of existence proof
that says it's possible.
Keep at it.
I should also say that brain-inspired AI is often overhyped
and use this fodder, just as quantum computing
for a marketing speak,
but I'm not afraid of exploring these
sometimes overhyped areas since where there's smoke,
there's sometimes fire.
Quick summary of the ads.
Three sponsors, Babbel, Raycon Earbuds, and Masterclass.
Please consider supporting this podcast
by clicking the special links in the description
to get the discount.
It really is the best way to support this podcast.
If you enjoy this thing, subscribe on YouTube,
review it with the five stars on Apple Podcasts,
support on Patreon, or connect with me on Twitter
at Lex Friedman.
As usual, I'll do a few minutes of ads now
and never any ads in the middle
that can break the flow of the conversation.
This show is sponsored by Babbel,
an app and website that gets you speaking
in a new language within weeks.
Go to babbel.com and use code Lex to get three months free.
They offer 14 languages, including Spanish, French,
Italian, German, and yes, Russian.
Daily lessons are 10 to 15 minutes, super easy, effective,
designed by over 100 language experts.
Let me read a few lines from the Russian poem,
"Noch, ulitza, fanar, apteka," by Alexander Bloch,
that you'll start to understand if you sign up to Babbel.
(speaking in foreign language)
Now, I say that you'll only start to understand this poem
because Russian starts with a language
and ends with a vodka.
Now, the latter part is definitely not endorsed
or provided by Babbel
and will probably lose me the sponsorship,
but once you graduate from Babbel,
you can enroll in my advanced course
of late night Russian conversation over vodka.
I have not yet developed an app for that.
It's in progress.
So get started by visiting babbel.com
and use code Lex to get three months free.
This show is sponsored by Raycon earbuds.
Get them at buyraycon.com/lex.
They've become my main method of listening to podcasts,
audio books, and music when I run, do pushups and pull-ups,
or just living life.
In fact, I often listen to Brown Noise with them
when I'm thinking deeply about something,
it helps me focus.
They're super comfortable, pair easily, great sound,
great bass, six hours of playtime.
I've been putting in a lot of miles to get ready
for a potential ultra marathon
and listening to audio books on World War II.
The sound is rich and really comes in clear.
So again, get them at buyraycon.com/lex.
This show is sponsored by Masterclass.
Sign up at masterclass.com/lex to get a discount
and to support this podcast.
When I first heard about Masterclass,
I thought it was too good to be true.
I still think it's too good to be true.
For 180 bucks a year, you get an all access pass
to watch courses from, to list some of my favorites,
Chris Hadfield on space exploration,
Neil deGrasse Tyson on scientific thinking
and communication, Will Wright,
creator of SimCity and Sims on game design.
Every time I do this read,
I really want to play a city builder game.
Carlos Santana on guitar, Garry Kasparov on chess,
Daniel Negrano on poker and many more.
Chris Hadfield explaining how rockets work
and the experience of being launched into space alone
is worth the money.
By the way, you can watch it on basically any device.
Once again, sign up at masterclass.com to get a discount
and to support this podcast.
And now here's my conversation with Dilip George.
Do you think we need to understand the brain
in order to build it?
- Yes, if you want to build the brain,
we definitely need to understand how it works.
So Blue Brain or Henry Markram's project
is trying to build a brain without understanding it,
like just trying to put details of the brain
from neuroscience experiments into a giant simulation.
By putting more and more neurons, more and more details.
But that is not going to work
because when it doesn't perform as what you expect it to do,
then what do you do?
You do, you just keep adding more details.
How do you debug it?
So unless you understand,
unless you have a theory about
how the system is supposed to work,
how the pieces are supposed to fit together,
what they're going to contribute, you can't build it.
- At the functional level, understand.
So can you actually linger on
and describe the Blue Brain project?
It's kind of a fascinating principle,
an idea to try to simulate the brain.
We're talking about the human brain, right?
- Right, human brains and rat brains or cat brains
have lots in common.
That the cortex, the neocortex structure is very similar.
So initially they were trying to just simulate a cat brain
and--
- To understand the nature of evil.
- To understand the nature of evil.
Or as it happens in most of the simulations,
you easily get one thing out, which is oscillations.
If you simulate a large number of neurons, they oscillate.
And you can adjust the parameters and say that,
oh, oscillations match the rhythm
that we see in the brain, et cetera.
But--
- Oh, I see.
So the idea is, is the simulation
at the level of individual neurons?
- Yeah, so the Blue Brain project,
the original idea as proposed was,
you put very detailed biophysical neurons,
biophysical models of neurons.
And you interconnect them according to the statistics
of connections that we have found
from real neuroscience experiments.
And then turn it on.
And see what happens.
And these neural models are incredibly complicated
in themselves, right?
Because these neurons are modeled using this idea
called Hodgkin-Huxley models,
which are about how signals propagate in a cable.
And there are active dendrites, all those phenomena,
which those phenomena themselves,
we don't understand that well.
And then we put in connectivity,
which is part guesswork, part observed.
And of course, if we do not have any theory
about how it is supposed to work,
we just have to take whatever comes out of it as,
okay, this is something interesting.
- But in your sense, these models
of the way signal travels along,
like with the axons and all the basic models,
they're too crude?
- Oh, well, actually, they are pretty detailed
and pretty sophisticated.
And they do replicate the neural dynamics.
If you take a single neuron,
and you try to turn on the different channels,
the calcium channels and the different receptors,
and see what the effect of turning on
or off those channels are in the neuron's spike output,
people have built pretty sophisticated models of that.
And they are, I would say, in the regime of correct.
- Well, see, the correctness, that's interesting,
'cause you've mentioned it at several levels.
The correctness is measured
by looking at some kind of aggregate statistics?
- It would be more the spiking dynamics of the--
- Spiking dynamics of the single neuron, okay.
- Yeah, and yeah, these models,
because they are going to the level of mechanism, right?
So they are basically looking at,
okay, what is the effect of turning on an ion channel?
And you can model that using electric circuits.
And then, so they are model,
so it is not just a function fitting,
it is people are looking at the mechanism underlying it,
and putting that in terms of electric circuit theory,
signal propagation theory, and modeling that.
And so those models are sophisticated,
but getting a single neurons model 99% right
does not still tell you how to,
§02Transcript · 09:38–18:44
you know, it would be the analog
of getting a transistor model right,
and now trying to build a microprocessor.
And if you just observe, you know,
if you did not understand how a microprocessor works,
but you say, oh, I now can model one transistor well,
and now I will just try to interconnect the transistors
according to whatever I could, you know,
guess from the experiments and try to simulate it,
then it is very unlikely
that you will produce a functioning microprocessor.
You want to, you know,
when you want to produce a functioning microprocessor,
you want to understand Boolean logic,
how does, how do the gates work, all those things,
and then, you know,
understand how do those gates get implemented
using transistors.
- Yeah, there's actually,
I remember this reminds me, there's a paper,
maybe you're familiar with it,
that I remember going through in a reading group
that approaches a microprocessor
from a perspective of a neuroscientist.
I think it basically,
it uses all the tools that we have of neuroscience
to try to understand,
like as if we just aliens showed up to study computers,
- Yeah.
- And to see if those tools can be used
to get any kind of sense of how the microprocessor works.
I think the final, the takeaway from,
at least this initial exploration is that we're screwed.
There's no way that the tools of neuroscience
would be able to get us to anything,
like not even Boolean logic.
I mean, it's just,
any aspect of the architecture of the function
of the processes involved,
the clocks, the timing, all that,
you can't figure that out from the tools of neuroscience.
- Yeah, so I'm very familiar with this particular paper.
I think it was called,
Can a Neuroscientist Understand a Microprocessor?
- Yeah.
- Something like that.
Following the methodology in that paper,
even an electrical engineer
would not understand microprocessors.
So I could, so,
(both laughing)
So I don't think it is that bad in the sense of saying,
neuroscientists do find valuable things
by observing the brain.
They do find good insights,
but those insights cannot be put together
just as a simulation.
You have to investigate
what are the computational underpinnings
of those findings.
How do all of them fit together
from an information processing perspective?
You have to, somebody has to painstakingly
put those things together and build hypothesis.
So I don't want to diss all of neuroscientists saying,
oh, they're not finding anything.
No, that paper almost went to that level of,
neuroscientists will never understand.
No, that's not true.
I think they do find lots of useful things,
but it has to be put together in a computational framework.
- Yeah, I mean, but just the AI systems
will be listening to this podcast 100 years from now,
and they will probably,
there's some non-zero probability
they'll find your words laughable.
They're like, I remember humans thought
they understood something about the brain
and they were totally clueless.
There's a sense about neuroscience
that we may be in the very, very early days
of understanding the brain.
But I mean, that's one perspective.
In your perspective,
how far are we into understanding
any aspect of the brain?
So the dynamics of the individual neurocommunication
to the, how in a collective sense,
how they're able to store information,
transfer information,
how the intelligence then emerges,
all that kind of stuff.
Where are we on that timeline?
- Yeah, so timelines are very, very hard to predict,
and you can, of course, be wrong.
And it can be wrong on either side.
We know that when we look back,
the first flight was in 1903.
In 1900, there was a New York Times article
on flying machines that do not fly.
And humans might not fly for another 100 years.
That was what that article stated.
And so, but no, they flew three years after that.
So it's very hard to, so-
- Well, and on that point, one of the Wright brothers,
I think two years before,
said that, he said some number, like 50 years,
he has become convinced that it's impossible.
- Even during their experimentation, yeah, yeah, yeah.
- I mean, that's a tribute to when,
that's like the entrepreneurial battle of depression,
of going through just thinking this is impossible.
But there, yeah, there's something,
even the person that's in it
is not able to see, estimate correctly.
- Exactly, but I can tell from the point of,
objectively, what are the things that we know about the brain
and how that can be used to build AI models,
which can then go back and inform how the brain works.
So my way of understanding the brain
would be to basically say,
look at the insights neuroscientists have found,
understand that from a computational angle,
information processing angle, build models using that.
And then building that model, which functions,
which is a functional model,
which is doing the task that we want the model to do.
It is not just trying to model a phenomena in the brain.
It is trying to do what the brain is trying to do
on the whole functional level.
And building that model will help you
fill in the missing pieces that,
biology just gives you the hints.
And building the model,
fills in the rest of the pieces of the puzzle.
And then you can go and connect that back to biology
and say, okay, now it makes sense
that this part of the brain is doing this,
or this layer in the cortical circuit is doing this.
And then continue this iteratively,
because now that will inform new experiments
in neuroscience.
And of course, building the model
and verifying that in the real world
will also tell you more about,
does the model actually work?
And you can refine the model,
find better ways of putting
these neuroscience insights together.
So I would say it is,
so neuroscientists alone, just from experimentation,
will not be able to build a model of the brain,
or a functional model of the brain.
So there's lots of efforts,
which are very impressive efforts
in collecting more and more connectivity data
from the brain.
How are the microcircuits of the brain
connected with each other?
- Those are beautiful, by the way.
- Those are beautiful.
And at the same time, those do not itself,
by themselves, convey the story of how does it work.
And somebody has to understand,
okay, why are they connected like that?
And what are those things doing?
And we do that by building models in AI,
using hints from neuroscience, and repeat the cycle.
- So what aspect of the brain
are useful in this whole endeavor?
Which, by the way, I should say,
you're both a neuroscientist and an AI person.
I guess the dream is to both understand the brain
and to build AGI systems.
So you're, it's like an engineer's perspective
of trying to understand the brain.
So what aspects of the brain, functionally speaking,
like you said, do you find interesting?
- Yeah, quite a lot of things.
So one is, if you look at the visual cortex,
and the visual cortex is a large part of the brain.
I forgot the exact fraction,
but it's a huge part of our brain area
is occupied by just vision.
So vision, visual cortex is not just
a feed-forward cascade of neurons.
There are a lot more feedback connections in the brain
compared to the feed-forward connections.
And it is surprising to the level of detail
neuroscientists have actually studied this.
If you go into neuroscience literature
and poke around and ask, have they studied
what will be the effect of poking a neuron
in level IT in level V1?
And have they studied that?
And you will say, yes, they have studied that.
- So every possible combination has been studied.
- I mean, it's not a random exploration at all.
It's very hypothesis-driven, right?
They are very, experimental neuroscientists
§03Transcript · 18:46–28:13
are very, very systematic in how they probe the brain.
Because experiments are very costly to conduct.
They take a lot of preparation.
They need a lot of control.
So they are very hypothesis-driven
in how they probe the brain.
And often what I find is that when we have a question
in AI about, has anybody probed
how lateral connections in the brain works?
And when you go and read the literature,
yes, people have probed it,
and people have probed it very systematically.
And they have hypothesis about how those lateral connections
are supposedly contributing to visual processing.
But of course, they haven't built
very, very functional detailed models of it.
- By the way, how do they, in those studies,
sorry to interrupt, do they stimulate like a neuron
in one particular area of the visual cortex
and then see how the signal travels kind of thing?
- Fascinating, very, very fascinating experiments.
So I can give you one example I was impressed with.
This is, so before going to that,
let me give you an overview
of how the layers in the cortex are organized.
Visual cortex is organized
into roughly four hierarchical levels.
Okay, so V1, V2, V4, IT.
And in V1-
- What happened to V3?
- Well, yeah, there's another pathway.
Okay, so there is, this is,
I'm talking about just object recognition pathway.
- All right, cool.
- And then in V1 itself,
so it's, there is a very detailed microcircuit in V1 itself.
There is organization within a level itself.
The cortical sheet is organized into multiple layers,
and there are columnar structure.
And this layer-wise and columnar structure
is repeated in V1, V2, V4, IT, all of them, right?
And the connections between these layers within a level,
in V1 itself, there are six layers, roughly,
and the connections between them,
there is a particular structure to them.
And now, so one example of an experiment people did is,
when you present a stimulus, which is,
let's say, requires separating the foreground
from the background of an object.
So it's a textured triangle on a textured background.
And you can check, does the surface settle first,
or does the contour settle first?
- Settle?
- Settle in the sense that the,
so when you finally form the percept of the triangle,
you understand where the contours of the triangle are,
and you also know where the inside of the triangle is,
right, that's when you form the final percept.
Now, you can ask, what is the dynamics
of forming that final percept?
Do the neurons first find the edges
and converge on where the edges are,
and then they find the inner surfaces,
or does it go the other way around?
- The other way around.
So what's the answer?
- In this case, it turns out that it first settles
on the edges, it converges on the edge hypothesis first,
and then the surfaces are filled in
from the edges to the inside.
- That's fascinating.
- And the detail to which you can study this,
it's amazing that you can actually not only find
the temporal dynamics of when this happens,
and then you can also find which layer in V1,
which layer is encoding the edges,
which layer is encoding the surfaces,
and which layer is encoding the feedback,
which layer is encoding the feedforward,
and what's the combination of them
that produces the final percept.
And these kinds of experiments stand out
when you try to explain illusions.
One example of a favorite illusion of mine
is the Kanitsa triangle,
I don't know that you are familiar with this one.
So this is an example where it's a triangle,
but only the corners of the triangle
are shown in the stimulus.
So they look like kind of Pac-Man.
- Oh, the black Pac-Man, yeah.
- And then your visual system hallucinates the edges.
And when you look at it, you will see a faint edge.
And you can go inside the brain and look,
do actually neurons signal the presence of this edge?
And if they signal, how do they do it?
Because they are not receiving anything from the input.
The input is black for those neurons, right?
So how do they signal it?
When does the signaling happen?
So if a real contour is present in the input,
then the neurons immediately signal,
oh, okay, there is an edge here.
When it is an illusory edge,
it is clearly not in the input,
it is coming from the context.
So those neurons fire later.
And you can say that, okay,
it's the feedback connections that is causing them to fire.
And they happen later,
and you can find the dynamics of them.
So these studies are pretty impressive and very detailed.
- So by the way, just a step back,
you said that there may be more feedback connections
than feedforward connections.
First of all, just for like a machine learning folks,
I mean, that's crazy
that there's all these feedback connections.
We often think about,
thanks to deep learning,
you start to think about the human brain
as a kind of feedforward mechanism.
So what the heck are these feedback connections?
- Yeah.
- What's the dynamics?
What are we supposed to think about them?
- Yeah, so this fits into a very beautiful picture
about how the brain works, right?
So the beautiful picture of how the brain works
is that our brain is building a model of the world.
I know, so our visual system is building a model
of how objects behave in the world.
And we are constantly projecting
that model back onto the world.
So what we are seeing is not just a feedforward thing
that just gets interpreted in a feedforward part.
We are constantly projecting our expectations
onto the world.
And what the final percept is a combination
of what we project onto the world,
combined with what the actual sensory input is.
- Almost like trying to calculate the difference
and then trying to interpret the difference.
- Yeah, I wouldn't put it as calculating the difference.
It's more like what is the best explanation
for the input stimulus based on the model
of the world I have?
- Got it, got it.
And that's where all the illusions come in.
But that's an incredibly efficient process.
So the feedback mechanism, it just helps you constantly,
yeah, to hallucinate how the world should be
based on your world model.
And then just looking at if there's novelty,
like trying to explain it.
Hence, that's why movement,
we detect movement really well.
There's all these kinds of things.
And this is like at all different levels
of the cortex you're saying.
This happens at the lowest level, at the highest level.
- Yes, yeah.
In fact, feedback connections are more prevalent
in everywhere in the cortex.
And so one way to think about it,
and there's a lot of evidence for this,
is inference.
So basically, if you have a model of the world,
and when some evidence comes in,
what you are doing is inference.
You are trying to now explain this evidence
using your model of the world.
And this inference includes projecting your model
onto the evidence and taking the evidence
back into the model and doing an iterative procedure.
And this iterative procedure is what happens
using the feedforward feedback propagation.
And feedback affects what you see in the world,
and it also affects feedforward propagation.
And examples are everywhere.
We see these kinds of things everywhere.
The idea that there can be multiple competing hypothesis
in our model, trying to explain the same evidence,
and then you have to kind of make them compete.
And one hypothesis will explain away the other hypothesis
through this competition process.
- Wait, what?
So you have competing models of the world
that try to explain, what do you mean by explain away?
- So this is a classic example in graphical models,
probabilistic models.
So if you-- - Oh, what are those?
- Okay. (laughs)
- I think it's useful to mention
because we'll talk about them more.
- Yeah, yeah.
So neural networks are one class
of machine learning models.
You have distributed set of nodes,
which are called the neurons.
§04Transcript · 28:15–37:32
Each one is doing a dot product
and you can approximate any function
using this multilevel network of neurons.
So that's a class of models
which are useful for function approximation.
There is another class of models in machine learning
called probabilistic graphical models.
And you can think of them as each node in that model
is variable, which is talking about something.
It can be a variable representing,
is an edge present in the input or not?
And at the top of the network,
a node can be representing,
is there an object present in the world or not?
And then, so it is another way of encoding knowledge.
And then once you encode the knowledge,
you can do inference in the right way.
What is the best way to explain some set of evidence
using this model that you encoded?
So when you encode the model,
you are encoding the relationship
between these different variables.
How is the edge connected to the model of the object?
How is the surface connected to the model of the object?
And then, of course,
this is a very distributed, complicated model.
And inference is, how do you explain a piece of evidence
when a set of stimulus comes in?
If somebody tells me there is a 50% probability
that there is an edge here in this part of the model,
how does that affect my belief on whether I should think
that there should be a square present in the image?
So this is the process of inference.
So one example of inference is having this
expanding effect between multiple causes.
So graphical models can be used
to represent causality in the world.
So let's say, you know, your alarm at home
can be triggered by a burglar getting into your house,
or it can be triggered by an earthquake.
Both can be causes of the alarm going off.
So now, you're in your office,
you heard burglar alarm going off,
you are heading home,
thinking that there's a burglar, got it.
But while driving home, if you hear on the radio
that there was an earthquake in the vicinity,
now your strength of evidence
for a burglar getting into their house is diminished.
Because now that piece of evidence is explained
by the earthquake being present.
So if you think about these two causes
explaining at lower level variable, which is alarm,
now what we're seeing is that
increasing the evidence for some cause,
there is evidence coming in from below
for alarm being present.
And initially it was flowing to a burglar being present,
but now since there is side evidence for this other cause,
it explains away this evidence
and evidence will now flow to the other cause.
This is two competing causal things
trying to explain the same evidence.
- And the brain has a similar kind of mechanism
for doing so.
That's kind of interesting.
How's that all encoded in the brain?
Like where's the storage of information?
Are we talking, just maybe to get it
a little bit more specific,
is it in the hardware of the actual connections?
Is it in chemical communication?
Is it electrical communication?
Do we know?
- So this is a paper that we are bringing out soon.
- Which one is this?
- This is the cortical microcircuits paper
that I sent you a draft of.
Of course, a lot of it is still hypothesis.
One hypothesis is that you can think of a cortical column
as encoding a concept.
A concept, think of it as an example of a concept
is an edge present or not,
or is an object present or not.
Okay, so you can think of it as a binary variable,
a binary random variable,
the presence of an edge or not,
or the presence of an object or not.
So each cortical column can be thought of
as representing that one concept, one variable.
And then the connections between these cortical columns
are basically encoding the relationship
between these random variables.
And then there are connections within the cortical column.
Each cortical column is implemented
using multiple layers of neurons
with very, very, very rich structure there.
There are thousands of neurons in a cortical column.
- But that structure is similar
across the different cortical columns.
- Correct, correct.
And also these cortical columns
connect to a substructure called thalamus.
So all cortical columns pass through this substructure.
So our hypothesis is that
the connections between the cortical columns
implement this, that's where the knowledge is stored
about how these different concepts connect to each other.
And then the neurons inside this cortical column
and in thalamus in combination
implement this actual computations in data for inference,
which includes explaining away
and competing between the different hypothesis.
And it is all very,
so what is amazing is that
neuroscientists have actually done experiments
to the tune of showing these things.
They might not be putting it
in the overall inference framework,
but they will show things like,
if I poke this higher level neuron,
it will inhibit through this complicated loop
through the thalamus,
it will inhibit this other column.
So they will do such experiments.
- Do they use terminology of concepts, for example?
So, I mean,
is it something where,
it's easy to anthropomorphize and think about concepts,
like you started moving into logic-based
kind of reasoning systems.
So how would you think of concepts in that kind of way?
Or is it a lot messier, a lot more gray area,
you know, even more gray,
even more messy than the artificial neural network kinds,
kinds of abstractions?
- Easiest way to think of it is a variable, right?
It's a binary variable,
which is showing the presence or absence of something.
- But I guess what I'm asking is,
is that something that we're supposed to think of something
that's human interpretable, of that something?
- It doesn't need to be.
It doesn't need to be human interpretable.
There's no need for it to be human interpretable.
But it's almost like,
you will be able to find some interpretation of it
because it is connected to the other things
that you know about.
- And the point is it's useful somehow.
It's useful as an entity in the graphic,
in connecting to the other entities that are,
let's call them concepts.
Okay, so by the way, are these the cortical microcircuits?
- Correct, these are the cortical microcircuits.
That's what neuroscientists use to talk about the circuits
within a level of the cortex.
So you can think of, let's think of a neural network,
artificial neural network terms.
People talk about the architecture of the,
so how many layers they build,
what is the fan in, fan out, et cetera.
That is the macro architecture.
And then within a layer of the neural network,
you can, the cortical neural network
is much more structured within a level.
There's a lot more intricate structure there.
But even within an artificial neural network,
you can think of feature detection plus pooling as one level.
And so that is kind of a microcircuit.
It's much more complex in the real brain.
So within a level, whatever is that circuitry
within a column of the cortex
and between the layers of the cortex,
that's the microcircuitry.
- I love that terminology.
Machine learning people don't use the circuit terminology,
but they should.
It's nice.
So, okay.
Okay, so that's the cortical microcircuits.
So what's interesting about, what can we say,
what is the paper that you're working on,
propose about the ideas around these cortical microcircuits?
- So this is a fully functional model
for the microcircuits of the visual cortex.
- So the paper focuses, and your idea in our discussions now
is focusing on vision.
- Yeah.
- The visual cortex.
Okay.
So this is a model, this is a full model.
This is how vision works.
- Well, this is a model of--
- A hypothesis.
- Yeah.
- A hypothesis.
- Okay, so let me step back a bit.
So we looked at neuroscience for insights
on how to build a vision model.
§05Transcript · 37:35–45:49
- Right.
- And we synthesized all those insights
into a computational model.
This is called the recursive cortical network model
that we used for breaking CAPTCHAs.
And we are using the same model for robotic picking
and tracking of objects.
- And that again is a vision system.
- That's a vision system.
- Computer vision system.
- That's a computer vision system.
- Takes in images and outputs what?
- On one side, it outputs the class of the image
and also segments the image.
And you can also ask it further queries.
Where is the edge of the object?
Where is the interior of the object?
So it's a model that you build to answer multiple questions.
So you're not trying to build a model
for just classification or just segmentation, et cetera.
It's a joint model that can do multiple things.
And so that's the model that we built
using insights from neuroscience.
And some of those insights are,
what is the role of feedback connections?
What is the role of lateral connections?
So all those things went into the model.
The model actually uses feedback connections.
- All these ideas from neuroscience.
- Yeah.
- So what the heck is a recursive cortical network?
Like what are the architecture approaches,
interesting aspects here,
which is essentially a brain inspired approach
to a computer vision?
- Yeah.
So there are multiple layers to this question.
I can go from the very, very top and then zoom in.
So one important thing, constraint that went into the model
is that you should not think vision,
think of vision as something in isolation.
We should not think perception as something
as a pre-processor for cognition.
Perception and cognition are interconnected.
And so you should not think of one problem
in separation from the other problem.
And so that means if you finally want to have a system
that understand concepts about the world
and can learn in a very conceptual model of the world
and can reason and connect to language,
all of those things,
you need to have, think all the way through
and make sure that your perception system
is compatible with your cognition system
and language system and all of them.
And one aspect of that is top-down controllability.
- What does that mean?
- So that means,
- In this context.
- So think of, you can close your eyes
and think about the details of one object.
I can zoom in further and further.
So think of the bottle in front of me.
And now you can think about,
okay, what the cap of that bottle looks.
I know we can think about what's the texture
on that bottle of the cap.
You can think about what will happen if something hits that.
So you can manipulate your visual knowledge
in cognition driven ways.
- Yes.
- And so this top-down controllability
and being able to simulate scenarios in the world.
- So you're not just a passive player
in this perception game.
You can control it.
You have imagination.
- Correct, correct.
So basically, having a generative network,
which is a model,
and it is not just some arbitrary generative network.
It has to be built in a way that it is controllable top-down.
It is not just trying to generate a whole picture at once.
It's not trying to generate
photorealistic things of the world.
You don't have good photorealistic models of the world.
Human brains do not have.
If I, for example, ask you the question,
what is the color of the letter E in the Google logo?
You have no idea.
- No idea.
- You probably have seen it millions of times.
(laughing)
Or not millions of times, hundreds of times.
(laughing)
So it's not, our model is not photorealistic.
But it has other properties that we can manipulate it.
And you can think about filling in
a different color in that logo.
You can think about expanding the letter E.
So you can imagine the consequence of actions
that you have never performed.
So these are the kind of characteristics
the generative model need to have.
So this is one constraint that went into our model.
So this is, when you read the,
just the perception side of the paper,
it is not obvious that this was a constraint
into the, that went into the model,
this top-down controllability of the generative model.
- So what does top-down controllability in a model look like?
It's a really interesting concept, fascinating concept.
What is that?
Is that the recursiveness gives you that?
Or how do you do it?
- Quite a few things.
It's like, what does the model factor,
factorize, what are the,
what is the model representing as different pieces
in the puzzle?
So in the RCN network, it thinks of the world,
so what I say, the background of an image
is modeled separately from the foreground of the image.
So the objects are separate from the background.
They are different entities.
- So there's a kind of segmentation
that's built in fundamentally to the structure.
And then even that object is composed of parts.
And also, another one is the shape of the object
is differently modeled from the texture of the object.
- Got it.
So there's like these, I've been,
you know who Francois Chollet is?
- Yeah, yeah.
- He's, so there's, he developed this IQ test
type of thing for ARC challenge for,
and it's kind of cool that there's these concepts,
priors that he defines that you bring to the table
in order to be able to reason about basic shapes
and things in an IQ test.
So here you're making it quite explicit
that here are the things that you should be,
these are like distinct things
that you should be able to model in this.
- Keep in mind that you can derive this
from much more general principles.
It doesn't, you don't need to explicitly put it as,
oh, objects versus foreground versus background,
the surface versus texture.
No, these are derivable from more fundamental principles
of how, you know, what's the property
of continuity of natural signals?
- What's the property of continuity of natural signals?
- Yeah.
- By the way, that sounds very poetic, but yeah.
So you're saying that's a,
there's some low-level properties
from which emerges the idea that shapes
should be different than,
like there should be a parts of an object,
there should be, I mean,
- Exactly.
- Kind of like Francois, I mean, there's objectness,
there's all these things that it's kind of crazy
that we humans, I guess, evolved to have
because it's useful for us to perceive the world.
- Correct, correct.
And it derives mostly from the properties
of natural signals.
And so-
- Natural signals.
So natural signals are the kind of things
we'll perceive in the natural world.
- Correct.
- I don't know why that sounds so beautiful,
natural signals, yeah.
- As opposed to a QR code, right?
Which is an artificial signal that we created.
Humans are not very good at classifying QR codes.
We are very good at saying something is a cat or a dog,
but not very good at, you know,
where computers are very good at classifying QR codes.
So our visual system is tuned for natural signals.
And there are fundamental assumptions in the architecture
that are derived from natural signals properties.
- I wonder when you take a hallucinogenic drugs,
does that go into natural or is that closer to QR code?
- It's still natural.
- It's still natural?
- Yeah, because it is still operating using our brains.
- By the way, on that topic, I mean,
I haven't been following,
I think they're becoming legalized in certain,
I can't wait until they become legalized to a degree
that you, like vision science researchers could study it.
- Yeah.
- And then through medical, chemical ways, modify it.
There could be ethical concerns,
but that's another way to study the brain
is to be able to chemically modify it.
§06Transcript · 45:53–54:05
There's probably very long a way
to figure out how to do it ethically.
- Yeah, but I think there are studies on that already.
- Already?
- Yeah, I think so.
Because it's not unethical to give it to rats.
- Oh, that's true, that's true.
(both laughing)
There's a lot of drugged up rats out there.
Okay, cool, sorry, sorry to, so okay.
So there's these low level things from natural signals
that can--
- From which these properties will emerge.
- Yes.
- But it is still a very hard problem
on how to encode that.
So you don't, there is no,
so you mentioned the priors Franscho wanted to encode
in the abstract reasoning challenge,
but it is not straightforward how to encode those priors.
So some of those challenges,
like the object recognition, completion challenges
are things that we purely use our visual system to do.
It looks like abstract reasoning,
but it is purely an output of the vision system.
For example, completing the corners
of that Kaninsa triangle,
completing the lines of that Kaninsa triangle.
It's a purely a visual system property.
There is no abstract reasoning involved.
It uses all these priors,
but it is stored in our visual system in a particular way
that is amenable to inference.
And that is one of the things that we tackled in the,
basically saying, okay, these are the prior knowledge
which will be derived from the world,
but then how is that prior knowledge represented
in the model such that inference,
when some piece of evidence comes in,
can be done very efficiently and in a very distributed way.
Because it is very,
there are so many ways of representing knowledge,
which is not amenable to very quick inference,
you know, quick lookups.
And so that's one core part of what we tackled
in the RCN model.
How do you encode visual knowledge
to do very quick inference?
And yeah.
- Can you maybe comment on,
so folks listening to this in general
may be familiar with different kinds of architectures
of neural networks.
What are we talking about with the RCN?
What does the architecture look like?
What are the different components?
Is it close to neural networks?
Is it far away from neural networks?
What does it look like?
- Yeah, so you can think of the delta between the model
and a convolutional neural network,
if people are familiar with convolutional neural networks.
So convolutional neural networks
have this feed-forward processing cascade,
which is called feature detectors and pooling.
And that is repeated in the hierarchy,
in a multi-level system.
And if you want an intuitive idea of what is happening,
feature detectors are, you know,
detecting interesting co-occurrences in the input.
It can be a line, a corner, an eye,
or a piece of texture, et cetera.
And the pooling neurons are doing
some local transformation of that
and making it invariant to local transformations.
So this is what the structure
of convolutional neural network is.
Recursive cortical network has a similar structure
when you look at just the feed-forward pathway.
But in addition to that,
it is also structured in a way that it is generative.
So that it can run it backward
and combine the forward with the backward.
Another aspect that it has is it has lateral connections.
These lateral connections, which is between,
so if you have an edge here and an edge here,
it has connections between these edges.
It is not just feed-forward connections.
It is something between these edges,
which is the nodes representing these edges,
which is to enforce compatibility between them.
So otherwise what will happen is that-
- Like constraints?
- It's a constraint.
It's basically, if you do just feature detection
followed by pooling,
then your transformations in different parts
of the visual field are not coordinated.
And so you will create jagged,
when you generate from the model,
you will create jagged things
and uncoordinated transformations.
So these lateral connections
are enforcing the transformations.
- Is the whole thing still differentiable?
- No.
- Okay.
- No.
(laughs)
It's not trained using a backprop.
- Okay, that's really important.
So there's these feed-forward,
there's feedback mechanisms.
There's some interesting connectivity things.
It's still layered?
Like-
- Yes, there are multiple levels.
- Multiple layers.
Okay, very, very interesting.
And yeah, okay, so the interconnection between adjacent,
so connections across service constraints
that keep the thing stable.
- Correct.
- Okay, so what else?
- And then there's this idea of doing inference.
A neural network does not do inference on the fly.
So an example of why this inference is important is,
so one of the first applications
that we showed in the paper was to crack text-based captures.
- What are captures, by the way?
(laughs)
- Yeah.
- By the way, one of the most awesome,
like the people don't use this term anymore,
it's human computation, I think.
I love this term.
The guy who created captures,
I think came up with this term.
- Yeah.
- I love it.
Anyway, what are captures?
- So captures are those strings that you fill in
when you're, you know,
if you're opening a new account in Google,
they show you a picture, you know,
usually it used to be set of garbled letters
that you have to kind of figure out
what is that string of characters and type it.
And the reason captures exist is because, you know,
Google or Twitter do not want
automatic creation of accounts.
You can use a computer to create millions of accounts
and use that for nefarious purposes.
So you want to make sure that, to the extent possible,
the interaction that their system is having
is with a human.
So it's called a human interaction proof.
A capture is a human interaction proof.
So this is, captures are by design,
things that are easy for humans to solve,
but hard for computers.
- Hard for robots, yeah.
- Yeah.
So, and text-based captures was the one which is prevalent
until around 2014,
because at that time, text-based captures
were hard for computers to crack.
Even now, they are actually,
in the sense of an arbitrary text-based capture
will be unsolvable even now.
But with the techniques that we have developed,
it can be, you know, you can quickly develop a mechanism
that solves the capture.
- They've probably gotten a lot harder too.
The people, they've been getting cleverer and cleverer
generating these text captures.
- Correct, correct.
- So, okay, so that was one of the things you've tested it on
is these kinds of captures in 2014, '15, that kind of stuff.
So what, I mean, by the way, why captures?
- Yeah, yeah.
Even now, I would say capture
is a very, very good challenge problem.
If you want to understand how human perception works
and if you want to build systems that work
like the human brain.
And I wouldn't say capture is a solved problem.
We have cracked the fundamental defense of captures,
but it is not solved in the way that humans solve it.
So I can give an example.
I can take a five-year-old child
who has just learned characters
and show them any new capture that we create.
They will be able to solve it.
I can show you pretty much any new capture
from any new website.
You'll be able to solve it
without getting any training examples
from that particular style of capture.
- You're assuming I'm human, yeah.
§07Transcript · 54:07–1:02:55
- Yes, yeah, that's right.
So if you are human,
otherwise I will be able to figure that out using this one.
- This whole podcast is just a touring test,
a long touring test.
Anyway, I'm sorry.
So yeah, so humans can figure it out
with very few examples.
- Or no training examples.
No training examples from that particular style of capture.
And so even now this is unreachable
for the current deep learning system.
So basically there is no,
I don't think a system exists
where you can basically say,
train on whatever you want.
And then now say,
hey, I will show you a new capture,
which I did not show you in the training setup.
Will the system be able to solve it?
Still doesn't exist.
So that is the magic of human perception.
And Doug Hofstadter put this very beautifully
in one of his talks.
The central problem in AI is what is the letter A?
If you can build a system that reliably can detect
all the variations of the letter A,
you don't even need to go to the...
- The B and the C.
- Yeah, you don't even need to go to the B and the C
or the strings of characters.
And so that is the spirit at which,
with which we tackle that problem.
- What does he mean by that?
I mean, is it like without training examples,
try to figure out the fundamental elements
that make up the letter A in all of its forms?
- In all of its forms.
It can be, A can be made with two humans standing,
leaning against each other, holding the hands.
And it can be made of leaves.
It can be...
- Yeah, you might have to understand everything
about this world in order to understand the letter A.
- Exactly.
- So it's common sense reasoning, essentially.
- Right.
So to finally, to really solve,
finally to say that you have solved capture,
you have to solve the whole problem.
(both laughing)
- Yeah, okay.
So how does this kind of the RCN architecture
help us to get, do a better job of that kind of thing?
- Yeah, so as I mentioned,
one of the important things was being able to do inference,
being able to dynamically do inference.
- Can you clarify what you mean?
'Cause you said like neural networks don't do inference.
- Yeah.
- So what do you mean by inference in this context then?
- So, okay, so in captures,
what they do to confuse people
is to make these characters crowd together.
- Yes.
- Okay, and when you make the characters crowd together,
what happens is that you will now start seeing
combinations of characters as some other new character
or an existing character.
So you would put an R and N together,
it will start looking like an M.
And so locally, there is very strong evidence
for it being some incorrect character.
But globally, the only explanation that fits together
is something that is different
from what you can find locally.
- Yes.
- So this is inference.
You are basically taking local evidence
and putting it in the global context
and often coming to a conclusion locally,
which is conflicting with the local information.
- So actually, so you mean inference like
in the way it's used when you talk about reasoning,
for example, as opposed to like inference,
which is with artificial neural networks,
which is a single pass to the network.
- Correct.
- Okay.
So like you're basically doing some basic forms of reasoning
like integration of like how local things fit
into the global picture.
- And things like explaining away coming into this one
because you are explaining that piece of evidence
as something else because globally,
that's the only thing that makes sense.
So now you can amortize this inference by,
you know, in a neural network,
if you want to do this, you can brute force it.
You can just show it all combinations of things
that you want your reasoning to work over.
And you can, you know, like just train the hell
out of that neural network.
And it will look like it is doing, you know,
inference on the fly,
but it is really just doing amortized inference.
It is because you have shown it a lot of these combinations
during training time.
So what you want to do is be able to do dynamic inference
rather than just being able to show all those combinations
in the training time.
And that's something we emphasized in the model.
- What does it mean dynamic inference?
Is that that has to do with the feedback thing?
- Yes.
- Like what is dynamic?
I'm trying to visualize what dynamic inference
would be in this case.
Like, what is it doing with the input?
It's shown the input the first time.
- Yeah.
- And it's like, what's changing over temporarily?
What's the dynamics of this inference process?
- So you can think of it as you have at the top of the model,
the characters that you are trained on,
they are the causes.
You're trying to explain the pixels
using the characters as the causes.
The characters are the things that cause the pixels.
- Yeah, so there's this causality thing.
So the reason you mentioned causality, I guess,
is because there's a temporal aspect to this whole thing.
- In this particular case,
the temporal aspect is not important.
It is more like when, if I turn the character on,
the pixels will turn on.
Yeah, it will be after this a little bit, but yeah.
- Okay, so it's causality in the sense
of like a logic causality, like hence inference, okay.
- The dynamics is that even though locally,
it will look like, okay, this is an A.
And locally, just when I look at just that patch of the image,
it looks like an A.
But when I look at it in the context of all the other causes,
it might not, A is not something that makes sense.
So that is something you have to kind of,
you know, recursively figure out.
- Yeah, so, okay, so, and this thing performed pretty well
on the CAPTCHAs.
- Correct.
- And I mean, is there some kind of interesting intuition
you can provide why it did well?
Like, what did it look like?
Is there visualizations that could be human interpretable
to us humans?
- Yes, yeah, so the good thing about the model
is that it is extremely,
so it is not just doing a classification, right?
It is providing a full explanation for the scene.
So when it operates on a scene,
it is coming at back and saying,
look, this is the part is the A,
and these are the pixels that turned on,
these are the pixels in the input that tells,
makes me think that it is an A.
And also these are the portions I hallucinated.
It provides a complete explanation of that form.
And then these are the contours, this is the interior,
and this is in front of this other object.
So that's the kind of explanation
the inference network provides.
So that is useful and interpretable.
And then the kind of errors it makes are also,
I don't want to read too much into it,
but the kind of errors the network makes
are very similar to the kinds of errors
humans would make in a similar situation.
- So there's something about the structure
that feels reminiscent of the way
humans' visual system works.
Well, I mean, how hard-coded is this
to the capture problem, this idea?
- Not really hard-coded because it's the,
the assumptions, as I mentioned, are general, right?
It is more, and those themselves can be applied
in many situations which are natural signals.
So it's the foreground versus background factorization
and the factorization of the surfaces versus the contours.
So these are all generally applicable assumptions.
- In all vision.
So why capture, why attack the capture problem,
which is quite unique in the computer vision context
versus like the traditional benchmarks of ImageNet
and all those kinds of image classification
or even segmentation tasks and all that kind of stuff.
Do you feel like that's, I mean,
what's your thinking about those kinds of benchmarks
in this context?
- I mean, those benchmarks are useful
for deep learning kind of algorithms where you,
§08Transcript · 1:02:58–1:12:49
so the settings that deep learning works in are,
here is my huge training set and here is my test set.
So the training set is almost 100x, 1000x bigger
than the test set in many cases.
What we wanted to do was invert that.
The training set is way smaller than the test set.
- Yes.
- And, you know, capture is a problem that is by definition
hard for computers and it has these good properties
of strong generalization, strong out of training
distribution generalization.
If you are interested in studying that and putting,
having your model have that property,
then it's a good data set to tackle.
- So is there, have you attempted to,
which I think, I believe there's quite a growing body
of work on looking at MNIST and ImageNet without training.
So like taking like the basic challenges,
how, what tiny fraction of the training set can we take
in order to do a reasonable job of the classification task?
Have you explored that angle in these classic benchmarks?
- Yes, so we did do MNIST.
So, you know, so it's not just capture.
So there was also versions of, multiple versions of MNIST,
including the standard version,
which where we inverted the problem,
which is basically saying rather than train
on 60,000 training data, you know,
how quickly can you get to high level accuracy
with very little training data?
- Is there some performance that you remember,
like how well, how well did it do?
How many examples did it need?
- Yeah, I, you know, I remember that it was,
you know, on the order of tens or hundreds of examples
to get into 95% accuracy.
And it was, it was definitely better than the systems,
other systems out there at that time.
- At that time. - Yeah.
- Yeah, they're really pushing it.
I think that's a really interesting space, actually.
I think there's an actual name for MNIST that,
like there's different names
for the different sizes of training sets.
I mean, people are like attacking this problem.
I think it's super interesting.
It's funny how like the MNIST will probably be with us
all the way to AGI.
- Yes. (laughs)
- It's a data set that just sticks by.
It is, it's a clean, simple data set
to study the fundamentals of learning
with just like CAPTCHAs, it's interesting.
Not enough people, I don't know, maybe you can correct me,
but I feel like CAPTCHAs don't show up as often in papers
as they probably should.
- That's correct, yeah.
Because, you know, usually these things have a momentum,
you know, once something gets established
as a standard benchmark, there is a dynamics
of how graduate students operate
and how academic system works
that pushes people to track that benchmark.
- Yeah, to focus.
- Yeah.
- Nobody wants to think outside the box, okay.
- Yes.
- Okay, so good performance on the CAPTCHAs.
What else is there interesting on the RCN side
before we talk about the cortical microscope?
- Yeah, so the same model,
so the important part of the model
was that it trains very quickly
with very little training data.
And it's quite robust to out-of-distribution perturbations.
And we are using that very fruitfully
and advocatiously in many of the robotics tasks
we are solving.
- You're solving that.
Well, let me ask you this kind of touchy question.
I have to, I've spoken with your friend, colleague,
Jeff Hawkins, too.
I mean, I have to kind of ask, there is a bit of,
whenever you have brain-inspired stuff
and you make big claims, big sexy claims,
there's critics, I mean, machine learning subreddit.
Don't get me started on those people.
I mean, criticism is good, but they're a bit over the top.
There is quite a bit of sort of skepticism and criticism.
Is this work really as good as it promises to be?
Do you have thoughts on that kind of skepticism?
Do you have comments on the kind of criticism
we might've received about, is this approach legit?
Is this a promising approach?
Or at least as promising as it seems to be advertised as?
- Yeah, I can comment on it.
So our Arsene paper is published in Science,
which I would argue is a very high quality journal,
very hard to publish in.
And usually it is indicative of the quality of the work.
And I am very, very certain that the ideas
that we brought together in that paper
in terms of the importance of feedback connections,
recursive inference, lateral connections,
coming to best explanation of the scene
as the problem to solve,
trying to solve recognition, segmentation, all jointly
in a way that is compatible with higher level cognition,
top-down attention, all those ideas
that we brought together into something coherent
and workable in the world and tackling a challenging problem,
I think that will stay and that contribution I stand by.
Now, I can tell you a story which is funny
in the context of this.
So if you read the abstract of the paper
and the argument we are putting in,
look, current deep learning systems
take a lot of training data.
They don't use these insights.
And here is our new model,
which is not a deep neural network, it's a graphical model.
It does inference.
This is what the paper is, right?
Now, once the paper was accepted and everything,
it went to the press department in Science,
AAAS Science Office.
We didn't do any press release when it was published.
It went to the press department.
What was the press release that they wrote up?
A new deep learning model. (laughs)
- Solves CAPTCHAs.
- Solves CAPTCHAs.
And so you can see what was being hyped in that thing.
So it's like there is a dynamic in the community.
That especially happens when there are lots of new people
coming into the field and they get attracted to one thing.
And some people are trying to think different
compared to that.
So there is some, I think skepticism in science
is important and it is very much required.
But it's also, it's not skepticism usually,
it's mostly bandwagon effect that is happening
rather than-
- But that's not even that.
I mean, I'll tell you what they react to,
which is like, I'm sensitive to as well.
If you look at just companies, OpenAI, DeepMind,
Vicarious, I mean, there's a little bit of a race
to the top and hype, right?
It's like, it doesn't pay off to be humble.
(laughs)
So like, and the press is just irresponsible often.
They just, I mean, don't get me started
on the state of journalism today.
Like, it seems like the people who write articles
about these things, they literally have not even spent
an hour on the Wikipedia article
about what is neural networks.
They haven't invested just even the language to laziness.
It's like, robots beat humans.
Like, they write this kind of stuff that just,
and then of course the researchers are quite sensitive
to that because it gets a lot of attention.
They're like, why did this word get so much attention?
That's over the top and people get really sensitive.
- The same kind of criticism with,
OpenAI did work with Rubik's Cube with the robot
that people criticized.
Same with GPT-2 and 3, they criticize.
Same thing with DeepMinds with AlphaZero.
I mean, yeah, I'm sensitive to it, but,
and of course with your work, you mentioned deep learning,
but there's something super sexy to the public
about brain-inspired.
I mean, that immediately grabs people's imagination.
Not even like neural networks, but like really brain-inspired.
- Got it.
- Like brain-like neural networks.
That seems really compelling to people and to me as well,
to the world as a narrative.
And so people hook up, hook onto that,
and sometimes the skepticism engine turns on
in the research community and they're skeptical.
But I think putting aside the ideas
of the actual performance on CAPTCHAs
or performance on any dataset,
I mean, to me, all these datasets are useless anyway.
It's nice to have them, but in the grand scheme of things,
they're silly toy examples.
The point is, is there intuition about the ideas,
just like you mentioned,
bringing the ideas together in a unique way?
Is there something there?
Is there some value there?
And is it gonna stand the test of time?
- Yes.
- And that's the hope.
- Yes.
- That's the hope.
- My confidence there is very high.
§09Transcript · 1:12:51–1:22:33
I don't treat brain-inspired as a marketing term.
I am looking into the details of biology
and puzzling over those things,
and I am grappling with those things.
And so it is not a marketing term at all.
You can use it as a marketing term,
and people often use it.
And you can get combined with them.
And when people don't understand
how we are approaching the problem,
it is easy to be misunderstood
and think of it as purely marketing,
but that's not the way we are.
- So you really, I mean, as a scientist,
you believe that if we kinda just stick
to really understanding the brain,
that's the right, like you should constantly meditate
on the how does the brain do this?
'Cause that's going to be really helpful
for engineering intelligence systems.
- Yes, you need to, so I think it's one input,
and it is helpful, but you should know
when to deviate from it too.
So an example is convolutional neural networks, right?
Convolution is not an operation brain implements.
The visual cortex is not convolutional.
Visual cortex has local receptive fields,
local connectivity, but there is no translation
in invariance in the network weights in the visual cortex.
That is a computational trick,
which is a very good engineering trick
that we use for sharing the training
between the different nodes.
So, and that trick will be with us for some time.
It will go away when we have robots
with eyes and heads that move.
And so then that trick will go away.
It will not be useful at that time.
So--
- So the brain doesn't have translational invariance.
It has the focal point, like it has a thing it focuses on.
- Correct, it has a fovea, and because of the fovea,
the receptive fields are not like the copying of the weights,
like the weights in the center are very different
from the weights in the periphery.
- Yes, at the periphery.
I mean, I did this, actually wrote a paper
and just gotten a chance to really study peripheral vision,
which is a fascinating thing.
Very under understood thing of what the,
at every level the brain does with the periphery.
It does some funky stuff.
So it's another kind of trick than convolutional.
Like it does, it's, you know, convolutional,
convolution in neural networks is a trick for efficiency,
is an efficiency trick.
And the brain does a whole nother kind of thing, I guess.
- Correct, correct.
So you need to understand the principles
of processing so that you can still apply engineering tricks
where you want it to.
You don't want to be slavishly mimicking
all the things of the brain.
And so, yeah, so it should be one input
and I think it is extremely helpful,
but it should be the point of really understanding
so that you know when to deviate from it.
- So, okay, that's really cool.
That's work from a few years ago.
So you did work in Numenta with Jeff Hawkins
with hierarchical temporal memory.
How is your just, if you could just give a brief history,
how is your view of the way the models of the brain changed
over the past few years leading up to now?
Is there some interesting aspects
where there was an adjustment
to your understanding of the brain
or is it all just building on top of each other?
- In terms of the higher level ideas,
especially the ones Jeff wrote about in the book,
if you blur out, right, you know.
- Yeah, on intelligence.
- Right, on intelligence.
If you blur out the details and if you just zoom out
and at the higher level idea,
things are, I would say, consistent
with what he wrote about,
but many things will be consistent with that
because it's a blur, you know, when you,
deep learning systems are also, you know,
multi-level, hierarchical, all of those things, right?
But in terms of the detail, a lot of things are different
and those details matter a lot.
So one point of difference I had with Jeff
was how to approach, you know,
how much of biological plausibility and realism
do you want in the learning algorithms?
So when I was there, this was, you know,
almost 10 years ago now, so--
- Yeah, flies when you're having fun.
- I don't know what Jeff thinks now,
but 10 years ago, the difference was that
I did not want to be so constrained on saying,
my learning algorithms need to be biologically plausible
based on some filter of biological plausibility
available at that time.
To me, that is a dangerous cut to make
because we are, you know,
discovering more and more things about the brain
all the time.
New biophysical mechanisms, new channels
are being discovered all the time.
So I don't want to upfront kill off a learning algorithm
just because we don't really understand
the full biophysics or whatever of how the brain learns.
- Exactly, exactly.
- But let me ask, and I'm sorry to interrupt,
like, what's your sense, what's our best understanding
of how the brain learns?
- So things like back propagation, credit assignment,
so many of these algorithms have,
learning algorithms have things in common, right?
It is, back propagation is one way of credit assignment.
There is another algorithm called expectation maximization,
which is, you know, another weight adjustment algorithm.
- But is it your sense the brain does something like this?
- Has to, there is no way around it
in the sense of saying that you do have to adjust
the connections.
- So, and you're saying credit assignment,
you have to reward the connections that were useful
in making a correct prediction and not,
yeah, I guess, but yeah, it doesn't have to be differentiable.
- Yeah, it doesn't have to be differentiable.
- Yeah, but you have to have a, you know,
you have a model that you start with,
you have data comes in, and you have to have a way
of adjusting the model such that it better fits the data.
So that is all of learning, right?
And some of them can be using backprop to do that.
Some of it can be using, you know,
very local graph changes to do that.
That can, you know, many of these learning algorithms
have similar update properties locally
in terms of what the neurons need to do locally.
- I wonder if small differences in learning algorithms
can have huge differences in the actual effect.
So the dynamics of, I mean, sort of the reverse,
like spiking, like if credit assignment is like a lightning
versus like a rainstorm or something,
like whether there's like a looping local type of situation
with the credit assignment,
whether there is like regularization,
like how it injects robustness into the whole thing,
like whether it's chemical or electrical or mechanical,
all those kinds of things.
- Yes.
- I feel like it, that, yeah.
I feel like those differences could be essential, right?
- It could be.
It's just that you don't know enough to,
on the learning side, you don't know enough to say
that is definitely not the way the brain does it.
- Got it.
So you don't want to be stuck to it.
- Right.
- So that, yeah.
So you've been open-minded on that side of things.
- Correct.
On the inference side, on the recognition side,
I am much more amenable to being constrained
because it's much easier to do experiments
because it's like, okay, here's the stimulus.
How many steps did it get to take the answer?
I can trace it back.
I can understand the speed of that computation, et cetera,
much more readily on the inference side.
- Got it.
And then you can't do good experiments on the learning side.
- Correct.
- So let's go right into cortical microcircuits right back.
So what are these ideas
beyond recursive cortical network that you're looking at now?
- So we have made a pass through multiple of the steps
that as I mentioned earlier,
we were looking at perception from the angle of cognition.
It was not just perception for perception's sake.
How do you connect it to cognition?
How do you learn concepts?
And how do you learn abstract reasoning?
Similar to some of the things Francois talked about.
So we have taken one pass through it,
basically saying,
what is the basic cognitive architecture
that you need to have,
which has a perceptual system,
which has a system that learns dynamics of the world,
and then has something like a routine,
program learning system on top of it to learn concepts.
So we have built one, the version 0.1 of that system.
§10Transcript · 1:22:38–1:32:09
This was another science robotics paper.
It's the title of that paper was,
something like cognitive programs.
How do you build cognitive programs?
And-
- And the application there was on manipulation,
robotic manipulation?
- It was, so think of it like this.
Suppose you wanted to tell a new person that you met,
you don't know the language that person uses.
You want to communicate to that person
to achieve some task, right?
So I want to say,
hey, you need to pick up all the red cups
from the kitchen counter and put it here, right?
How do you communicate that, right?
You can show pictures.
You can basically say, look, this is the starting state.
The things are here, this is the ending state.
And what does the person need to understand from that?
The person need to understand
what conceptually happened in those pictures
from the input to the output, right?
So we are looking at pre-verbal conceptual understanding.
Without language, how do you have a set of concepts
that you can manipulate in your head?
And from a set of images of input and output,
can you infer what is happening in those images?
- Got it, with concepts that are pre-language, okay.
So what's it mean for a concept to be pre-language?
Like, why is language so important here?
- So I want to make a distinction between concepts
that are just learned from text,
by just feeding brute force text.
You can start extracting things like, okay,
cow is likely to be on grass.
So those kinds of things you can extract purely from text.
But that's kind of a simple association thing
rather than a concept as an abstraction of something
that happens in the real world, in a grounded way,
that I can simulate it in my mind
and connect it back to the real world.
- And you think kind of the visual world,
concepts in the visual world are somehow lower level
than just the language?
- The lower level kind of makes it feel like,
okay, that's unimportant.
Like, it's more like, I would say the concepts
in the visual and motor system
and the concept learning system,
which if you cut off the language part,
just what we learn by interacting with the world
and abstractions from that,
that is a prerequisite for any real language understanding.
- So you disagree with Chomsky,
'cause he says language is at the bottom of everything.
- No, yeah, I disagree with Chomsky completely
on so many levels, from universal grammar to, yeah.
- So that was a paper in Science B
on the recursive cortical network.
What other interesting problems are there,
the open problems in brain-inspired approaches
that you're thinking about?
- I mean, everything is open, right?
Like, no problem is solved, solved, right?
First, I think of perception as kind of
the first thing that you have to build,
but the last thing that you will be actually solved.
Because if you do not build perception system
in the right way,
you cannot build concept system in the right way.
So you have to build a perception system,
however wrong that might be,
you have to still build that and learn concepts from there
and then keep iterating.
And finally, perception will get solved fully
when perception, cognition, language,
all those things work together finally.
- So what, and that, so great,
we've talked a lot about perception,
but then maybe on the concept side and like common sense
or just general reasoning side,
is there some intuition you can draw from the brain
about how we can do that?
- So I have this classic example I give.
So suppose I give you a few sentences
and then ask you a question following that sentence.
This is a natural language processing problem, right?
So here it goes.
I'm telling you, Sally pounded a nail on the ceiling.
Okay, that's a sentence.
Now I'm asking you a question,
was the nail horizontal or vertical?
- Vertical.
- Okay, how did you answer that?
- Well, I imagined Sally,
it was kind of hard to imagine what the hell she was doing,
but I imagined a visual of the whole situation.
- Exactly, exactly.
So here, I posed a question in natural language.
The answer to that question was,
you got the answer from actually simulating the scene.
Now I can go more and more detail about,
okay, was Sally standing on something while doing this?
Could she have been standing on a light bulb to do this?
I could ask more and more questions about this
and I can ask, make you simulate the scene
in more and more detail, right?
Where is all that knowledge that you're accessing stored?
It is not in your language system.
It was not just by reading text you got that knowledge.
It is stored from the everyday experiences
that you have had from,
and by the age of five, you have pretty much all of this,
and it is stored in your visual system.
It is stored in your motor system in a way
such that it can be accessed through language.
- Got it.
I mean, right.
So here, the language is just,
almost serves as the query into the whole visual cortex
and that does the whole feedback thing.
But I mean, is all reasoning kind of connected
to the perception system in some way?
- You can do a lot of it.
You can still do a lot of it by quick associations
without having to go into the depth.
And by the time you will be right, right?
You can just do quick associations,
but I can easily create tricky situations for you
where that quick associations is wrong
and you have to actually run the simulation.
- So the figuring out how these concepts connect,
do I have a good idea of how to do that?
- That's exactly what-
- That's the-
- One of the problems that we are working on.
And the way we are approaching that is basically saying,
okay, you need to,
so the takeaway is that language is simulation control
and your perceptual plus motor system
is building a simulation of the world.
And so that's basically the way we are approaching it.
And the first thing that we built
was a controllable perceptual system.
And we built a schema networks,
which was a controllable dynamic system.
Then we built a concept learning system
that puts all these things together into programs,
as abstractions that you can run and simulate.
And now we are taking the step of connecting it to language.
And it will be very simple examples initially.
It will not be the GPT-3 like examples,
but it will be grounded simulation-based language.
- And for like the querying would be
like question answering kind of thing?
- Correct, correct.
And it will be in some simple world initially
on, you know, but it will be about,
okay, can the system connect the language
and ground it in the right way
and run the right simulations to come up with the answer.
- And the goal is to try to do things that,
for example, GPT-3 couldn't do.
- Correct.
- Speaking of which,
if we could talk about GPT-3 a little bit,
I think it's an interesting thought provoking
set of ideas that OpenAI is pushing forward.
I think it's good for us to talk about the limits
and the possibilities in neural network.
So in general, what are your thoughts
about this recently released very large
175 billion parameter language model?
- So I haven't directly evaluated it yet.
From what I have seen on Twitter
and other people evaluating it,
it looks very intriguing.
You know, I am very intrigued
by some of the properties it is displaying.
And of course the text generation part of that
was already evident in GPT-2,
you know, that it can generate coherent text
over long distances.
But of course the weaknesses are also pretty visible
in saying that, okay,
it is not really carrying a world state around.
And, you know, sometimes you get sentences like,
I went up the hill to reach the valley or the thing.
You know, some completely incompatible statements.
Or when you're traveling from one place to the other,
it doesn't take into account the time of travel,
things like that.
So those things I think will happen less in GPT-3
because it is trained on even more data.
And so, and it can do even more longer distance coherence.
But it will still have the fundamental limitations
that it doesn't have a world model
and it can't run simulations in its head
to find whether something is true in the world or not.
§11Transcript · 1:32:13–1:41:40
- Do you think within,
so it's taking a huge amount of text from the internet
and forming a compressed representation.
Do you think in that could emerge
something that's an approximation of a world model,
which essentially could be used for reasoning?
I mean, it's a,
I'm not talking about GPT-3,
I'm talking about GPT-4, 5, and GPT-10.
- Yeah, I mean, they will look more impressive than GPT-3.
So you can, if you take that to the extreme,
then a Markov chain of just first order,
and if you go to,
I'm taking it the other extreme.
If you read Shannon's book, right?
He has a model of English text,
which is based on first order Markov chains,
second order Markov chains, third order Markov chains,
and saying that, okay, third order Markov chains
look better than first order Markov chains, right?
So does that mean a first order Markov chain
has a model of the world?
Yes, it does.
So yes, in that level,
when you go higher order models
or more sophisticated structure in the model,
like the transformer networks have,
yes, they have a model of the text world,
but that is not a model of the world.
It's a model of the text world,
and it will have interesting properties
and it will be useful,
but just scaling it up is not going to give us AGI
or natural language understanding or meaning.
- The question is whether being forced
to compress a very large amount of text
forces you to construct things that are very much like,
'cause the ideas of concepts and meaning is a spectrum.
So in order to form that kind of compression,
maybe it will be forced to figure out abstractions
which look awfully a lot like the kind of things
that we think about as concepts,
as world models, as common sense.
Is that possible?
- No, I don't think it is possible
because the information is not there.
- The information is there behind the text, right?
- No, unless somebody has written down all the details
about how everything works in the world
to the absurd amounts like,
okay, it is easier to walk forward than backward,
that you have to open the door to go out of the thing,
doctors wear underwear,
unless all these things somebody has written down somewhere
or somehow the program found it to be useful
for compression from some other text,
the information is not there.
- That's an argument that like text is a lot lower fidelity
than the experience of our physical world.
- Correct, correct.
Picture is worth a thousand words,
like that kind of thing.
- Well, in this case, pictures aren't really,
so the richest aspect of the physical world
isn't even just pictures,
it's the interactivity with the world.
- Exactly.
- It's being able to interact.
It's almost like,
it's almost like if you could interact,
so I disagree, well, maybe I agree with you
that pictures worth a thousand words,
but a thousand--
- It's still, yeah, you could say,
you could capture it with a GPT-X.
- So I wonder if there's some interactive element
where a system could live in text world
where it could be part of the chat,
be part of talking to people.
It's interesting, I mean, fundamentally,
so you're making a statement about the limitation of text.
Okay, so let's say we have a text corpus
that includes basically every experience
we could possibly have.
I mean, just a very large corpus of text
and also interactive components.
I guess the question is whether
the neural network architecture,
these very simple transformers,
but if they had like hundreds of trillions
or whatever comes after trillion parameters,
whether that could store the information needed,
that's architecturally.
Do you have thoughts about the limitation
on that side of things with neural networks?
- I mean, so transformer is still
a feed-forward neural network.
It has a very interesting architecture
which is good for text modeling
and probably some aspects of video modeling,
but it is still a feed-forward architecture.
- You believe in the feedback mechanism, recursion.
- Oh, and also causality,
being able to do counterfactual reasoning,
being able to do interventions,
which is actions in the world.
So all those things require different kinds
of models to be built.
I don't think transformers captures that family.
It is very good at statistical modeling of text
and it will become better and better with more data,
bigger models, but that is only going to get so far.
Finally, when you, so I had this joke on Twitter
saying that, "Hey, this is a model that has read
"all of quantum mechanics and theory of relativity
"and we are asking it to do text completion
"or we are asking it to solve simple puzzles."
That's, when you have AGI,
that's not what you ask the system to do.
If it does, we'll ask the system to do experiments,
what should, and come up with hypothesis
and revise the hypothesis based on evidence
from experiments, all those things, right?
Those are the things that we want the system to do
when we have AGI, not solve simple puzzles.
- Like impressive demo, somebody generating a red button
in HTML.
- Which are all useful,
like there's no dissing the usefulness of it.
- So I get, by the way, I'm playing a little bit
of a devil's advocate, so calm down internet.
So I just, I'm curious almost,
in which ways will a dumb,
but large neural network will surprise us?
- Yeah.
I'm, it's kind of your,
I completely agree with your intuition,
it's just that I don't want to dogmatically,
like 100% put all the chips there.
We've been surprised so much,
even the current GPT-2 and 3 are so surprising.
- Yeah.
- The self-play mechanisms of alpha zero
are really surprising.
And I, reinforcement, the fact that reinforcement learning
works at all to me is really surprising.
The fact that neural networks work at all
is quite surprising, given how nonlinear the space is,
the fact that it's able to find local minima
that are at all reasonable, it's very surprising.
So it's, I wonder sometimes
whether us humans just want it to not,
for AGI not to be such a dumb thing.
(laughing)
So I just,
'cause exactly what you're saying is like,
the ideas of concepts and be able to reason
with those concepts and connect those concepts
in like hierarchical ways,
and then to be able to have world models.
I mean, just everything we're describing
in human language in this poetic way seems to make sense,
that that is what intelligence and reasoning are like.
I wonder if at the core of it, it could be much dumber.
- Well, finally it is still connections
and messages passing over, right?
- Right.
- So in that way it's dumb.
(laughing)
- So I guess the recursion, the feedback mechanism,
that does seem to be a fundamental kind of thing.
Yeah, yeah.
The idea of concepts, also memory.
- Correct.
Yeah, having an episodic memory.
- Yeah.
That seems to be an important thing.
- So how do we get memory?
- So yeah, we have another piece of work
which came out recently on how do you form episodic memories
and form abstractions from them?
And we haven't figured out all the connections of that
to the overall cognitive architecture, but-
- Well, yeah, what are your ideas
about how you could have episodic memory?
- So at least it's very clear that you need
to have two kinds of memory, right?
That's very, very clear, right?
There are things that happen
as statistical patterns in the world,
but then there is the one timeline of things
that happen only once in your life, right?
And this day is not going to happen ever again.
And that needs to be stored as just a stream of strings,
right?
This is my experience.
And then the question is about how do you take
that experience and connect it to the statistical part
of it?
How do you now say that, okay, I experienced this thing.
§12Transcript · 1:41:43–1:50:54
Now I want to be careful about similar situations.
And so you need to be able to index that similarity
using your other giant statistics,
the model of the world that you have learned.
Although the situation came from the episode,
you need to be able to index the other one.
So the episodic memory being implemented
as an indexing over the other model that you're building.
- So the memories remain and they're an index
into this, like the statistical thing that you formed.
- Yeah, statistical causal structural model
that you've built over time.
So it's basically the idea is that the hippocampus
is just storing or sequencing in a set of pointers
that happens over time.
And then whenever you want to reconstitute that memory
and evaluate the different aspects of it,
whether it was good, bad,
do I need to encounter the situation again?
You need the cortex to re-instantiate,
to replay that memory.
- So how do you find that memory?
Which direction is the important direction?
- Both directions are again bidirectional.
- I guess, how do you retrieve the memory?
- So this is again hypothesis, right?
We're making this up.
So when you come to a new situation,
your cortex is doing inference over in the new situation.
And then of course, hippocampus is connected
to different parts of the cortex.
And you have this deja vu situation, right?
Okay, I have seen this thing before.
And then in the hippocampus,
you can have an index of,
okay, this is when it happened as a timeline.
Then you can use the hippocampus
to drive the similar timelines to say,
now I am rather than being driven
by my current input stimuli,
I am going back in time
and rewinding my experience from there.
- Replaying it.
- But putting back into the cortex.
And then putting it back into the cortex,
of course affects what you're going to see next
in your current situation.
- Got it, yeah.
So that's the whole thing,
having a world model and then, yeah.
Connecting to the perception.
Yeah, it does seem to be that that's what's happening.
It'd be, on the neural network side,
it's interesting to think of how we actually do that.
- Yeah, yeah.
- To have a knowledge base.
- Yes, it is possible that you can put many
of these structures into neural networks
and we will find ways of combining properties
of neural networks and graphical models.
So, I mean, it's already started happening.
Graph neural networks are kind of a merge between them.
And there will be more of that.
So, but to me, the direction is pretty clear.
I mean, looking at biology and the history
of evolutionary history of intelligence,
it is pretty clear that, okay,
what we need is more structure in the models
and modeling of the world
and supporting dynamic inference.
- Well, let me ask you, there's a guy named Elon Musk.
There's a company called Neuralink
and there's a general field called
Brain-Computer Interfaces.
It's kind of an interface between your two loves.
- Yes.
- The brain and the intelligence.
So, there's like very direct applications
of Brain-Computer Interfaces for people
with different conditions, more in the short term.
But there's also these sci-fi futuristic kinds of ideas
of AI systems being able to communicate
in a high bandwidth way with the brain, bi-directional.
What are your thoughts about Neuralink
and BCI in general as a possibility?
- So, I think BCI is a cool research area.
And in fact, when I got interested in brains initially,
when I was enrolled at Stanford
and when I got interested in brains,
it was through a brain-computer interface talk
that Krishna Shenoy gave.
That's when I even started thinking about the problem.
So, it is definitely a fascinating research area
and the applications are enormous, right?
So, there's the science fiction scenario
of brains directly communicating.
Let's keep that aside for the time being.
Even just the intermediate milestones they're pursuing,
which are very reasonable as far as I can see,
being able to control an external limb
using direct connections from the brain
and being able to write things into the brain.
So, those are all good steps to take
and they have enormous applications.
People losing limbs being able to control prosthetics,
quadriplegics being able to control something,
so, and therapeutics.
And I also know about another company
working in this space called Paradromics.
They're based on a different electrode array,
but trying to attack some of the same problems.
So, I think it's a very-
- Also surgery?
- Correct, surgically implanted electrodes, yeah.
So, yeah, I think of it as a very, very promising field,
especially when it is helping people
overcome some limitations.
Now, at some point, of course,
it will advance the level of being able to communicate.
- How hard is that problem, do you think?
Like, so, okay, let's say we magically solve
what I think is a really hard problem
of doing all of this safely.
- Yeah.
- So, like, being able to connect electrodes,
and not just thousands, but like millions to the brain.
- I think it's very, very hard
because you also do not know
what will happen to the brain with that, right?
In the sense of how does the brain adapt
to something like that?
- And it's, you know, as we were learning,
the brain is quite, in terms of neuroplasticity,
is pretty malleable.
- Correct.
- So, it's gonna adjust.
- Correct.
- So, the machine learning side,
the computer side is gonna adjust,
and then the brain's gonna adjust.
- Exactly, and then what soup does this land us into?
- The kind of hallucinations you might get from this
that might be pretty intense.
- Yeah, yeah.
- Just connecting to all of Wikipedia.
It's interesting whether we need to be able to figure out
the basic protocol of the brain's communication schemes
in order to get them to, the machine and the brain to talk.
'Cause another possibility is the brain
actually just adjusts to whatever the heck
the computer is doing.
- Exactly, that's the way I think,
I find that to be a more promising way.
It's basically saying, you know,
okay, attach electrodes to some part of the cortex, okay?
And maybe if it is done from birth,
the brain will adapt it such that, you know,
that part is not damaged, it was not used for anything.
These electrodes are attached there, right?
And now, you train that part of the brain
to do this high bandwidth communication
between something else, right?
And if you do it like that, then it is brain adapting to,
and of course, your external system is designed
such that it is adaptable.
Just like we design computers or mouse, keyboard,
all of them to be interacting with humans.
So of course, that feedback system is designed
to be human compatible, but now it is not trying to record
from all of the brain and now, you know,
two system trying to adapt to each other.
It's the brain adapting into one way.
- That's fascinating.
The brain is connected to like the internet.
It's connected.
- Yeah.
- Just imagine, it's connecting it to Twitter
and just taking that stream of information.
Yeah, but again, if we take a step back,
I don't know what your intuition is.
I feel like that is not as hard of a problem
as doing it safely.
There's a huge barrier to surgery.
- Right.
- 'Cause the biological system,
it's a mush of like weird stuff.
- Correct.
So that, the surgery part of it, biology part of it,
the long-term repercussions part of it,
again, I don't know what else will,
we often find after a long time in biology that,
okay, that idea was wrong, right?
So people used to cut off this,
the gland called the thymus or something.
And then they found that,
oh no, that actually causes cancer.
(both laughing)
- And then there's a subtle,
like millions of variables involved.
But this whole process, the nice thing,
§13Transcript · 1:50:56–2:00:40
just like, again, with Elon, just like colonizing Mars,
seems like a ridiculously difficult idea.
But in the process of doing it,
we might learn a lot about the biology,
the neurobiology of the brain,
the neuroscience side of things.
It's like, if you wanna learn something,
do the most difficult version of it.
- Yeah.
- And see what you learn.
- The intermediate steps that they are taking
sounded all very reasonable to me.
- Yeah, it's great.
Well, but like everything with Elon
is the timeline seems insanely fast, so.
- Right.
- That's the only awful question.
- Well, we've been talking about cognition a little bit,
so like reasoning.
We haven't mentioned the other C word,
which is consciousness.
Do you ever think about that one?
Is that useful at all in this whole context
of what it takes to create an intelligent reasoning being?
Or is that completely outside of your,
like the engineering perspective of intelligence?
- It is not outside the realm,
but it doesn't on a day-to-day basis inform what we do,
but it's more, so in many ways,
the company name is connected to this idea of consciousness.
- What's the company name?
- Vicarious, so Vicarious is the company name.
So what does Vicarious mean?
At the first level, it is about modeling the world,
and it is internalizing the external actions.
So you interact with the world
and learn a lot about the world.
And now, after having learned a lot about the world,
you can run those things in your mind
without actually having to act in the world.
So you can run things vicariously, just in your brain.
And similarly, you can experience another person's thoughts
by having a model of how that person works
and running there, putting yourself
in some other person's shoes.
So that is being vicarious.
Now, it's the same modeling apparatus
that you're using to model the external world
or some other person's thoughts.
You can turn it to yourself.
You can, if that same modeling thing
is applied to your own modeling apparatus,
then that is what gives rise to consciousness, I think.
- Well, that's more like self-awareness.
There's the hard problem of consciousness,
which is when the model feels like something,
when this whole process is like,
it's like you really are in it.
You feel like an entity in this world.
Not just you know that you're an entity,
but it feels like something to be that entity.
And thereby, we attribute this,
then it starts to be where in something
that has consciousness can suffer.
You start to have these kinds of things
that we can reason about that.
- Yes.
- It's much heavier.
It seems like there's much greater cost of your decisions.
And mortality is tied up into that.
The fact that these things end.
- Right.
- First of all, I end at some point,
and then other things end.
And that somehow seems to be,
at least for us humans, a deep motivator.
- Yes.
- And that idea of motivation in general,
we talk about goals in AI,
but goals aren't quite the same thing as our mortality.
It feels like first of all, humans don't have a goal.
And they just kind of create goals at different levels.
They like make up goals,
because we're terrified by the mystery
of the thing that gets us all.
So we make these goals up.
So we're like a goal generation machine,
as opposed to a machine which optimizes
the trajectory towards a singular goal.
So it feels like that's an important part of cognition,
that whole mortality thing.
- Well, it is a part of human cognition,
but there is no reason for that mortality
to come to the equation for an artificial system,
because we can copy the artificial system.
The problem with humans is that I can't clone you.
Even if I clone you as the hardware,
your experience that was stored in your brain,
or your episodic memory,
all those will not be captured in the new clone.
But that's not the same with an AI system.
- But it's also possible that the thing
that you mentioned with us humans
is actually of fundamental importance for intelligence.
So the fact that you can copy an AI system
means that that AI system is not yet an AGI.
- So if you look at existence proof,
if we reason based on existence proof,
you could say that it doesn't feel like death
is a fundamental property of an intelligent system,
but we don't yet, give me an example
of an immortal intelligent being, we don't have those.
It's very possible that that is a fundamental property
of intelligence is a thing that has a deadline for itself.
(both laughing)
- You can think of it like this.
Suppose you invent a way to freeze people for a long time.
It's not dying, right?
So you can be frozen and woken up
thousands of years from now.
So it's no fear of death.
- Well, no, you're still, it's not about time.
It's about the knowledge that it's temporary.
And that aspect of it, the finiteness of it,
I think creates a kind of urgency.
- Correct, for us, for humans.
- Yeah, for humans.
- Yes, and that is part of our drives.
But, and that's why I'm not too worried about AI,
having motivations to kill all humans
and those kinds of things.
Why, just wait.
(both laughing)
Why do you need to do that?
- Yeah, I've never heard that before.
That's a good point.
Yeah, just murder seems like a lot of work.
Just wait it out.
They'll probably hurt themselves.
Let me ask you, people often kind of wonder,
world-class researchers such as yourself,
what kind of books, technical, fiction, philosophical,
were, had an impact on you in your life
and maybe ones you could possibly recommend
that others read, maybe if you have three books
that pop into mind.
- Yeah, so I definitely liked Judea Pearl's book,
Probabilistic Reasoning and Intelligent Systems.
It's a very deep technical book,
but what I liked is that,
so there are many places where you can learn
about probabilistic graphical models from.
But throughout this book, Judea Pearl kind of sprinkles
his philosophical observations and he thinks about,
connects us to how the brain thinks and attentions
and resources, all those things.
So that whole thing makes it more interesting to read.
- He emphasizes the importance of causality.
- So that was in his later book.
So this was the first book,
Probabilistic Reasoning and Intelligent Systems.
He mentions causality, but he hadn't really
sunk his teeth into, like, you know,
how do you actually formalize--
- Got it.
- Yeah, and the second book, Causality,
it was the one in 2000, that one is really hard.
So I wouldn't recommend that.
- Oh yeah, so that looks at the, like,
the mathematical, like, his model of--
- Due calculus.
- Due calculus, yeah, it was pretty dense mathematically.
- Right, right, right.
The book of Y is definitely more enjoyable.
- Oh, for sure.
- Yeah, so yeah, so I would recommend
Probabilistic Reasoning and Intelligent Systems.
Another book I liked was one from Doug Hofstadter.
This was a long time ago.
He has a book, he had a book, I think,
called, it was called The Mind's Eye.
It was probably Hofstadter and Daniel Dennett together.
- Yeah, and I actually was, I bought that book.
It's on my shelf, I haven't read it yet.
But I couldn't get an electronic version of it,
which is annoying, 'cause you read everything on Kindle.
- Oh, okay.
- So you have to actually purchase the physical.
It's like one of the only physical books I have,
'cause anyway, a lot of people recommended it highly, so.
- Yeah, and the third one I would definitely
recommend reading is, this is not a technical book.
It is history.
It's called, the name of the book, I think,
is Bishop's Boys.
It's about Wright brothers and their path
and how it was, there are multiple books on this topic
and all of them are great.
It's fascinating how flight was treated
as an unsolvable problem.
§14Transcript · 2:00:44–2:09:49
And also, what aspects did people emphasize?
People thought, oh, it is all about just powerful engines.
Just need to have powerful, lightweight engines.
And so, some people thought of it as,
how far can we just throw the thing?
Just throw it.
- Like a catapult.
- Yeah, so it's a very fascinating,
and even after they made the invention,
people not believing it.
- The social aspect of it, yeah.
- The social aspect, it's very fascinating.
- Do you draw any parallels between birds fly?
So there's the natural approach to flight
and then there's the engineered approach.
Do you see the same kind of thing with the brain
and our trying to engineer intelligence?
- Yeah, it's a good analogy to have.
Of course, all analogies have their--
- Limits, for sure.
- So people in AI often use airplanes as an example of,
hey, we didn't learn anything from birds.
Look, but the funny thing is that,
and the saying is, airplanes don't flap wings.
This is what they say.
The funny thing and the ironic thing is that,
that you don't need to flap to fly
is something Wright Brothers found by observing birds.
(both laughing)
In their notebook, in some of these books,
they show their notebook drawings,
they make detailed notes about buzzards
just soaring over thermals.
And they basically say, look,
flapping is not the important,
propulsion is not the important problem to solve here.
We want to solve control.
And once you solve control,
propulsion will fall into place.
All of these are people,
they realize this by observing birds.
(both laughing)
- Beautifully put.
That's actually brilliant.
'Cause people do use that analogy a lot.
I'm gonna have to remember that one.
Do you have advice for people
interested in artificial intelligence,
like young folks today?
I talk to undergraduate students all the time.
Interested in neuroscience,
interested in understanding how the brain works.
Is there advice you would give them about their career,
maybe about their life in general?
- Sure.
I think every piece of advice
should be taken with a pinch of salt, of course.
Because each person is different,
their motivations are different.
But I can definitely say,
if your goal is to understand the brain
from the angle of wanting to build one,
then being an experimental neuroscientist
might not be the way to go about it.
A better way to pursue it might be through computer science,
electrical engineering, machine learning, and AI.
And of course, you have to study up the neuroscience,
but that you can do on your own.
If you are more attracted by finding something intriguing,
discovering something intriguing about the brain,
then of course it is better to be an experimentalist.
So find that motivation, what are you intrigued by?
And of course, find your strengths too.
Some people are very good experimentalists,
and they enjoy doing that.
- And it's interesting to see which department,
if you're picking in terms of your education path,
whether to go with, at MIT it's brain and computer,
no, it'd be CS.
- Yeah.
- Brain and cognitive sciences, yeah.
Or the CS side of things.
And actually, the brain folks, the neuroscience folks
are more and more now embracing of learning TensorFlow,
PyTorch, right?
They see the power of trying to engineer ideas
that they get from the brain into,
and then explore how those could be used
to create intelligent systems.
So that might be the right department actually.
- Yeah.
So this was a question in one of the
Redwood Neuroscience Institute workshops
that Jeff Hawkins organized almost 10 years ago.
This question was put to a panel, right?
What should be the undergrad major
you should take if you want to understand the brain?
And the majority opinion in that one
was electrical engineering.
- Interesting.
- Because, I mean, I'm a W undergrad,
so I got lucky in that way.
But I think it does have some of the right ingredients,
because you learn about circuits.
You learn about how you can construct circuits
to approach, do functions.
You learn about microprocessors.
You learn information theory.
You learn signal processing.
You learn continuous math.
So in that way, it's a good step
to if you want to go to computer science or neuroscience,
you could, it's a good step.
- The downside, you're more likely to be forced to use MATLAB.
(laughing)
- One of the interesting things about,
I mean, this is changing.
The world is changing.
But certain departments lagged
on the programming side of things,
on developing good habits in software engineering.
But I think that's more and more changing.
And students can take that into their own hands,
like learn to program.
I feel like everybody should learn to program,
because it, like everyone in the sciences,
'cause it empowers, it puts the data at your fingertips.
So you can organize it.
You can find all kinds of things in the data.
And then you can also, for the appropriate sciences,
build systems that, like based on that.
So like then engineer intelligence systems.
We already talked about mortality.
So we hit a ridiculous point.
But let me ask you the,
one of the things about intelligence is it's goal-driven.
And you study the brain.
So the question is like,
what's the goal that the brain is operating under?
What's the meaning of it all for us humans, in your view?
What's the meaning of life?
(laughing)
- The meaning of life is whatever you construct out of it.
It's completely open.
- It's open?
- Yeah.
- So there's nothing, like you mentioned,
you like constraints.
So there's, what's, it's wide open.
Is there some useful aspect that you think about
in terms of like the openness of it
and just the basic mechanisms of generating goals
in studying cognition in the brain that you think about?
Or is it just about, 'cause everything we've talked about,
kind of the perception system,
is to understand the environment.
That's like to be able to like not die.
- Correct, exactly.
- Like not fall over and like be able to,
you don't think we need to think about
anything bigger than that?
- Yeah, I think so.
Because it's basically being able to understand
the machinery of the world,
such that you can pursue whatever goals you want, right?
- So the machinery of the world is really ultimately
what we should be striving to understand.
The rest is just whatever the heck you wanna do,
or whatever fun you--
- World is culturally popular, you know?
(laughing)
- I think that's beautifully put.
I don't think there's a better way to end it.
I'm so honored that you would show up here
and waste your time with me.
It's been an awesome conversation.
Thanks so much for talking today.
- Oh, thank you so much.
This was so much more fun than I expected.
(laughing)
Thank you.
- Thanks for listening to this conversation
with Delete George.
And thank you to our sponsors,
Babbel, Raycon Earbuds, and Masterclass.
Please consider supporting this podcast
by going to babbel.com and use code Lex,
going to buyraycon.com/lex,
and signing up at masterclass.com/lex.
Click the links, get the discount.
It really is the best way to support this podcast.
If you enjoy this thing, subscribe on YouTube,
review the Five Stars on Apple Podcast,
support it on Patreon, or connect with me on Twitter,
@LexFriedman, spelled, yes, without the E,
just F-R-I-D-M-A-N.
And now let me leave you with some words
from Marcus Aurelius.
You have power over your mind, not outside events.
Realize this and you will find strength.
§15Transcript · 2:09:52–2:09:59
Thank you for listening and hope to see you next time.
(upbeat music)
(upbeat music)