back to indexRisto Miikkulainen: Neuroevolution and Evolutionary Computation | Lex Fridman Podcast #177
link |
The following is a conversation with Risto Michaelainen,
link |
a computer scientist at University of Texas at Austin
link |
and Associate Vice President
link |
of Evolutionary Artificial Intelligence at Cognizant.
link |
He specializes in evolutionary computation,
link |
but also many other topics in artificial intelligence,
link |
cognitive science, and neuroscience.
link |
Quick mention of our sponsors,
link |
Jordan Harbin's show, Grammarly, Belcampo, and Indeed.
link |
Check them out in the description to support this podcast.
link |
As a side note, let me say that nature inspired algorithms
link |
from ant colony optimization to genetic algorithms
link |
to cellular automata to neural networks
link |
have always captivated my imagination,
link |
not only for their surprising power
link |
in the face of long odds,
link |
but because they always opened up doors
link |
to new ways of thinking about computation.
link |
It does seem that in the long arc of computing history,
link |
running toward biology, not running away from it
link |
is what leads to long term progress.
link |
This is the Lex Friedman podcast,
link |
and here is my conversation with Risto Michaelainen.
link |
If we ran the Earth experiment,
link |
this fun little experiment we're on,
link |
over and over and over and over a million times
link |
and watch the evolution of life as it pans out,
link |
how much variation in the outcomes of that evolution
link |
do you think we would see?
link |
Now, we should say that you are a computer scientist.
link |
That's actually not such a bad question
link |
for a computer scientist,
link |
because we are building simulations of these things,
link |
and we are simulating evolution,
link |
and that's a difficult question to answer in biology,
link |
but we can build a computational model
link |
and run it million times and actually answer that question.
link |
How much variation do we see when we simulate it?
link |
And that's a little bit beyond what we can do today,
link |
but I think that we will see some regularities,
link |
and it took evolution also a really long time
link |
and then things accelerated really fast towards the end.
link |
But there are things that need to be discovered,
link |
and they probably will be over and over again,
link |
like manipulation of objects,
link |
and also some way to communicate,
link |
maybe orally, like when you have speech,
link |
it might be some other kind of sounds,
link |
and decision making, but also vision.
link |
Eye has evolved many times.
link |
Various vision systems have evolved.
link |
So we would see those kinds of solutions,
link |
I believe, emerge over and over again.
link |
They may look a little different,
link |
but they get the job done.
link |
The really interesting question is,
link |
would we have primates?
link |
Would we have humans or something that resembles humans?
link |
And would that be an apex of evolution after a while?
link |
We don't know where we're going from here,
link |
but we certainly see a lot of tool use
link |
and building, constructing our environment.
link |
So I think that we will get that.
link |
We get some evolution producing,
link |
some agents that can do that,
link |
manipulate the environment and build.
link |
What do you think is special about humans?
link |
Like if you were running the simulation
link |
and you observe humans emerge,
link |
like these tool makers,
link |
they start a fire and all this stuff,
link |
start running around, building buildings,
link |
and then running for president and all those kinds of things.
link |
What would be, how would you detect that?
link |
Cause you're like really busy
link |
as the creator of this evolutionary system.
link |
So you don't have much time to observe,
link |
like detect if any cool stuff came up, right?
link |
How would you detect humans?
link |
Well, you are running the simulation.
link |
So you also put in visualization
link |
and measurement techniques there.
link |
So if you are looking for certain things like communication,
link |
you'll have detectors to find out whether that's happening,
link |
even if it's a large simulation.
link |
And I think that that's what we would do.
link |
We know roughly what we want,
link |
intelligent agents that communicate, cooperate, manipulate,
link |
and we would build detections
link |
and visualizations of those processes.
link |
Yeah, and there's a lot of,
link |
we'd have to run it many times
link |
and we have plenty of time to figure out
link |
how we detect the interesting things.
link |
But also, I think we do have to run it many times
link |
because we don't quite know what shape those will take
link |
and our detectors may not be perfect for them
link |
Well, that seems really difficult to build a detector
link |
of intelligent or intelligent communication.
link |
Sort of, if we take an alien perspective,
link |
observing earth, are you sure that they would be able
link |
to detect humans as the special thing?
link |
Wouldn't they be already curious about other things?
link |
There's way more insects by body mass, I think,
link |
than humans by far, and colonies.
link |
Obviously, dolphins is the most intelligent creature
link |
on earth, we all know this.
link |
So it could be the dolphins that they detect.
link |
It could be the rockets that we seem to be launching.
link |
That could be the intelligent creature they detect.
link |
It could be some other trees.
link |
Trees have been here a long time.
link |
I just learned that sharks have been here
link |
400 million years and that's longer
link |
than trees have been here.
link |
So maybe it's the sharks, they go by age.
link |
Like there's a persistent thing.
link |
Like if you survive long enough,
link |
especially through the mass extinctions,
link |
that could be the thing your detector is detecting.
link |
Humans have been here for a very short time
link |
and we're just creating a lot of pollution,
link |
but so is the other creatures.
link |
So I don't know, do you think you'd be able
link |
Like how would you go about detecting
link |
in the computational sense?
link |
Maybe we can leave humans behind.
link |
In the computational sense, detect interesting things.
link |
Do you basically have to have a strict objective function
link |
by which you measure the performance of a system
link |
or can you find curiosities and interesting things?
link |
Yeah, well, I think that the first measurement
link |
would be to detect how much of an effect
link |
you can have in your environment.
link |
So if you look around, we have cities
link |
and that is constructed environments.
link |
And that's where a lot of people live, most people live.
link |
So that would be a good sign of intelligence
link |
that you don't just live in an environment,
link |
but you construct it to your liking.
link |
And that's something pretty unique.
link |
I mean, there are certainly birds build nests
link |
but they don't build quite cities.
link |
Termites build mounds and ice and things like that.
link |
But the complexity of the human construction cities,
link |
I think would stand out even to an external observer.
link |
Of course, that's what a human would say.
link |
Yeah, and you know, you can certainly say
link |
that sharks are really smart
link |
because they've been around so long
link |
and they haven't destroyed their environment,
link |
which humans are about to do,
link |
which is not a very smart thing.
link |
But we'll get over it, I believe.
link |
And we can get over it by doing some construction
link |
that actually is benign
link |
and maybe even enhances the resilience of nature.
link |
So you mentioned the simulation that we run over and over
link |
might start, it's a slow start.
link |
So do you think how unlikely, first of all,
link |
I don't know if you think about this kind of stuff,
link |
but how unlikely is step number zero,
link |
which is the springing up,
link |
like the origin of life on earth?
link |
And second, how unlikely is the,
link |
anything interesting happening beyond that?
link |
So like the start that creates
link |
all the rich complexity that we see on earth today.
link |
Yeah, there are people who are working
link |
on exactly that problem from primordial soup.
link |
How do you actually get self replicating molecules?
link |
And they are very close.
link |
With a little bit of help, you can make that happen.
link |
So of course we know what we want,
link |
so they can set up the conditions
link |
and try out conditions that are conducive to that.
link |
For evolution to discover that, that took a long time.
link |
For us to recreate it probably won't take that long.
link |
And the next steps from there,
link |
I think also with some handholding,
link |
I think we can make that happen.
link |
But with evolution, what was really fascinating
link |
was eventually the runaway evolution of the brain
link |
that created humans and created,
link |
well, also other higher animals,
link |
that that was something that happened really fast.
link |
And that's a big question.
link |
Is that something replicable?
link |
Is that something that can happen?
link |
And if it happens, does it go in the same direction?
link |
That is a big question to ask.
link |
Even in computational terms,
link |
I think that it's relatively possible to come up here,
link |
create an experiment where we look at the primordial soup
link |
and the first couple of steps
link |
of multicellular organisms even.
link |
But to get something as complex as the brain,
link |
we don't quite know the conditions for that.
link |
And how do you even get started
link |
and whether we can get this kind of runaway evolution
link |
From a detector perspective,
link |
if we're observing this evolution,
link |
what do you think is the brain?
link |
What do you think is the, let's say, what is intelligence?
link |
So in terms of the thing that makes humans special,
link |
we seem to be able to reason,
link |
we seem to be able to communicate.
link |
But the core of that is this something
link |
in the broad category we might call intelligence.
link |
So if you put your computer scientist hat on,
link |
is there a favorite ways you like to think about
link |
that question of what is intelligence?
link |
Well, my goal is to create agents that are intelligent.
link |
Not to define what.
link |
And that is a way of defining it.
link |
And that means that it's some kind of an object
link |
or a program that has limited sensory
link |
and effective capabilities interacting with the world.
link |
And then also a mechanism for making decisions.
link |
So with limited abilities like that, can it survive?
link |
Survival is the simplest goal,
link |
but you could also give it other goals.
link |
Can it solve problems that you give it?
link |
And that is quite a bit less than human intelligence.
link |
There are, animals would be intelligent, of course,
link |
with that definition.
link |
And you might have even some other forms of life, even.
link |
So intelligence in that sense is a survival skill
link |
given resources that you have and using your resources
link |
so that you will stay around.
link |
Do you think death, mortality is fundamental to an agent?
link |
So like there's, I don't know if you're familiar,
link |
there's a philosopher named Ernest Becker
link |
who wrote The Denial of Death and his whole idea.
link |
And there's folks, psychologists, cognitive scientists
link |
that work on terror management theory.
link |
And they think that one of the special things about humans
link |
is that we're able to sort of foresee our death, right?
link |
We can realize not just as animals do,
link |
sort of constantly fear in an instinctual sense,
link |
respond to all the dangers that are out there,
link |
but like understand that this ride ends eventually.
link |
And that in itself is the force behind
link |
all of the creative efforts of human nature.
link |
That's the philosophy.
link |
I think that makes sense, a lot of sense.
link |
I mean, animals probably don't think of death the same way,
link |
but humans know that your time is limited
link |
and you wanna make it count.
link |
And you can make it count in many different ways,
link |
but I think that has a lot to do with creativity
link |
and the need for humans to do something
link |
beyond just surviving.
link |
And now going from that simple definition
link |
to something that's the next level,
link |
I think that that could be the second level of definition,
link |
that intelligence means something,
link |
that you do something that stays behind you,
link |
that's more than your existence.
link |
You create something that is useful for others,
link |
is useful in the future, not just for yourself.
link |
And I think that's the nicest definition of intelligence
link |
within a next level.
link |
And it's also nice because it doesn't require
link |
that they are humans or biological.
link |
They could be artificial agents that are intelligence.
link |
They could achieve those kind of goals.
link |
So particular agent, the ripple effects of their existence
link |
on the entirety of the system is significant.
link |
So like they leave a trace where there's like a,
link |
yeah, like ripple effects.
link |
But see, then you go back to the butterfly
link |
with the flap of a wing and then you can trace
link |
a lot of like nuclear wars
link |
and all the conflicts of human history,
link |
somehow connected to that one butterfly
link |
that created all of the chaos.
link |
So maybe that's not, maybe that's a very poetic way
link |
to think that that's something we humans
link |
in a human centric way wanna hope we have this impact.
link |
Like that is the secondary effect of our intelligence.
link |
We've had the long lasting impact on the world,
link |
but maybe the entirety of physics in the universe
link |
has a very long lasting effects.
link |
Sure, but you can also think of it.
link |
What if like the wonderful life, what if you're not here?
link |
Will somebody else do this?
link |
Is it something that you actually contributed
link |
because you had something unique to compute?
link |
That contribute, that's a pretty high bar though.
link |
So, you have to be Mozart or something to actually
link |
reach that level that nobody would have developed that,
link |
but other people might have solved this equation
link |
if you didn't do it, but also within limited scope.
link |
I mean, during your lifetime or next year,
link |
you could contribute something that unique
link |
that other people did not see.
link |
And then that could change the way things move forward
link |
So, I don't think we have to be Mozart
link |
to be called intelligence,
link |
but we have this local effect that is changing.
link |
If you weren't there, that would not have happened.
link |
And it's a positive effect, of course,
link |
you want it to be a positive effect.
link |
Do you think it's possible to engineer
link |
into computational agents, a fear of mortality?
link |
Like, does that make any sense?
link |
So, there's a very trivial thing where it's like,
link |
you could just code in a parameter,
link |
which is how long the life ends,
link |
but more of a fear of mortality,
link |
like awareness of the way that things end
link |
and somehow encoding a complex representation of that fear,
link |
which is like, maybe as it gets closer,
link |
you become more terrified.
link |
I mean, there seems to be something really profound
link |
about this fear that's not currently encodable
link |
in a trivial way into our programs.
link |
Well, I think you're referring to the emotion of fear,
link |
something, because we have cognitively,
link |
we know that we have limited lifespan
link |
and most of us cope with it by just,
link |
hey, that's what the world is like
link |
and I make the most of it.
link |
But sometimes you can have like a fear that's not healthy,
link |
that paralyzes you, that you can't do anything.
link |
And somewhere in between there,
link |
not caring at all and getting paralyzed because of fear
link |
is a normal response,
link |
which is a little bit more than just logic
link |
So now the question is, what good are emotions?
link |
I mean, they are quite complex
link |
and there are multiple dimensions of emotions
link |
and they probably do serve a survival function,
link |
heightened focus, for instance.
link |
And fear of death might be a really good emotion
link |
when you are in danger, that you recognize it,
link |
even if it's not logically necessarily easy to derive
link |
and you don't have time for that logical deduction,
link |
you may be able to recognize the situation is dangerous
link |
and this fear kicks in and you all of a sudden perceive
link |
the facts that are important for that.
link |
And I think that's generally is the role of emotions.
link |
It allows you to focus what's relevant for your situation.
link |
And maybe if fear of death plays the same kind of role,
link |
but if it consumes you and it's something that you think
link |
in normal life when you don't have to,
link |
then it's not healthy and then it's not productive.
link |
Yeah, but it's fascinating to think
link |
how to incorporate emotion into a computational agent.
link |
It almost seems like a silly statement to make,
link |
but it perhaps seems silly because we have
link |
such a poor understanding of the mechanism of emotion,
link |
of fear, of, I think at the core of it
link |
is another word that we know nothing about,
link |
but say a lot, which is consciousness.
link |
Do you ever in your work, or like maybe on a coffee break,
link |
think about what the heck is this thing consciousness
link |
and is it at all useful in our thinking about AI systems?
link |
Yes, it is an important question.
link |
You can build representations and functions,
link |
I think into these agents that act like emotions
link |
and consciousness perhaps.
link |
So I mentioned emotions being something
link |
that allow you to focus and pay attention,
link |
filter out what's important.
link |
Yeah, you can have that kind of a filter mechanism
link |
and it puts you in a different state.
link |
Your computation is in a different state.
link |
Certain things don't really get through
link |
and others are heightened.
link |
Now you label that box emotion.
link |
I don't know if that means it's an emotion,
link |
but it acts very much like we understand
link |
what emotions are.
link |
And we actually did some work like that,
link |
modeling hyenas who were trying to steal a kill from lions,
link |
which happens in Africa.
link |
I mean, hyenas are quite intelligent,
link |
but not really intelligent.
link |
And they have this behavior
link |
that's more complex than anything else they do.
link |
They can band together, if there's about 30 of them or so,
link |
they can coordinate their effort
link |
so that they push the lions away from a kill.
link |
Even though the lions are so strong
link |
that they could kill a hyena by striking with a paw.
link |
But when they work together and precisely time this attack,
link |
the lions will leave and they get the kill.
link |
And probably there are some states
link |
like emotions that the hyenas go through.
link |
The first, they call for reinforcements.
link |
They really want that kill, but there's not enough of them.
link |
So they vocalize and there's more people,
link |
more hyenas that come around.
link |
And then they have two emotions.
link |
They're very afraid of the lion, so they want to stay away,
link |
but they also have a strong affiliation between each other.
link |
And then this is the balance of the two emotions.
link |
And also, yes, they also want the kill.
link |
So it's both repelled and attractive.
link |
But then this affiliation eventually is so strong
link |
that when they move, they move together,
link |
they act as a unit and they can perform that function.
link |
So there's an interesting behavior
link |
that seems to depend on these emotions strongly
link |
and makes it possible, coordinate the actions.
link |
And I think a critical aspect of that,
link |
the way you're describing is emotion there
link |
is a mechanism of social communication,
link |
of a social interaction.
link |
Maybe humans won't even be that intelligent
link |
or most things we think of as intelligent
link |
wouldn't be that intelligent without the social component
link |
Maybe much of our intelligence
link |
is essentially an outgrowth of social interaction.
link |
And maybe for the creation of intelligent agents,
link |
we have to be creating fundamentally social systems.
link |
Yes, I strongly believe that's true.
link |
And yes, the communication is multifaceted.
link |
I mean, they vocalize and call for friends,
link |
but they also rub against each other and they push
link |
and they do all kinds of gestures and so on.
link |
So they don't act alone.
link |
And I don't think people act alone very much either,
link |
at least normal, most of the time.
link |
And social systems are so strong for humans
link |
that I think we build everything
link |
on top of these kinds of structures.
link |
And one interesting theory around that,
link |
bigotness theory, for instance, for language,
link |
but language origins is that where did language come from?
link |
And it's a plausible theory that first came social systems,
link |
that you have different roles in a society.
link |
And then those roles are exchangeable,
link |
that I scratch your back, you scratch my back,
link |
we can exchange roles.
link |
And once you have the brain structures
link |
that allow you to understand actions
link |
in terms of roles that can be changed,
link |
that's the basis for language, for grammar.
link |
And now you can start using symbols
link |
to refer to objects in the world.
link |
And you have this flexible structure.
link |
So there's a social structure
link |
that's fundamental for language to develop.
link |
Now, again, then you have language,
link |
you can refer to things that are not here right now.
link |
And that allows you to then build all the good stuff
link |
about planning, for instance, and building things and so on.
link |
So yeah, I think that very strongly humans are social
link |
and that gives us ability to structure the world.
link |
But also as a society, we can do so much more
link |
because one person does not have to do everything.
link |
You can have different roles
link |
and together achieve a lot more.
link |
And that's also something
link |
we see in computational simulations today.
link |
I mean, we have multi agent systems that can perform tasks.
link |
This fascinating demonstration, Marco Dorego,
link |
I think it was, these little robots
link |
that had to navigate through an environment
link |
and there were things that are dangerous,
link |
like maybe a big chasm or some kind of groove, a hole,
link |
and they could not get across it.
link |
But if they grab each other with their gripper,
link |
they formed a robot that was much longer under the team
link |
and this way they could get across that.
link |
So this is a great example of how together
link |
we can achieve things we couldn't otherwise.
link |
Like the hyenas, you know, alone they couldn't,
link |
but as a team they could.
link |
And I think humans do that all the time.
link |
We're really good at that.
link |
Yeah, and the way you described the system of hyenas,
link |
it almost sounds algorithmic.
link |
Like the problem with humans is they're so complex,
link |
it's hard to think of them as algorithms.
link |
But with hyenas, there's a, it's simple enough
link |
to where it feels like, at least hopeful
link |
that it's possible to create computational systems
link |
Yeah, that's exactly why we looked at that.
link |
As opposed to humans.
link |
Like I said, they are intelligent,
link |
but they are not quite as intelligent as say, baboons,
link |
which would learn a lot and would be much more flexible.
link |
The hyenas are relatively rigid in what they can do.
link |
And therefore you could look at this behavior,
link |
like this is a breakthrough in evolution about to happen.
link |
That they've discovered something about social structures,
link |
communication, about cooperation,
link |
and it might then spill over to other things too
link |
in thousands of years in the future.
link |
Yeah, I think the problem with baboons and humans
link |
is probably too much is going on inside the head.
link |
We won't be able to measure it if we're observing the system.
link |
With hyenas, it's probably easier to observe
link |
the actual decision making and the various motivations
link |
that are involved.
link |
Yeah, they are visible.
link |
And we can even quantify possibly their emotional state
link |
because they leave droppings behind.
link |
And there are chemicals there that can be associated
link |
with neurotransmitters.
link |
And we can separate what emotions they might have
link |
experienced in the last 24 hours.
link |
What to you is the most beautiful, speaking of hyenas,
link |
what to you is the most beautiful nature inspired algorithm
link |
in your work that you've come across?
link |
Something maybe early on in your work or maybe today?
link |
I think evolution computation is the most amazing method.
link |
So what fascinates me most is that with computers
link |
is that you can get more out than you put in.
link |
I mean, you can write a piece of code
link |
and your machine does what you told it.
link |
I mean, this happened to me in my freshman year.
link |
It did something very simple and I was just amazed.
link |
I was blown away that it would get the number
link |
and it would compute the result.
link |
And I didn't have to do it myself.
link |
But if you push that a little further,
link |
you can have machines that learn and they might learn patterns.
link |
And already say deep learning neural networks,
link |
they can learn to recognize objects, sounds,
link |
patterns that humans have trouble with.
link |
And sometimes they do it better than humans.
link |
And that's so fascinating.
link |
And now if you take that one more step,
link |
you get something like evolutionary algorithms
link |
that discover things, they create things,
link |
they come up with solutions that you did not think of.
link |
And that just blows me away.
link |
It's so great that we can build systems, algorithms
link |
that can be in some sense smarter than we are,
link |
that they can discover solutions that we might miss.
link |
A lot of times it is because we have as humans,
link |
we have certain biases,
link |
we expect the solutions to be certain way
link |
and you don't put those biases into the algorithm
link |
so they are more free to explore.
link |
And evolution is just absolutely fantastic explorer.
link |
And that's what really is fascinating.
link |
Yeah, I think I get made fun of a bit
link |
because I currently don't have any kids,
link |
but you mentioned programs.
link |
I mean, do you have kids?
link |
So maybe you could speak to this,
link |
but there's a magic to the creative process.
link |
Like with Spot, the Boston Dynamics Spot,
link |
but really any robot that I've ever worked on,
link |
it just feels like the similar kind of joy
link |
I imagine I would have as a father.
link |
Not the same perhaps level,
link |
but like the same kind of wonderment.
link |
Like there's exactly this,
link |
which is like you know what you had to do initially
link |
to get this thing going.
link |
Let's speak on the computer science side,
link |
like what the program looks like,
link |
but something about it doing more
link |
than what the program was written on paper
link |
is like that somehow connects to the magic
link |
of this entire universe.
link |
Like that's like, I feel like I found God.
link |
Every time I like, it's like,
link |
because you've really created something that's living.
link |
Even if it's a simple program.
link |
It has a life of its own, it has the intelligence of its own.
link |
It's beyond what you actually thought.
link |
And that is, I think it's exactly spot on.
link |
That's exactly what it's about.
link |
You created something and it has a ability
link |
to live its life and do good things
link |
and you just gave it a starting point.
link |
So in that sense, I think it's,
link |
that may be part of the joy actually.
link |
But you mentioned creativity in this context,
link |
especially in the context of evolutionary computation.
link |
So, we don't often think of algorithms as creative.
link |
So how do you think about creativity?
link |
Yeah, algorithms absolutely can be creative.
link |
They can come up with solutions that you don't think about.
link |
I mean, creativity can be defined.
link |
A couple of requirements has to be new.
link |
It has to be useful and it has to be surprising.
link |
And those certainly are true with, say,
link |
evolutionary computation discovering solutions.
link |
So maybe an example, for instance,
link |
we did this collaboration with MIT Media Lab,
link |
Caleb Harbus Lab, where they had
link |
a hydroponic food computer, they called it,
link |
environment that was completely computer controlled,
link |
nutrients, water, light, temperature,
link |
everything is controlled.
link |
Now, what do you do if you can control everything?
link |
Farmers know a lot about how to make plants grow
link |
in their own patch of land.
link |
But if you can control everything, it's too much.
link |
And it turns out that we don't actually
link |
know very much about it.
link |
So we built a system, evolutionary optimization system,
link |
together with a surrogate model of how plants grow
link |
and let this system explore recipes on its own.
link |
And initially, we were focusing on light,
link |
how strong, what wavelengths, how long the light was on.
link |
And we put some boundaries which we thought were reasonable.
link |
For instance, that there was at least six hours of darkness,
link |
like night, because that's what we have in the world.
link |
And very quickly, the system, evolution,
link |
pushed all the recipes to that limit.
link |
We were trying to grow basil.
link |
And we initially had some 200, 300 recipes,
link |
exploration as well as known recipes.
link |
But now we are going beyond that.
link |
And everything was pushed to that limit.
link |
So we look at it and say, well, we can easily just change it.
link |
Let's have it your way.
link |
And it turns out the system discovered
link |
that basil does not need to sleep.
link |
24 hours, lights on, and it will thrive.
link |
It will be bigger, it will be tastier.
link |
And this was a big surprise, not just to us,
link |
but also the biologists in the team
link |
that anticipated that there are some constraints
link |
that are in the world for a reason.
link |
It turns out that evolution did not have the same bias.
link |
And therefore, it discovered something that was creative.
link |
It was surprising, it was useful, and it was new.
link |
That's fascinating to think about the things we think
link |
that are fundamental to living systems on Earth today,
link |
whether they're actually fundamental
link |
or they somehow fit the constraints of the system.
link |
And all we have to do is just remove the constraints.
link |
Do you ever think about,
link |
I don't know how much you know
link |
about brain computer interfaces in your link.
link |
The idea there is our brains are very limited.
link |
And if we just allow, we plug in,
link |
we provide a mechanism for a computer
link |
to speak with the brain.
link |
So you're thereby expanding
link |
the computational power of the brain.
link |
The possibilities there,
link |
from a very high level philosophical perspective,
link |
But I wonder how limitless it is.
link |
Are the constraints we have features
link |
that are fundamental to our intelligence?
link |
Or is this just this weird constraint
link |
in terms of our brain size and skull
link |
and lifespan and senses?
link |
It's just the weird little quirk of evolution.
link |
And if we just open that up,
link |
like add much more senses,
link |
add much more computational power,
link |
the intelligence will expand exponentially.
link |
Do you have a sense about constraints,
link |
the relationship of evolution and computation
link |
to the constraints of the environment?
link |
Well, at first I'd like to comment on that,
link |
like changing the inputs to human brain.
link |
And flexibility of the brain.
link |
I think there's a lot of that.
link |
There are experiments that are done in animals
link |
like Mikangazuru at MIT,
link |
switching the auditory and visual information
link |
and going to the wrong part of the cortex.
link |
And the animal was still able to hear
link |
and perceive the visual environment.
link |
And there are kids that are born with severe disorders
link |
and sometimes they have to remove half of the brain,
link |
like one half, and they still grow up.
link |
They have the functions migrate to the other parts.
link |
There's a lot of flexibility like that.
link |
So I think it's quite possible to hook up the brain
link |
with different kinds of sensors, for instance,
link |
and something that we don't even quite understand
link |
or have today on different kinds of wavelengths
link |
or whatever they are.
link |
And then the brain can learn to make sense of it.
link |
And that I think is this good hope
link |
that these prosthetic devices, for instance, work,
link |
not because we make them so good and so easy to use,
link |
but the brain adapts to them
link |
and can learn to take advantage of them.
link |
And so in that sense, if there's a trouble, a problem,
link |
I think the brain can be used to correct it.
link |
Now going beyond what we have today, can you get smarter?
link |
That's really much harder to do.
link |
Giving the brain more input probably might overwhelm it.
link |
It would have to learn to filter it and focus
link |
and in order to use the information effectively
link |
and augmenting intelligence
link |
with some kind of external devices like that
link |
might be difficult, I think.
link |
But replacing what's lost, I think is quite possible.
link |
Right, so our intuition allows us to sort of imagine
link |
that we can replace what's been lost,
link |
but expansion beyond what we have,
link |
I mean, we're already one of the most,
link |
if not the most intelligent things on this earth, right?
link |
So it's hard to imagine.
link |
But if the brain can hold up with an order of magnitude
link |
greater set of information thrown at it,
link |
if it can do, if it can reason through that.
link |
Part of me, this is the Russian thing, I think,
link |
is I tend to think that the limitations
link |
is where the superpower is,
link |
that immortality and a huge increase in bandwidth
link |
of information by connecting computers with the brain
link |
is not going to produce greater intelligence.
link |
It might produce lesser intelligence.
link |
So I don't know, there's something about the scarcity
link |
being essential to fitness or performance,
link |
but that could be just because we're so limited.
link |
No, exactly, you make do with what you have,
link |
but you don't have to be a genius
link |
but you don't have to pipe it directly to the brain.
link |
I mean, we already have devices like phones
link |
where we can look up information at any point.
link |
And that can make us more productive.
link |
You don't have to argue about, I don't know,
link |
what happened in that baseball game or whatever it is,
link |
because you can look it up right away.
link |
And I think in that sense, we can learn to utilize tools.
link |
And that's what we have been doing for a long, long time.
link |
And we are already, the brain is already drinking
link |
the water, firehose, like vision.
link |
There's way more information in vision
link |
that we actually process.
link |
So brain is already good at identifying what matters.
link |
And that we can switch that from vision
link |
to some other wavelength or some other kind of modality.
link |
But I think that the same processing principles
link |
probably still apply.
link |
But also indeed this ability to have information
link |
more accessible and more relevant,
link |
I think can enhance what we do.
link |
I mean, kids today at school, they learn about DNA.
link |
I mean, things that were discovered
link |
just a couple of years ago.
link |
And it's already common knowledge
link |
and we are building on it.
link |
And we don't see a problem where
link |
there's too much information that we can absorb and learn.
link |
Maybe people become a little bit more narrow
link |
in what they know, they are in one field.
link |
But this information that we have accumulated,
link |
it is passed on and people are picking up on it
link |
and they are building on it.
link |
So it's not like we have reached the point of saturation.
link |
We have still this process that allows us to be selective
link |
and decide what's interesting, I think still works
link |
even with the more information we have today.
link |
Yeah, it's fascinating to think about
link |
like Wikipedia becoming a sensor.
link |
Like, so the fire hose of information from Wikipedia.
link |
So it's like you integrated directly into the brain
link |
to where you're thinking, like you're observing the world
link |
with all of Wikipedia directly piping into your brain.
link |
So like when I see a light,
link |
I immediately have like the history of who invented
link |
electricity, like integrated very quickly into.
link |
So just the way you think about the world
link |
might be very interesting
link |
if you can integrate that kind of information.
link |
What are your thoughts, if I could ask on early steps
link |
on the Neuralink side?
link |
I don't know if you got a chance to see,
link |
but there was a monkey playing pong
link |
through the brain computer interface.
link |
And the dream there is sort of,
link |
you're already replacing the thumbs essentially
link |
that you would use to play video game.
link |
The dream is to be able to increase further
link |
the interface by which you interact with the computer.
link |
Are you impressed by this?
link |
Are you worried about this?
link |
What are your thoughts as a human?
link |
I think it's wonderful.
link |
I think it's great that we could do something
link |
I mean, there are devices that read your EEG for instance,
link |
and humans can learn to control things
link |
using just their thoughts in that sense.
link |
And I don't think it's that different.
link |
I mean, those signals would go to limbs,
link |
they would go to thumbs.
link |
Now the same signals go through a sensor
link |
to some computing system.
link |
It still probably has to be built on human terms,
link |
not to overwhelm them, but utilize what's there
link |
and sense the right kind of patterns
link |
that are easy to generate.
link |
But, oh, that I think is really quite possible
link |
and wonderful and could be very much more efficient.
link |
Is there, so you mentioned surprising
link |
being a characteristic of creativity.
link |
Is there something, you already mentioned a few examples,
link |
but is there something that jumps out at you
link |
as was particularly surprising
link |
from the various evolutionary computation systems
link |
you've worked on, the solutions that were
link |
come up along the way?
link |
Not necessarily the final solutions,
link |
but maybe things that would even discarded.
link |
Is there something that just jumps to mind?
link |
It happens all the time.
link |
I mean, evolution is so creative,
link |
so good at discovering solutions you don't anticipate.
link |
A lot of times they are taking advantage of something
link |
that you didn't think was there,
link |
like a bug in the software, for instance.
link |
A lot of, there's a great paper,
link |
the community put it together
link |
about surprising anecdotes about evolutionary computation.
link |
A lot of them are indeed, in some software environment,
link |
there was a loophole or a bug
link |
and the system utilizes that.
link |
By the way, for people who want to read it,
link |
it's kind of fun to read.
link |
It's called The Surprising Creativity of Digital Evolution,
link |
a collection of anecdotes from the evolutionary computation
link |
and artificial life research communities.
link |
And there's just a bunch of stories
link |
from all the seminal figures in this community.
link |
You have a story in there that released to you,
link |
at least on the Tic Tac Toe memory bomb.
link |
So can you, I guess, describe that situation
link |
if you think that's still?
link |
Yeah, that's a quite a bit smaller scale
link |
than our basic doesn't need to sleep surprise,
link |
but it was actually done by students in my class,
link |
in a neural nets evolution computation class.
link |
There was an assignment.
link |
It was perhaps a final project
link |
where people built game playing AI, it was an AI class.
link |
And this one, and it was for Tic Tac Toe
link |
or five in a row in a large board.
link |
And this one team evolved a neural network
link |
to make these moves.
link |
And they set it up, the evolution.
link |
They didn't really know what would come out,
link |
but it turned out that they did really well.
link |
Evolution actually won the tournament.
link |
And most of the time when it won,
link |
it won because the other teams crashed.
link |
And then when we look at it, like what was going on
link |
was that evolution discovered that if it makes a move
link |
that's really, really far away,
link |
like millions of squares away,
link |
the other teams, the other programs has expanded memory
link |
in order to take that into account
link |
until they run out of memory and crashed.
link |
And then you win a tournament
link |
by crashing all your opponents.
link |
I think that's quite a profound example,
link |
which probably applies to most games,
link |
from even a game theoretic perspective,
link |
that sometimes to win, you don't have to be better
link |
within the rules of the game.
link |
You have to come up with ways to break your opponent's brain,
link |
if it's a human, like not through violence,
link |
but through some hack where the brain just is not,
link |
you're basically, how would you put it?
link |
You're going outside the constraints
link |
of where the brain is able to function.
link |
Expectations of your opponent.
link |
I mean, this was even Kasparov pointed that out
link |
that when Deep Blue was playing against Kasparov,
link |
that it was not playing the same way as Kasparov expected.
link |
And this has to do with not having the same biases.
link |
And that's really one of the strengths of the AI approach.
link |
Can you at a high level say,
link |
what are the basic mechanisms
link |
of evolutionary computation algorithms
link |
that use something that could be called
link |
an evolutionary approach?
link |
Like how does it work?
link |
What are the connections to the,
link |
what are the echoes of the connection to his biological?
link |
A lot of these algorithms really do take motivation
link |
from biology, but they are caricatures.
link |
You try to essentialize it
link |
and take the elements that you believe matter.
link |
So in evolutionary computation,
link |
it is the creation of variation
link |
and then the selection upon that.
link |
So the creation of variation,
link |
you have to have some mechanism
link |
that allow you to create new individuals
link |
that are very different from what you already have.
link |
That's the creativity part.
link |
And then you have to have some way of measuring
link |
how well they are doing and using that measure to select
link |
who goes to the next generation and you continue.
link |
So first you also, you have to have
link |
some kind of digital representation of an individual
link |
that can be then modified.
link |
So I guess humans in biological systems
link |
have DNA and all those kinds of things.
link |
And so you have to have similar kind of encodings
link |
in a computer program.
link |
Yes, and that is a big question.
link |
How do you encode these individuals?
link |
So there's a genotype, which is that encoding
link |
and then a decoding mechanism gives you the phenotype,
link |
which is the actual individual that then performs the task
link |
and in an environment can be evaluated how good it is.
link |
So even that mapping is a big question
link |
and how do you do it?
link |
But typically the representations are,
link |
either they are strings of numbers
link |
or they are some kind of trees.
link |
Those are something that we know very well
link |
in computer science and we try to do that.
link |
But they, and DNA in some sense is also a sequence
link |
and it's a string.
link |
So it's not that far from it,
link |
but DNA also has many other aspects
link |
that we don't take into account necessarily
link |
like there's folding and interactions
link |
that are other than just the sequence itself.
link |
And lots of that is not yet captured
link |
and we don't know whether they are really crucial.
link |
Evolution, biological evolution has produced
link |
wonderful things, but if you look at them,
link |
it's not necessarily the case that every piece
link |
is irreplaceable and essential.
link |
There's a lot of baggage because you have to construct it
link |
and it has to go through various stages
link |
and we still have appendix and we have tail bones
link |
and things like that that are not really that useful.
link |
If you try to explain them now,
link |
it would make no sense, very hard.
link |
But if you think of us as productive evolution,
link |
you can see where they came from.
link |
They were useful at one point perhaps
link |
and no longer are, but they're still there.
link |
So that process is complex
link |
and your representation should support it.
link |
And that is quite difficult if we are limited
link |
with strings or trees,
link |
and then we are pretty much limited
link |
what can be constructed.
link |
And one thing that we are still missing
link |
in evolutionary computation in particular
link |
is what we saw in biology, major transitions.
link |
So that you go from, for instance,
link |
single cell to multi cell organisms
link |
and eventually societies.
link |
There are transitions of level of selection
link |
and level of what a unit is.
link |
And that's something we haven't captured
link |
in evolutionary computation yet.
link |
Does that require a dramatic expansion
link |
of the representation?
link |
Is that what that is?
link |
Most likely it does, but it's quite,
link |
we don't even understand it in biology very well
link |
where it's coming from.
link |
So it would be really good to look at major transitions
link |
in biology, try to characterize them
link |
a little bit more in detail, what the processes are.
link |
How does a, so like a unit, a cell is no longer
link |
It's evaluated as part of a community,
link |
a multi cell organism.
link |
Even though it could reproduce, now it can't alone.
link |
It has to have that environment.
link |
So there's a push to another level, at least a selection.
link |
And how do you make that jump to the next level?
link |
Yes, how do you make the jump?
link |
As part of the algorithm.
link |
So we haven't really seen that in computation yet.
link |
And there are certainly attempts to have open ended evolution.
link |
Things that could add more complexity
link |
and start selecting at a higher level.
link |
But it is still not quite the same
link |
as going from single to multi to society,
link |
for instance, in biology.
link |
So there essentially would be,
link |
as opposed to having one agent,
link |
those agent all of a sudden spontaneously decide
link |
to then be together.
link |
And then your entire system would then be treating them
link |
Something like that.
link |
Some kind of weird merger building.
link |
But also, so you mentioned,
link |
I think you mentioned selection.
link |
So basically there's an agent and they don't get to live on
link |
if they don't do well.
link |
So there's some kind of measure of what doing well is
link |
And does mutation come into play at all in the process
link |
and what in the world does it serve?
link |
Yeah, so, and again, back to what the computational
link |
mechanisms of evolution computation are.
link |
So the way to create variation,
link |
you can take multiple individuals, two usually,
link |
but you could do more.
link |
And you exchange the parts of the representation.
link |
You do some kind of recombination.
link |
Could be crossover, for instance.
link |
In biology, you do have DNA strings that are cut
link |
and put together again.
link |
We could do something like that.
link |
And it seems to be that in biology, the crossover
link |
is really the workhorse in biological evolution.
link |
In computation, we tend to rely more on mutation.
link |
And that is making random changes
link |
into parts of the chromosome.
link |
You can try to be intelligent and target certain areas
link |
of it and make the mutations also follow some principle.
link |
Like you collect statistics of performance and correlations
link |
and try to make mutations you believe
link |
are going to be helpful.
link |
That's where evolution computation has moved
link |
in the last 20 years.
link |
I mean, evolution computation has been around for 50 years,
link |
but a lot of the recent...
link |
Success comes from mutation.
link |
Yes, comes from using statistics.
link |
It's like the rest of machine learning based on statistics.
link |
We use similar tools to guide evolution computation.
link |
And in that sense, it has diverged a bit
link |
from biological evolution.
link |
And that's one of the things I think we could look at again,
link |
having a weaker selection, more crossover,
link |
large populations, more time,
link |
and maybe a different kind of creativity
link |
would come out of it.
link |
We are very impatient in evolution computation today.
link |
We want answers right now, right, quickly.
link |
And if somebody doesn't perform, kill it.
link |
And biological evolution doesn't work quite that way.
link |
And it's more patient.
link |
Yes, much more patient.
link |
So I guess we need to add some kind of mating,
link |
some kind of like dating mechanisms,
link |
like marriage maybe in there.
link |
So into our algorithms to improve the combination
link |
as opposed to all mutation doing all of the work.
link |
Yeah, and many ways of being successful.
link |
Usually in evolution computation, we have one goal,
link |
play this game really well compared to others.
link |
But in biology, there are many ways of being successful.
link |
You can build niches.
link |
You can be stronger, faster, larger, or smarter,
link |
or eat this or eat that.
link |
So there are many ways to solve the same problem of survival.
link |
And that then breeds creativity.
link |
And it allows more exploration.
link |
And eventually you get solutions
link |
that are perhaps more creative
link |
rather than trying to go from initial population directly
link |
or more or less directly to your maximum fitness,
link |
which you measure as just one metric.
link |
So in a broad sense, before we talk about neuroevolution,
link |
do you see evolutionary computation
link |
as more effective than deep learning in a certain context?
link |
Machine learning, broadly speaking.
link |
Maybe even supervised machine learning.
link |
I don't know if you want to draw any kind of lines
link |
and distinctions and borders
link |
where they rub up against each other kind of thing,
link |
where one is more effective than the other
link |
in the current state of things.
link |
Yes, of course, they are very different
link |
and they address different kinds of problems.
link |
And the deep learning has been really successful
link |
in domains where we have a lot of data.
link |
And that means not just data about situations,
link |
but also what the right answers were.
link |
So labeled examples, or they might be predictions,
link |
maybe weather prediction where the data itself becomes labels.
link |
What happened, what the weather was today
link |
and what it will be tomorrow.
link |
So they are very effective deep learning methods
link |
on that kind of tasks.
link |
But there are other kinds of tasks
link |
where we don't really know what the right answer is.
link |
Game playing, for instance,
link |
but many robotics tasks and actions in the world,
link |
decision making and actual practical applications,
link |
like treatments and healthcare
link |
or investment in stock market.
link |
Many tasks are like that.
link |
We don't know and we'll never know
link |
what the optimal answers were.
link |
And there you need different kinds of approach.
link |
Reinforcement learning is one of those.
link |
Reinforcement learning comes from biology as well.
link |
Agents learn during their lifetime.
link |
They eat berries and sometimes they get sick
link |
and then they don't and get stronger.
link |
And then that's how you learn.
link |
And evolution is also a mechanism like that
link |
at a different timescale because you have a population,
link |
not an individual during his lifetime,
link |
but an entire population as a whole
link |
can discover what works.
link |
And there you can afford individuals that don't work out.
link |
They will, you know, everybody dies
link |
and you have a next generation
link |
and they will be better than the previous one.
link |
So that's the big difference between these methods.
link |
They apply to different kinds of problems.
link |
And in particular, there's often a comparison
link |
that's kind of interesting and important
link |
between reinforcement learning and evolutionary computation.
link |
And initially, reinforcement learning
link |
was about individual learning during their lifetime.
link |
And evolution is more engineering.
link |
You don't care about the lifetime.
link |
You don't care about all the individuals that are tested.
link |
You only care about the final result.
link |
The last one, the best candidate that evolution produced.
link |
In that sense, they also apply to different kinds of problems.
link |
And now that boundary is starting to blur a bit.
link |
You can use evolution as an online method
link |
and reinforcement learning to create engineering solutions,
link |
but that's still roughly the distinction.
link |
And from the point of view of what algorithm you wanna use,
link |
if you have something where there is a cost for every trial,
link |
reinforcement learning might be your choice.
link |
Now, if you have a domain
link |
where you can use a surrogate perhaps,
link |
so you don't have much of a cost for trial,
link |
and you want to have surprises,
link |
you want to explore more broadly,
link |
then this population based method is perhaps a better choice
link |
because you can try things out that you wouldn't afford
link |
when you're doing reinforcement learning.
link |
There's very few things as entertaining
link |
as watching either evolutionary computation
link |
or reinforcement learning teaching a simulated robot to walk.
link |
Maybe there's a higher level question
link |
that could be asked here,
link |
but do you find this whole space of applications
link |
in the robotics interesting for evolution computation?
link |
Yeah, yeah, very much.
link |
And indeed, there are fascinating videos of that.
link |
And that's actually one of the examples
link |
where you can contrast the difference.
link |
Between reinforcement learning and evolution.
link |
Yes, so if you have a reinforcement learning agent,
link |
it tries to be conservative
link |
because it wants to walk as long as possible and be stable.
link |
But if you have evolutionary computation,
link |
it can afford these agents that go haywire.
link |
They fall flat on their face and they could take a step
link |
and then they jump and then again fall flat.
link |
And eventually what comes out of that
link |
is something like a falling that's controlled.
link |
You take another step and another step
link |
and you no longer fall.
link |
Instead you run, you go fast.
link |
So that's a way of discovering something
link |
that's hard to discover step by step incrementally.
link |
Because you can afford these evolutionist dead ends,
link |
although they are not entirely dead ends
link |
in the sense that they can serve as stepping stones.
link |
When you take two of those, put them together,
link |
you get something that works even better.
link |
And that is a great example of this kind of discovery.
link |
Yeah, learning to walk is fascinating.
link |
I talked quite a bit to Russ Tedrick who's at MIT.
link |
There's a community of folks
link |
who just roboticists who love the elegance
link |
and beauty of movement.
link |
And walking bipedal robotics is beautiful,
link |
but also exceptionally dangerous
link |
in the sense that like you're constantly falling essentially
link |
if you want to do elegant movement.
link |
And the discovery of that is,
link |
I mean, it's such a good example
link |
of that the discovery of a good solution
link |
sometimes requires a leap of faith and patience
link |
and all those kinds of things.
link |
I wonder what other spaces
link |
where you have to discover those kinds of things in.
link |
Yeah, another interesting direction
link |
is learning for virtual creatures, learning to walk.
link |
We did a study in simulation, obviously,
link |
that you create those creatures,
link |
not just their controller, but also their body.
link |
So you have cylinders, you have muscles,
link |
you have joints and sensors,
link |
and you're creating creatures that look quite different.
link |
Some of them have multiple legs.
link |
Some of them have no legs at all.
link |
And then the goal was to get them to move, to walk, to run.
link |
And what was interesting is that
link |
when you evolve the controller together with the body,
link |
you get movements that look natural
link |
because they're optimized for that physical setup.
link |
And these creatures, you start believing them
link |
that they're alive because they walk in a way
link |
that you would expect somebody
link |
with that kind of a setup to walk.
link |
Yeah, there's something subjective also about that, right?
link |
I've been thinking a lot about that,
link |
especially in the human robot interaction context.
link |
You know, I mentioned Spot, the Boston Dynamics robot.
link |
There is something about human robot communication.
link |
Let's say, let's put it in another context,
link |
something about human and dog context,
link |
like a living dog,
link |
where there's a dance of communication.
link |
First of all, the eyes, you both look at the same thing
link |
and the dogs communicate with their eyes as well.
link |
Like if you're a human,
link |
if you and a dog want to deal with a particular object,
link |
you will look at the person,
link |
the dog will look at you and then look at the object
link |
and look back at you, all those kinds of things.
link |
But there's also just the elegance of movement.
link |
I mean, there's the, of course, the tail
link |
and all those kinds of mechanisms of communication
link |
and it all seems natural and often joyful.
link |
And for robots to communicate that,
link |
it's really difficult how to figure that out
link |
because it's almost seems impossible to hard code in.
link |
You can hard code it for demo purpose or something like that,
link |
but it's essentially choreographed.
link |
Like if you watch some of the Boston Dynamics videos
link |
where they're dancing,
link |
all of that is choreographed by human beings.
link |
But to learn how to, with your movement,
link |
demonstrate a naturalness and elegance, that's fascinating.
link |
Of course, in the physical space,
link |
that's very difficult to do to learn the kind of scale
link |
that you're referring to,
link |
but the hope is that you could do that in simulation
link |
and then transfer it into the physical space
link |
if you're able to model the robot sufficiently naturally.
link |
Yeah, and sometimes I think that that requires
link |
a theory of mind on the side of the robot
link |
that they understand what you're doing
link |
because they themselves are doing something similar.
link |
And that's a big question too.
link |
We talked about intelligence in general
link |
and the social aspect of intelligence.
link |
And I think that's what is required
link |
that we humans understand other humans
link |
because we assume that they are similar to us.
link |
We have one simulation we did a while ago.
link |
Ken Stanley did that.
link |
Two robots that were competing simulation, like I said,
link |
they were foraging for food to gain energy.
link |
And then when they were really strong,
link |
they would bounce into the other robot
link |
and win if they were stronger.
link |
And we watched evolution discover
link |
more and more complex behaviors.
link |
They first went to the nearest food
link |
and then they started to plot a trajectory
link |
so they get more, but then they started to pay attention
link |
what the other robot was doing.
link |
And in the end, there was a behavior
link |
where one of the robots, the most sophisticated one,
link |
sensed where the food pieces were
link |
and identified that the other robot
link |
was close to two of a very far distance
link |
and there was one more food nearby.
link |
So it faked, now I'm using anthropomorphizing terms,
link |
but it made a move towards those other pieces
link |
in order for the other robot to actually go and get them
link |
because it knew that the last remaining piece of food
link |
was close and the other robot would have to travel
link |
a long way, lose its energy
link |
and then lose the whole competition.
link |
So there was like emergence of something
link |
like a theory of mind,
link |
knowing what the other robot would do,
link |
to guide it towards bad behavior in order to win.
link |
So we can get things like that happen in simulation as well.
link |
But that's a complete natural emergence
link |
of a theory of mind.
link |
But I feel like if you add a little bit of a place
link |
for a theory of mind to emerge like easier,
link |
then you can go really far.
link |
I mean, some of these things with evolution, you know,
link |
you add a little bit of design in there, it'll really help.
link |
And I tend to think that a very simple theory of mind
link |
will go a really long way for cooperation between agents
link |
and certainly for human robot interaction.
link |
Like it doesn't have to be super complicated.
link |
I've gotten a chance in the autonomous vehicle space
link |
to watch vehicles interact with pedestrians
link |
or pedestrians interacting with vehicles in general.
link |
I mean, you would think that there's a very complicated
link |
theory of mind thing going on, but I have a sense,
link |
it's not well understood yet,
link |
but I have a sense it's pretty dumb.
link |
Like it's pretty simple.
link |
There's a social contract there between humans,
link |
a human driver and a human crossing the road
link |
where the human crossing the road trusts
link |
that the human in the car is not going to murder them.
link |
And there's something about, again,
link |
back to that mortality thing.
link |
There's some dance of ethics and morality that's built in,
link |
that you're mapping your own morality
link |
onto the person in the car.
link |
And even if they're driving at a speed where you think
link |
if they don't stop, they're going to kill you,
link |
you trust that if you step in front of them,
link |
they're going to hit the brakes.
link |
And there's that weird dance that we do
link |
that I think is a pretty simple model,
link |
but of course it's very difficult to introspect what it is.
link |
And autonomous robots in the human robot interaction
link |
context have to build that.
link |
Current robots are much less than what you're describing.
link |
They're currently just afraid of everything.
link |
They're more, they're not the kind that fall
link |
and discover how to run.
link |
They're more like, please don't touch anything.
link |
Don't hurt anything.
link |
Stay as far away from humans as possible.
link |
Treat humans as ballistic objects that you can't,
link |
that you do with a large spatial envelope,
link |
make sure you do not collide with.
link |
That's how, like you mentioned,
link |
Elon Musk thinks about autonomous vehicles.
link |
I tend to think autonomous vehicles need to have
link |
a beautiful dance between human and machine,
link |
where it's not just the collision avoidance problem,
link |
but a weird dance.
link |
Yeah, I think these systems need to be able to predict
link |
what will happen, what the other agent is going to do,
link |
and then have a structure of what the goals are
link |
and whether those predictions actually meet the goals.
link |
And you can go probably pretty far
link |
with that relatively simple setup already,
link |
but to call it a theory of mind, I don't think you need to.
link |
I mean, it doesn't matter whether the pedestrian
link |
has a mind, it's an object,
link |
and we can predict what we will do.
link |
And then we can predict what the states will be
link |
in the future and whether they are desirable states.
link |
Stay away from those that are undesirable
link |
and go towards those that are desirable.
link |
So it's a relatively simple functional approach to that.
link |
Where do we really need the theory of mind?
link |
Maybe when you start interacting
link |
and you're trying to get the other agent to do something
link |
and jointly, so that you can jointly,
link |
collaboratively achieve something,
link |
then it becomes more complex.
link |
Well, I mean, even with the pedestrians,
link |
you have to have a sense of where their attention,
link |
actual attention in terms of their gaze is,
link |
but also there's this vision science,
link |
people talk about this all the time.
link |
Just because I'm looking at it
link |
doesn't mean I'm paying attention to it.
link |
So figuring out what is the person looking at?
link |
What is the sensory information they've taken in?
link |
And the theory of mind piece comes in is
link |
what are they actually attending to cognitively?
link |
And also what are they thinking about?
link |
Like what is the computation they're performing?
link |
And you have probably maybe a few options
link |
for the pedestrian crossing.
link |
It doesn't have to be,
link |
it's like a variable with a few discrete states,
link |
but you have to have a good estimation
link |
which of the states that brain is in
link |
for the pedestrian case.
link |
And the same is for attending with a robot.
link |
If you're collaborating to pick up an object,
link |
you have to figure out is the human,
link |
like there's a few discrete states
link |
that the human could be in.
link |
You have to predict that by observing the human.
link |
And that seems like a machine learning problem
link |
to figure out what's the human up to.
link |
It's not as simple as sort of planning
link |
just because they move their arm
link |
means the arm will continue moving in this direction.
link |
You have to really have a model
link |
of what they're thinking about
link |
and what's the motivation behind the movement of the arm.
link |
Here we are talking about relatively simple physical actions,
link |
but you can take that the higher levels also
link |
like to predict what the people are going to do,
link |
you need to know what their goals are.
link |
What are they trying to, are they exercising?
link |
Are they just starting to get somewhere?
link |
But even higher level, I mean,
link |
you are predicting what people will do in their career,
link |
what their life themes are.
link |
Do they want to be famous, rich, or do good?
link |
And that takes a lot more information,
link |
but it allows you to then predict their actions,
link |
what choices they might make.
link |
So how does evolution and computation apply
link |
to the world of neural networks?
link |
I've seen quite a bit of work from you and others
link |
in the world of neural evolution.
link |
So maybe first, can you say, what is this field?
link |
Yeah, neural evolution is a combination of neural networks
link |
and evolution computation in many different forms,
link |
but the early versions were simply using evolution
link |
as a way to construct a neural network
link |
instead of say, stochastic gradient descent
link |
or backpropagation.
link |
Because evolution can evolve these parameters,
link |
weight values in a neural network,
link |
just like any other string of numbers, you can do that.
link |
And that's useful because some cases you don't have
link |
those targets that you need to backpropagate from.
link |
And it might be an agent that's running a maze
link |
or a robot playing a game or something.
link |
You don't, again, you don't know what the right answers are,
link |
you don't have backprop,
link |
but this way you can still evolve a neural net.
link |
And neural networks are really good at these tasks,
link |
because they recognize patterns
link |
and they generalize, interpolate between known situations.
link |
So you want to have a neural network in such a task,
link |
even if you don't have a supervised targets.
link |
So that's a reason and that's a solution.
link |
And also more recently,
link |
now when we have all this deep learning literature,
link |
it turns out that we can use evolution
link |
to optimize many aspects of those designs.
link |
The deep learning architectures have become so complex
link |
that there's little hope for us little humans
link |
to understand their complexity
link |
and what actually makes a good design.
link |
And now we can use evolution to give that design for you.
link |
And it might mean optimizing hyperparameters,
link |
like the depth of layers and so on,
link |
or the topology of the network,
link |
how many layers, how they're connected,
link |
but also other aspects like what activation functions
link |
you use where in the network during the learning process,
link |
or what loss function you use,
link |
you could generalize that.
link |
You could generate that, even data augmentation,
link |
all the different aspects of the design
link |
of deep learning experiments could be optimized that way.
link |
So that's an interaction between two mechanisms.
link |
But there's also, when we get more into cognitive science
link |
and the topics that we've been talking about,
link |
you could have learning mechanisms
link |
at two level timescales.
link |
So you do have an evolution
link |
that gives you baby neural networks
link |
that then learn during their lifetime.
link |
And you have this interaction of two timescales.
link |
And I think that can potentially be really powerful.
link |
Now, in biology, we are not born with all our faculties.
link |
We have to learn, we have a developmental period.
link |
In humans, it's really long and most animals have something.
link |
And probably the reason is that evolution of DNA
link |
is not detailed enough or plentiful enough to describe them.
link |
We can describe how to set the brain up,
link |
but we can, evolution can decide on a starting point
link |
and then have a learning algorithm
link |
that will construct the final product.
link |
And this interaction of intelligent, well,
link |
evolution that has produced a good starting point
link |
for the specific purpose of learning from it
link |
with the interaction with the environment,
link |
that can be a really powerful mechanism
link |
for constructing brains and constructing behaviors.
link |
I like how you walk back from intelligence.
link |
So optimize starting point, maybe.
link |
Yeah, okay, there's a lot of fascinating things to ask here.
link |
And this is basically this dance between neural networks
link |
and evolution and computation
link |
could go into the category of automated machine learning
link |
to where you're optimizing,
link |
whether it's hyperparameters of the topology
link |
or hyperparameters taken broadly.
link |
But the topology thing is really interesting.
link |
I mean, that's not really done that effectively
link |
or throughout the history of machine learning
link |
has not been done.
link |
Usually there's a fixed architecture.
link |
Maybe there's a few components you're playing with,
link |
but to grow a neural network, essentially,
link |
the way you grow an organism is really fascinating space.
link |
How hard is it, do you think, to grow a neural network?
link |
And maybe what kind of neural networks
link |
are more amenable to this kind of idea than others?
link |
I've seen quite a bit of work on recurrent neural networks.
link |
Is there some architectures that are friendlier than others?
link |
And is this just a fun, small scale set of experiments
link |
or do you have hope that we can be able to grow
link |
powerful neural networks?
link |
And most of the work up to now
link |
is taking architectures that already exist
link |
that humans have designed and try to optimize them further.
link |
And you can totally do that.
link |
A few years ago, we did an experiment.
link |
We took a winner of the image captioning competition
link |
and the architecture and just broke it into pieces
link |
and took the pieces.
link |
And that was our search base.
link |
See if you can do better.
link |
And we indeed could, 15% better performance
link |
by just searching around the network design
link |
that humans had come up with,
link |
Oreo vinyls and others.
link |
So, but that's starting from a point
link |
that humans have produced,
link |
but we could do something more general.
link |
It doesn't have to be that kind of network.
link |
The hard part is, there are a couple of challenges.
link |
One of them is to define the search base.
link |
What are your elements and how you put them together.
link |
And the space is just really, really big.
link |
So you have to somehow constrain it
link |
and have some hunch what will work
link |
because otherwise everything is possible.
link |
And another challenge is that in order to evaluate
link |
how good your design is, you have to train it.
link |
I mean, you have to actually try it out.
link |
And that's currently very expensive, right?
link |
I mean, deep learning networks may take days to train
link |
while imagine you having a population of a hundred
link |
and have to run it for a hundred generations.
link |
It's not yet quite feasible computationally.
link |
It will be, but also there's a large carbon footprint
link |
I mean, we are using a lot of computation for doing it.
link |
So intelligent methods and intelligent,
link |
I mean, we have to do some science
link |
in order to figure out what the right representations are
link |
and right operators are, and how do we evaluate them
link |
without having to fully train them.
link |
And that is where the current research is
link |
and we're making progress on all those fronts.
link |
So yes, there are certain architectures
link |
that are more amenable to that approach,
link |
but also I think we can create our own architecture
link |
and all representations that are even better at that.
link |
And do you think it's possible to do like a tiny baby network
link |
that grows into something that can do state of the art
link |
on like even the simple data set like MNIST,
link |
and just like it just grows into a gigantic monster
link |
that's the world's greatest handwriting recognition system?
link |
Yeah, there are approaches like that.
link |
Esteban Real and Cochlear for instance,
link |
I worked on evolving a smaller network
link |
and then systematically expanding it to a larger one.
link |
Your elements are already there and scaling it up
link |
will just give you more power.
link |
So again, evolution gives you that starting point
link |
and then there's a mechanism that gives you the final result
link |
and a very powerful approach.
link |
But you could also simulate the actual growth process.
link |
And like I said before, evolving a starting point
link |
and then evolving or training the network,
link |
there's not that much work that's been done on that yet.
link |
We need some kind of a simulation environment
link |
so the interactions at will,
link |
the supervised environment doesn't really,
link |
it's not as easily usable here.
link |
Sorry, the interaction between neural networks?
link |
Yeah, the neural networks that you're creating,
link |
interacting with the world
link |
and learning from these sequences of interactions,
link |
perhaps communication with others.
link |
We would like to get there,
link |
but just the task of simulating something
link |
is at that level is very hard.
link |
It's very difficult.
link |
I mean, one of the powerful things about evolution
link |
on Earth is the predators and prey emerged.
link |
And like there's just like,
link |
there's bigger fish and smaller fish
link |
and it's fascinating to think
link |
that you could have neural networks competing
link |
against each other in one neural network
link |
being able to destroy another one.
link |
There's like wars of neural networks competing
link |
to solve the MNIST problem, I don't know.
link |
Oh, totally, yeah, yeah, yeah.
link |
And we actually simulated also that prey
link |
and it was interesting what happened there,
link |
Padmini Rajagopalan did this
link |
and Kay Holkamp was a zoologist.
link |
we had simulated hyenas, simulated zebras.
link |
And initially, the hyenas just tried to hunt them
link |
and when they actually stumbled upon the zebra,
link |
they ate it and were happy.
link |
And then the zebras learned to escape
link |
and the hyenas learned to team up.
link |
And actually two of them approached
link |
in different directions.
link |
And now the zebras, their next step,
link |
they generated a behavior where they split
link |
in different directions,
link |
just like actually gazelles do
link |
when they are being hunted.
link |
They confuse the predator
link |
by going in different directions.
link |
That emerged and then more hyenas joined
link |
and kind of circled them.
link |
And then when they circled them,
link |
they could actually herd the zebras together
link |
and eat multiple zebras.
link |
So there was like an arms race of predators and prey.
link |
And they gradually developed more complex behaviors,
link |
some of which we actually do see in nature.
link |
And this kind of coevolution,
link |
that's competitive coevolution,
link |
it's a fascinating topic
link |
because there's a promise or possibility
link |
that you will discover something new
link |
that you don't already know.
link |
You didn't build it in.
link |
It came from this arms race.
link |
It's hard to keep the arms race going.
link |
It's hard to have rich enough simulation
link |
that supports all of these complex behaviors.
link |
But at least for several steps,
link |
we've already seen it in this predator prey scenario, yeah.
link |
First of all, it's fascinating to think about this context
link |
in terms of evolving architectures.
link |
So I've studied Tesla autopilot for a long time.
link |
It's one particular implementation of an AI system
link |
that's operating in the real world.
link |
I find it fascinating because of the scale
link |
at which it's used out in the real world.
link |
And I'm not sure if you're familiar with that system much,
link |
but, you know, Andre Kapathy leads that team
link |
on the machine learning side.
link |
And there's a multitask network, multiheaded network,
link |
where there's a core, but it's trained on particular tasks.
link |
And there's a bunch of different heads
link |
that are trained on that.
link |
Is there some lessons from evolutionary computation
link |
or neuroevolution that could be applied
link |
to this kind of multiheaded beast
link |
that's operating in the real world?
link |
Yes, it's a very good problem for neuroevolution.
link |
And the reason is that when you have multiple tasks,
link |
they support each other.
link |
So let's say you're learning to classify X ray images
link |
to different pathologies.
link |
So you have one task is to classify this disease
link |
and another one, this disease, another one, this one.
link |
And when you're learning from one disease,
link |
that forces certain kinds of internal representations
link |
and embeddings, and they can serve
link |
as a helpful starting point for the other tasks.
link |
So you are combining the wisdom of multiple tasks
link |
into these representations.
link |
And it turns out that you can do better
link |
in each of these tasks
link |
when you are learning simultaneously other tasks
link |
than you would by one task alone.
link |
Which is a fascinating idea in itself, yeah.
link |
Yes, and people do that all the time.
link |
I mean, you use knowledge of domains that you know
link |
in new domains, and certainly neural network can do that.
link |
When neuroevolution comes in is that,
link |
what's the best way to combine these tasks?
link |
Now there's architectural design that allow you to decide
link |
where and how the embeddings,
link |
the internal representations are combined
link |
and how much you combine them.
link |
And there's quite a bit of research on that.
link |
And my team, Elliot Meyerson has worked on that
link |
in particular, like what is a good internal representation
link |
that supports multiple tasks?
link |
And we're getting to understand how that's constructed
link |
and what's in it, so that it is in a space
link |
that supports multiple different heads, like you said.
link |
And that I think is fundamentally
link |
how biological intelligence works as well.
link |
You don't build a representation just for one task.
link |
You try to build something that's general,
link |
not only so that you can do better in one task
link |
or multiple tasks, but also future tasks
link |
and future challenges.
link |
So you learn the structure of the world
link |
and that helps you in all kinds of future challenges.
link |
And so you're trying to design a representation
link |
that will support an arbitrary set of tasks
link |
in a particular sort of class of problem.
link |
Yeah, and also it turns out,
link |
and that's again, a surprise that Elliot found
link |
was that those tasks don't have to be very related.
link |
You know, you can learn to do better vision
link |
by learning language or better language
link |
by learning about DNA structure.
link |
No, somehow the world.
link |
The world rhymes, even if it's very disparate fields.
link |
I mean, on that small topic, let me ask you,
link |
because you've also on the competition neuroscience side,
link |
you worked on both language and vision.
link |
What's the connection between the two?
link |
What's more, maybe there's a bunch of ways to ask this,
link |
but what's more difficult to build
link |
from an engineering perspective
link |
and evolutionary perspective,
link |
the human language system or the human vision system
link |
or the equivalent of in the AI space language and vision,
link |
or is it the best as the multitask idea
link |
that you're speaking to
link |
that they need to be deeply integrated?
link |
Yeah, absolutely the latter.
link |
Learning both at the same time,
link |
I think is a fascinating direction in the future.
link |
So we have data sets where there's visual component
link |
as well as verbal descriptions, for instance,
link |
and that way you can learn a deeper representation,
link |
a more useful representation for both.
link |
But it's still an interesting question
link |
of which one is easier.
link |
I mean, recognizing objects
link |
or even understanding sentences, that's relatively possible,
link |
but where it becomes, where the challenges are
link |
is to understand the world.
link |
Like the visual world, the 3D,
link |
what are the objects doing
link |
and predicting what will happen, the relationships.
link |
That's what makes vision difficult.
link |
And language, obviously it's what is being said,
link |
what the meaning is.
link |
And the meaning doesn't stop at who did what to whom.
link |
There are goals and plans and themes,
link |
and you eventually have to understand
link |
the entire human society and history
link |
in order to understand the sentence very much fully.
link |
There are plenty of examples of those kinds
link |
of short sentences when you bring in
link |
all the world knowledge to understand it.
link |
And that's the big challenge.
link |
Now we are far from that,
link |
but even just bringing in the visual world
link |
together with the sentence will give you already
link |
a lot deeper understanding of what's happening.
link |
And I think that that's where we're going very soon.
link |
I mean, we've had ImageNet for a long time,
link |
and now we have all these text collections,
link |
but having both together and then learning
link |
a semantic understanding of what is happening,
link |
I think that that will be the next step
link |
in the next few years.
link |
Yeah, you're starting to see that
link |
with all the work with Transformers,
link |
was the community, the AI community
link |
starting to dip their toe into this idea
link |
of having language models that are now doing stuff
link |
with images, with vision, and then connecting the two.
link |
I mean, right now it's like these little explorations
link |
we're literally dipping the toe in,
link |
but maybe at some point we'll just dive into the pool
link |
and it'll just be all seen as the same thing.
link |
I do still wonder what's more fundamental,
link |
whether vision is, whether we don't think
link |
about vision correctly.
link |
Maybe the fact, because we're humans
link |
and we see things as beautiful and so on,
link |
and because we have cameras that are taking pixels
link |
as a 2D image, that we don't sufficiently think
link |
about vision as language.
link |
Maybe Chomsky is right all along,
link |
that vision is fundamental to,
link |
sorry, that language is fundamental to everything,
link |
to even cognition, to even consciousness.
link |
The base layer is all language,
link |
not necessarily like English, but some weird
link |
abstract representation, linguistic representation.
link |
Yeah, well, earlier we talked about the social structures
link |
and that may be what's underlying the language,
link |
and that's the more fundamental part,
link |
and then language has been added on top of that.
link |
Language emerges from the social interaction.
link |
Yeah, that's a very good guess.
link |
We are visual animals, though.
link |
A lot of the brain is dedicated to vision,
link |
and also, when we think about various abstract concepts,
link |
we usually reduce that to vision and images,
link |
and that's, you know, we go to a whiteboard,
link |
you draw pictures of very abstract concepts.
link |
So we tend to resort to that quite a bit,
link |
and that's a fundamental representation.
link |
It's probably possible that it predated language even.
link |
I mean, animals, a lot of, they don't talk,
link |
but they certainly do have vision,
link |
and language is interesting development
link |
in from mastication, from eating.
link |
You develop an organ that actually can produce sound
link |
to manipulate them.
link |
Maybe that was an accident.
link |
Maybe that was something that was available
link |
and then allowed us to do the communication,
link |
or maybe it was gestures.
link |
Sign language could have been the original proto language.
link |
We don't quite know, but the language is more fundamental
link |
than the medium in which it's communicated,
link |
and I think that it comes from those representations.
link |
Now, in current world, they are so strongly integrated,
link |
it's really hard to say which one is fundamental.
link |
You look at the brain structures and even visual cortex,
link |
which is supposed to be very much just vision.
link |
Well, if you are thinking of semantic concepts,
link |
you're thinking of language, visual cortex lights up.
link |
It's still useful, even for language computations.
link |
So there are common structures underlying them.
link |
So utilize what you need.
link |
And when you are understanding a scene,
link |
you're understanding relationships.
link |
Well, that's not so far from understanding relationships
link |
between words and concepts.
link |
So I think that that's how they are integrated.
link |
Yeah, and there's dreams, and once we close our eyes,
link |
there's still a world in there somehow operating
link |
and somehow possibly the visual system somehow integrated
link |
I tend to enjoy thinking about aliens
link |
and thinking about the sad thing to me
link |
about extraterrestrial intelligent life,
link |
that if it visited us here on Earth,
link |
or if we came on Mars or maybe another solar system,
link |
another galaxy one day,
link |
that us humans would not be able to detect it
link |
or communicate with it or appreciate,
link |
like it'd be right in front of our nose
link |
and we were too self obsessed to see it.
link |
Not self obsessed, but our tools,
link |
our frameworks of thinking would not detect it.
link |
As a good movie, Arrival and so on,
link |
where Stephen Wolfram and his son,
link |
I think were part of developing this alien language
link |
of how aliens would communicate with humans.
link |
Do you ever think about that kind of stuff
link |
where if humans and aliens would be able to communicate
link |
with each other, like if we met each other at some,
link |
okay, we could do SETI, which is communicating
link |
from across a very big distance,
link |
but also just us, if you did a podcast with an alien,
link |
do you think we'd be able to find a common language
link |
and a common methodology of communication?
link |
I think from a computational perspective,
link |
the way to ask that is you have very fundamentally
link |
different creatures, agents that are created,
link |
would they be able to find a common language?
link |
Yes, I do think about that.
link |
I mean, I think a lot of people who are in computing,
link |
they, and AI in particular, they got into it
link |
because they were fascinated with science fiction
link |
and all of these options.
link |
I mean, Star Trek generated all kinds of devices
link |
that we have now, they envisioned it first
link |
and it's a great motivator to think about things like that.
link |
And I, so one, and again, being a computational scientist
link |
and trying to build intelligent agents,
link |
what I would like to do is have a simulation
link |
where the agents actually evolve communication,
link |
not just communication, we've done that,
link |
people have done that many times,
link |
that they communicate, they signal and so on,
link |
but actually develop a language.
link |
And language means grammar, it means all these
link |
social structures and on top of that,
link |
grammatical structures.
link |
And we do it under various conditions
link |
and actually try to identify what conditions
link |
are necessary for it to come out.
link |
And then we can start asking that kind of questions.
link |
Are those languages that emerge
link |
in those different simulated environments,
link |
are they understandable to us?
link |
Can we somehow make a translation?
link |
We can make it a concrete question.
link |
So machine translation of evolved languages.
link |
And so like languages that evolve come up with,
link |
can we translate, like I have a Google translate
link |
for the evolved languages.
link |
Yes, and if we do that enough,
link |
we have perhaps an idea what an alien language
link |
might be like, the space of where those languages can be.
link |
Because we can set up their environment differently.
link |
It doesn't need to be gravity.
link |
You can have all kinds of, societies can be different.
link |
They may have no predators.
link |
They may have all, everybody's a predator.
link |
All kinds of situations.
link |
And then see what the space possibly is
link |
where those languages are and what the difficulties are.
link |
That'd be really good actually to do that
link |
before the aliens come here.
link |
Yes, it's good practice.
link |
On the similar connection,
link |
you can think of AI systems as aliens.
link |
Is there ways to evolve a communication scheme
link |
for, there's a field you can call it explainable AI,
link |
for AI systems to be able to communicate.
link |
So you evolve a bunch of agents,
link |
but for some of them to be able to talk to you also.
link |
So to evolve a way for agents to be able to communicate
link |
about their world to us humans.
link |
Do you think that there's possible mechanisms
link |
We can certainly try.
link |
And if it's an evolution competition system,
link |
for instance, you reward those solutions
link |
that are actually functional.
link |
That communication makes sense.
link |
It allows us to together again, achieve common goals.
link |
I think that's possible.
link |
But even from that paper that you mentioned,
link |
the anecdotes, it's quite likely also
link |
that the agents learn to lie and fake
link |
and do all kinds of things like that.
link |
I mean, we see that in even very low level,
link |
like bacterial evolution.
link |
There are cheaters.
link |
And who's to say that what they say
link |
is actually what they think.
link |
But that's what I'm saying,
link |
that there would have to be some common goal
link |
so that we can evaluate whether that communication
link |
is at least useful.
link |
They may be saying things just to make us feel good
link |
or get us to do what we want,
link |
but they would not turn them off or something.
link |
But so we would have to understand
link |
their internal representations much better
link |
to really make sure that that translation is critical.
link |
But it can be useful.
link |
And I think it's possible to do that.
link |
There are examples where visualizations
link |
are automatically created
link |
so that we can look into the system
link |
and that language is not that far from it.
link |
I mean, it is a way of communicating and logging
link |
what you're doing in some interpretable way.
link |
I think a fascinating topic, yeah, to do that.
link |
Yeah, you're making me realize
link |
that it's a good scientific question
link |
whether lying is an effective mechanism
link |
for integrating yourself and succeeding
link |
in a social network, in a world that is social.
link |
I tend to believe that honesty and love
link |
are evolutionary advantages in an environment
link |
where there's a network of intelligent agents.
link |
But it's also very possible that dishonesty
link |
and manipulation and even violence,
link |
all those kinds of things might be more beneficial.
link |
That's the old open question about good versus evil.
link |
But I tend to, I mean, I don't know if it's a hopeful,
link |
maybe I'm delusional, but it feels like karma is a thing,
link |
which is like long term, the agents,
link |
they're just kind to others sometimes for no reason
link |
In a society that's not highly constrained on resources.
link |
So like people start getting weird
link |
and evil towards each other and bad
link |
when the resources are very low relative
link |
to the needs of the populace,
link |
especially at the basic level, like survival, shelter,
link |
food, all those kinds of things.
link |
But I tend to believe that once you have
link |
those things established, then, well, not to believe,
link |
I guess I hope that AI systems will be honest.
link |
But it's scary to think about the Turing test,
link |
AI systems that will eventually pass the Turing test
link |
will be ones that are exceptionally good at lying.
link |
That's a terrifying concept.
link |
I mean, I don't know.
link |
First of all, sort of from somebody who studied language
link |
and obviously are not just a world expert in AI,
link |
but somebody who dreams about the future of the field.
link |
Do you hope, do you think there'll be human level
link |
or superhuman level intelligences in the future
link |
that we eventually build?
link |
Well, I definitely hope that we can get there.
link |
One, I think important perspective
link |
is that we are building AI to help us.
link |
That it is a tool like cars or language
link |
or communication, AI will help us be more productive.
link |
And that is always a condition.
link |
It's not something that we build and let run
link |
and it becomes an entity of its own
link |
that doesn't care about us.
link |
Now, of course, really find the future,
link |
maybe that might be possible,
link |
but not in the foreseeable future when we are building it.
link |
And therefore we always in a position of limiting
link |
what it can or cannot do.
link |
And your point about lying is very interesting.
link |
Even in these hyenas societies, for instance,
link |
when a number of these hyenas band together
link |
and they take a risk and steal the kill,
link |
there are always hyenas that hang back
link |
and don't participate in that risky behavior,
link |
but they walk in later and join the party
link |
And there are even some that may be ineffective
link |
and cause others to have harm.
link |
So, and like I said, even bacteria cheat.
link |
And we see it in biology,
link |
there's always some element on opportunity.
link |
If you have a society, I think that is just because
link |
if you have a society,
link |
in order for society to be effective,
link |
you have to have this cooperation
link |
and you have to have trust.
link |
And if you have enough of agents
link |
who are able to trust each other,
link |
you can achieve a lot more.
link |
But if you have trust,
link |
you also have opportunity for cheaters and liars.
link |
And I don't think that's ever gonna go away.
link |
There will be hopefully a minority
link |
so that they don't get in the way.
link |
And we studied in these hyena simulations,
link |
like what the proportion needs to be
link |
before it is no longer functional.
link |
And you can point out that you can tolerate
link |
a few cheaters and a few liars
link |
and the society can still function.
link |
And that's probably going to happen
link |
when we build these systems at Autonomously Learn.
link |
The really successful ones are honest
link |
because that's the best way of getting things done.
link |
But there probably are also intelligent agents
link |
that find that they can achieve their goals
link |
by bending the rules or cheating.
link |
So that could be a huge benefit
link |
as opposed to having fixed AI systems.
link |
Say we build an AGI system and deploying millions of them,
link |
it'd be that are exactly the same.
link |
There might be a huge benefit to introducing
link |
sort of from like an evolution computation perspective,
link |
a lot of variation.
link |
Sort of like diversity in all its forms is beneficial
link |
even if some people are assholes
link |
or some robots are assholes.
link |
So like it's beneficial to have that
link |
because you can't always a priori know
link |
what's good, what's bad.
link |
But that's a fascinating.
link |
Diversity is the bread and butter.
link |
I mean, if you're running an evolution,
link |
you see diversity is the one fundamental thing
link |
And absolutely, also, it's not always good diversity.
link |
It may be something that can be destructive.
link |
We had in these hyenas simulations,
link |
we have hyenas that just are suicidal.
link |
They just run and get killed.
link |
But they form the basis of those
link |
who actually are really fast,
link |
but stop before they get killed
link |
and eventually turn into this mob.
link |
So there might be something useful there
link |
if it's recombined with something else.
link |
So I think that as long as we can tolerate some of that,
link |
it may turn into something better.
link |
You may change the rules
link |
because it's so much more efficient to do something
link |
that was actually against the rules before.
link |
And we've seen society change over time
link |
quite a bit along those lines.
link |
That there were rules in society
link |
that we don't believe are fair anymore,
link |
even though they were considered proper behavior before.
link |
So things are changing.
link |
And I think that in that sense,
link |
I think it's a good idea to be able to tolerate
link |
some of that cheating
link |
because eventually we might turn into something better.
link |
So yeah, I think this is a message
link |
to the trolls and the assholes of the internet
link |
that you too have a beautiful purpose
link |
in this human ecosystem.
link |
So I appreciate you very much.
link |
In moderate quantities, yeah.
link |
In moderate quantities.
link |
So there's a whole field of artificial life.
link |
I don't know if you're connected to this field,
link |
if you pay attention.
link |
Is, do you think about this kind of thing?
link |
Is there impressive demonstration to you
link |
of artificial life?
link |
Do you think of the agency you work with
link |
in the evolutionary computation perspective as life?
link |
And where do you think this is headed?
link |
Like, is there interesting systems
link |
that we'll be creating more and more
link |
that make us redefine, maybe rethink
link |
about the nature of life?
link |
Different levels of definition and goals there.
link |
I mean, at some level, artificial life
link |
can be considered multiagent systems
link |
that build a society that again, achieves a goal.
link |
And it might be robots that go into a building
link |
and clean it up or after an earthquake or something.
link |
You can think of that as an artificial life problem
link |
Or you can really think of it, artificial life,
link |
as a simulation of life and a tool to understand
link |
what life is and how life evolved on earth.
link |
And like I said, in artificial life conference,
link |
there are branches of that conference sessions
link |
of people who really worry about molecular designs
link |
and the start of life, like I said,
link |
primordial soup where eventually
link |
you get something self replicating.
link |
And they're really trying to build that.
link |
So it's a whole range of topics.
link |
And I think that artificial life is a great tool
link |
to understand life.
link |
And there are questions like sustainability,
link |
species, we're losing species.
link |
Is there a tipping point?
link |
And where are we going?
link |
I mean, like the hyena evolution,
link |
we may have understood that there's a pivotal point
link |
in their evolution.
link |
They discovered cooperation and coordination.
link |
Artificial life simulations can identify that
link |
and maybe encourage things like that.
link |
And also societies can be seen as a form of life itself.
link |
I mean, we're not talking about biological evolution,
link |
evolution of societies.
link |
Maybe some of the same phenomena emerge in that domain
link |
and having artificial life simulations and understanding
link |
could help us build better societies.
link |
Yeah, and thinking from a meme perspective
link |
of from Richard Dawkins,
link |
that maybe the organisms, ideas of the organisms,
link |
not the humans in these societies that from,
link |
it's almost like reframing what is exactly evolving.
link |
Maybe the interesting,
link |
the humans aren't the interesting thing
link |
as the contents of our minds is the interesting thing.
link |
And that's what's multiplying.
link |
And that's actually multiplying and evolving
link |
in a much faster timescale.
link |
And that maybe has more power on the trajectory
link |
of life on earth than does biological evolution
link |
is the evolution of these ideas.
link |
Yes, and it's fascinating, like I said before,
link |
that we can keep up somehow biologically.
link |
We evolved to a point where we can keep up
link |
with this meme evolution, literature, internet.
link |
We understand DNA and we understand fundamental particles.
link |
We didn't start that way a thousand years ago.
link |
And we haven't evolved biologically very much,
link |
but somehow our minds are able to extend.
link |
And therefore AI can be seen also as one such step
link |
that we created and it's our tool.
link |
And it's part of that meme evolution that we created,
link |
even if our biological evolution does not progress as fast.
link |
And us humans might only be able to understand so much.
link |
We're keeping up so far,
link |
or we think we're keeping up so far,
link |
but we might need AI systems to understand.
link |
Maybe like the physics of the universe is operating,
link |
look at strength theory.
link |
Maybe it's operating in much higher dimensions.
link |
Maybe we're totally, because of our cognitive limitations,
link |
are not able to truly internalize the way this world works.
link |
And so we're running up against the limitation
link |
And we have to create these next level organisms
link |
like AI systems that would be able to understand much deeper,
link |
like really understand what it means to live
link |
in a multi dimensional world
link |
that's outside of the four dimensions,
link |
the three of space and one of time.
link |
Translation, and generally we can deal with the world,
link |
even if you don't understand all the details,
link |
we can use computers, even though we don't,
link |
most of us don't know all the structure
link |
that's underneath or drive a car.
link |
I mean, there are many components,
link |
especially new cars that you don't quite fully know,
link |
but you have the interface, you have an abstraction of it
link |
that allows you to operate it and utilize it.
link |
And I think that that's perfectly adequate
link |
and we can build on it.
link |
And AI can play a similar role.
link |
I have to ask about beautiful artificial life systems
link |
or evolutionary computation systems.
link |
Cellular automata to me,
link |
I remember it was a game changer for me early on in life
link |
when I saw Conway's Game of Life
link |
who recently passed away, unfortunately.
link |
And it's beautiful
link |
how much complexity can emerge from such simple rules.
link |
I just don't, somehow that simplicity
link |
is such a powerful illustration
link |
and also humbling because it feels like I personally,
link |
from my perspective,
link |
understand almost nothing about this world
link |
because like my intuition fails completely
link |
how complexity can emerge from such simplicity.
link |
Like my intuition fails, I think,
link |
is the biggest problem I have.
link |
Do you find systems like that beautiful?
link |
Is there, do you think about cellular automata?
link |
Because cellular automata don't really have,
link |
and many other artificial life systems
link |
don't necessarily have an objective.
link |
Maybe that's a wrong way to say it.
link |
It's almost like it's just evolving and creating.
link |
And there's not even a good definition
link |
of what it means to create something complex
link |
and interesting and surprising,
link |
all those words that you said.
link |
Is there some of those systems that you find beautiful?
link |
And similarly, evolution does not have a goal.
link |
It is responding to current situation
link |
and survival then creates more complexity
link |
and therefore we have something that we perceive as progress
link |
but that's not what evolution is inherently set to do.
link |
And yeah, that's really fascinating
link |
how a simple set of rules or simple mappings can,
link |
how from such simple mappings, complexity can emerge.
link |
So it's a question of emergence and self organization.
link |
And the game of life is one of the simplest ones
link |
and very visual and therefore it drives home the point
link |
that it's possible that nonlinear interactions
link |
and this kind of complexity can emerge from them.
link |
And biology and evolution is along the same lines.
link |
We have simple representations.
link |
DNA, if you really think of it, it's not that complex.
link |
It's a long sequence of them, there's lots of them
link |
but it's a very simple representation.
link |
And similarly with evolutionary computation,
link |
whatever string or tree representation we have
link |
and the operations, the amount of code that's required
link |
to manipulate those, it's really, really little.
link |
And of course, game of life even less.
link |
So how complexity emerges from such simple principles,
link |
that's absolutely fascinating.
link |
The challenge is to be able to control it
link |
and guide it and direct it so that it becomes useful.
link |
And like game of life is fascinating to look at
link |
and evolution, all the forms that come out is fascinating
link |
but can we actually make it useful for us?
link |
And efficient because if you actually think about
link |
each of the cells in the game of life as a living organism,
link |
there's a lot of death that has to happen
link |
to create anything interesting.
link |
And so I guess the question is for us humans
link |
that are mortal and then life ends quickly,
link |
we wanna kinda hurry up and make sure we take evolution,
link |
the trajectory that is a little bit more efficient
link |
than the alternatives.
link |
And that touches upon something we talked about earlier
link |
that evolution competition is very impatient.
link |
We have a goal, we want it right away
link |
whereas this biology has a lot of time and deep time
link |
and weak pressure and large populations.
link |
One great example of this is the novelty search.
link |
So evolutionary computation
link |
where you don't actually specify a fitness goal,
link |
something that is your actual thing that you want
link |
but you just reward solutions that are different
link |
from what you've seen before, nothing else.
link |
And you know what?
link |
You actually discover things
link |
that are interesting and useful that way.
link |
Ken Stanley and Joel Lehmann did this one study
link |
where they actually tried to evolve walking behavior
link |
And that's actually, we talked about earlier
link |
where your robot actually failed in all kinds of ways
link |
and eventually discovered something
link |
that was a very efficient walk.
link |
And it was because they rewarded things that were different
link |
that you were able to discover something.
link |
And I think that this is crucial
link |
because in order to be really different
link |
from what you already have,
link |
you have to utilize what is there in a domain
link |
to create something really different.
link |
So you have encoded the fundamentals of your world
link |
and then you make changes to those fundamentals
link |
you get further away.
link |
So that's probably what's happening
link |
in these systems of emergence.
link |
That the fundamentals are there.
link |
And when you follow those fundamentals
link |
you get into points
link |
and some of those are actually interesting and useful.
link |
Now, even in that robotic Walker simulation
link |
there was a large set of garbage,
link |
but among them, there were some of these gems.
link |
And then those are the ones
link |
that somehow you have to outside recognize and make useful.
link |
But this kind of productive systems
link |
if you code them the right kind of principles
link |
I think that encode the structure of the domain
link |
then you will get to these solutions and discoveries.
link |
It feels like that might also be a good way to live life.
link |
So let me ask, do you have advice for young people today
link |
about how to live life or how to succeed in their career
link |
or forget career, just succeed in life
link |
from an evolution and computation perspective?
link |
Yes, yes, definitely.
link |
Explore, diversity, exploration and individuals
link |
take classes in music, history, philosophy,
link |
math, engineering, see connections between them,
link |
travel, learn a language.
link |
I mean, all this diversity is fascinating
link |
and we have it at our fingertips today.
link |
It's possible, you have to make a bit of an effort
link |
because it's not easy, but the rewards are wonderful.
link |
Yeah, there's something interesting
link |
about an objective function of new experiences.
link |
So try to figure out, I mean,
link |
what is the maximally new experience I could have today?
link |
And that sort of that novelty, optimizing for novelty
link |
for some period of time might be very interesting way
link |
to sort of maximally expand the sets of experiences you had
link |
and then ground from that perspective,
link |
like what will be the most fulfilling trajectory
link |
Of course, the flip side of that is where I come from.
link |
Again, maybe Russian, I don't know.
link |
But the choice has a detrimental effect, I think,
link |
at least from my mind where scarcity has an empowering effect.
link |
So if I sort of, if I have very little of something
link |
and only one of that something, I will appreciate it deeply
link |
until I came to Texas recently
link |
and I've been pigging out on delicious, incredible meat.
link |
I've been fasting a lot, so I need to do that again.
link |
But when you fast for a few days,
link |
that the first taste of a food is incredible.
link |
So the downside of exploration is that somehow,
link |
maybe you can correct me,
link |
but somehow you don't get to experience deeply
link |
any one of the particular moments,
link |
but that could be a psychology thing.
link |
That could be just a very human peculiar,
link |
Yeah, I didn't mean that you superficially explore.
link |
Yeah, so you don't have to explore 100 things,
link |
but maybe a few topics
link |
where you can take a deep enough dive
link |
that you gain an understanding.
link |
You yourself have to decide at some point
link |
that this is deep enough.
link |
And I obtained what I can from this topic
link |
and now it's time to move on.
link |
And that might take years.
link |
People sometimes switch careers
link |
and they may stay on some career for a decade
link |
and switch to another one.
link |
You're not pretty determined to stay where you are,
link |
but in order to achieve something,
link |
10,000 hours makes,
link |
you need 10,000 hours to become an expert on something.
link |
So you don't have to become an expert,
link |
but they even develop an understanding
link |
and gain the experience that you can use later.
link |
You probably have to spend, like I said, it's not easy.
link |
You've got to spend some effort on it.
link |
Now, also at some point then,
link |
when you have this diversity
link |
and you have these experiences, exploration,
link |
you may find something that you can't stay away from.
link |
Like for us, it was computers, it was AI.
link |
It was, you know, that I just have to do it.
link |
And I, you know, and then it will take decades maybe
link |
and you are pursuing it
link |
because you figured out that this is really exciting
link |
and you can bring in your experiences.
link |
And there's nothing wrong with that either,
link |
but you asked what's the advice for young people.
link |
That's the exploration part.
link |
And then beyond that, after that exploration,
link |
you actually can focus and build a career.
link |
And, you know, even there you can switch multiple times,
link |
but I think that diversity exploration is fundamental
link |
to having a successful career as is concentration
link |
and spending an effort where it matters.
link |
And, but you are in better position to make the choice
link |
when you have done your homework.
link |
So exploration precedes commitment, but both are beautiful.
link |
So again, from an evolutionary computation perspective,
link |
we'll look at all the agents that had to die
link |
in order to come up with different solutions in simulation.
link |
What do you think from that individual agent's perspective
link |
is the meaning of it all?
link |
So far as humans, you're just one agent
link |
who's going to be dead, unfortunately, one day too soon.
link |
What do you think is the why
link |
of why that agent came to be
link |
and eventually will be no more?
link |
Is there a meaning to it all?
link |
In evolution, there is meaning.
link |
Everything is a potential direction.
link |
Everything is a potential stepping stone.
link |
Not all of them are going to work out.
link |
Some of them are foundations for further improvement.
link |
And even those that are perhaps going to die out
link |
were potential energies, potential solutions.
link |
In biology, we see a lot of species die off naturally.
link |
And you know, like the dinosaurs,
link |
I mean, they were really good solution for a while,
link |
but then it didn't turned out to be
link |
not such a good solution in the long term.
link |
When there's an environmental change,
link |
you have to have diversity.
link |
Some other solutions become better.
link |
Doesn't mean that that was an attempt.
link |
It didn't quite work out or last,
link |
but there are still dinosaurs among us,
link |
at least their relatives.
link |
And they may one day again be useful, who knows?
link |
So from an individual's perspective,
link |
you got to think of a bigger picture
link |
that it is a huge engine that is innovative.
link |
And these elements are all part of it,
link |
potential innovations on their own.
link |
And also as raw material perhaps,
link |
or stepping stones for other things that could come after.
link |
But it still feels from an individual perspective
link |
that I matter a lot.
link |
But even if I'm just a little cog in a giant machine,
link |
is that just a silly human notion
link |
in an individualistic society, no, she'll let go of that?
link |
Do you find beauty in being part of the giant machine?
link |
Yeah, I think it's meaningful.
link |
I think it adds purpose to your life
link |
that you are part of something bigger.
link |
That said, do you ponder your individual agent's mortality?
link |
Do you think about death?
link |
Do you fear death?
link |
Well, certainly more now than when I was a youngster
link |
and did skydiving and paragliding and all these things.
link |
You've become wiser.
link |
There is a reason for this life arc
link |
that younger folks are more fearless in many ways.
link |
That's part of the exploration.
link |
They are the individuals who think,
link |
hmm, I wonder what's over those mountains
link |
or what if I go really far in that ocean?
link |
What would I find?
link |
I mean, older folks don't necessarily think that way,
link |
but younger do and it's kind of counterintuitive.
link |
So yeah, but logically it's like,
link |
you have a limited amount of time,
link |
what can you do with it that matters?
link |
So you try to, you have done your exploration,
link |
you committed to a certain direction
link |
and you become an expert perhaps in it.
link |
What can I do that matters
link |
with the limited resources that I have?
link |
That's how I think a lot of people, myself included,
link |
start thinking later on in their career.
link |
And like you said, leave a bit of a trace
link |
and a bit of an impact even though after the agent is gone.
link |
Yeah, that's the goal.
link |
Well, this was a fascinating conversation.
link |
I don't think there's a better way to end it.
link |
Thank you so much.
link |
So first of all, I'm very inspired
link |
of how vibrant the community at UT Austin and Austin is.
link |
It's really exciting for me to see it.
link |
And this whole field seems like profound philosophically,
link |
but also the path forward
link |
for the artificial intelligence community.
link |
So thank you so much for explaining
link |
so many cool things to me today
link |
and for wasting all of your valuable time with me.
link |
Oh, it was a pleasure.
link |
Thanks for listening to this conversation
link |
with Risto McAlignan.
link |
And thank you to the Jordan Harbinger Show,
link |
Grammarly, Belcampo, and Indeed.
link |
Check them out in the description to support this podcast.
link |
And now let me leave you with some words from Carl Sagan.
link |
Extinction is the rule.
link |
Survival is the exception.
link |
Thank you for listening.
link |
I hope to see you next time.