Every

day.

We live like.

That.

Like

it.

Used by the title of the work
of all the take up all of

the time.

From three point three thank you all
right so a pleasure to be here today.

And indeed most of the work that we do
you know Lab focuses on robotics and

some of the more well known work
that we've done has to do with

lots of interesting mechanical structure
like the cell from directly caving robots

there's a modular on one.

The.

Feedback feedback control so so
these are robots that can build identical

robots given the right supply of materials
so this is sort of an interesting take

on robotic ecologies we make all kinds
of crazy machines of various shapes and

sizes this is a ten segment the robot with
all kinds of cables an interesting thing

about this robot is that by pulling
on its own cables it can roll or

it can collapse into a flat thing and
crouch under the door do all kinds of

interesting Star Wars like stuff but
that's not what I'm going to tell you

about what I want to tell
you about is a different.

What I want to talk about
is a different sort of.

Story that started quite a while ago and
when I was doing my post.

Dog and
it's about making adaptive robots so

let me start of the beginning and you
know most of most of us to do research in

robotics and most of you
probably know that robots the day

are super human in almost any way you
want to measure it so they're fast

they are strong they're powerful they're
work twenty four seven they're precise so

almost any metric you want to put on it is
a robot that can do better than human but

there's one thing that robots
still can do very well and so

important for you to know in case
there's a robot uprising and

so also what is that it's their
ability to adapt to new situations so

most robots the day work very well as long
as they're in a structured environment

where everything is the right
time the right place but

once things get wounded big things
begin to change in unstructured way

most robotic systems today have problems
and a lot of our research is about

making adaptive systems that can
handle that kind of uncertainty so

that uncertainty becomes even worse when
the robot itself changes if a ball falls

off of this robot this game over most
of these robots cannot take care of

themselves now when you look at biology
it's exactly the other way around

the biology is not as fast not as
precise materials aren't as durable but

biology is all about adaptation and
adaptation over if illusionary time scale

and that a patient over the lifetime of
individuals that learns of the body heals

adapts recovers all forms of adaptations
and all scales so what can we learn from

biology and apply to robotics in order to
solve this adaptation problem so that was

the challenge I was given when
I started my postdoc at MIT and

at Brandeis this is a picture of Brandeis
complex system center and doesn't

look like it now it's all covered this No
but this is a good picture and when I came

there the idea that we had with
Professor Jordan power because my advise.

So the time is to use evolution so

the idea of evolution basically was to
breed robots instead of designing them

just them sitting at a desk and designing
it think about the robot how to program it

what components go into it how do we do
everything and think about all the details

we're just going to throw a lot of robotic
components into a big physics simulator

allow evolution allow random process to
connect them in the rain random ways and

then going to test them for whatever task
you want to do let's say their ability to

crawl across an infinite plane and
they're going to take the ones that do

less worse than others and
crossbreed them and throw them back into

the simulation in the next generation and
so forth so if we do this enough times

we're going to get lots of bars in wires
and neurons connected together then we'll

start getting some possibly interesting
things so that was the basic idea and

we wanted to see a far we
we could get with that so

we wrote a proposal for DARPA DARPA said
Sure here's two million dollars to try

it out never managed to replicate
that that kind of effect but

but it worked at that time and
we got ourselves a super fast computer and

on existing processor computer it's
a slower than your cell phone today but

this was a big deal I'm not talking about
the fifty's right this is two thousand

big deal of the time
a very fast computer and

we started running this thing so what you
see here is a plot over time of lots of

different robots each dot here is a robot
and you can see generations on the X.

axis and speed of this pile of
junk on the vertical axis so

you can see from many generations nothing
happened these are just static piles of

junk with wires and neurons and things
hanging off of them they're not going

anywhere speed is the euro after hundred
generations something happens they

begin to move and then nothing happens for
a long time it's only as something a disk.

Over the Ganser get this eclipse
punctuated equilibria development and

what if you look inside
you see these sort of.

These piles of junk that
don't do anything but

some only a pile of junk happens
to vibrate a little bit and

that vibration makes it move
a tiny amount it's not a lot but

it's infinitely better than the other
piles of junk not moving at all so

these vibrating piles of junk take over
and then they get better and better and

after a while you get interesting
sort of behaviors that are.

That basically are robotic motions that
here you see two of these very early

machines crawling across the infinite
plane in the simulator the one of

the right almost looks like it was
intelligibly designed it has this sort of

symmetry to it it wiggles its tail it's
a very interesting kind of the zine and

it's a cross for so one thing we did is to
take these designs and have these robots

to crawl out of the simulation into
the physical world using a three D.

printer so for you printers were
a new thing at the time and

here there are two of these robots
crawling across the floor three D.

printed with crude three printers
of that time and then tire about

this point in one shot so everything
there is printed pre-assembled So

when you look at these robots you're have
to remember these are very simple but

they were designed sort of automatically
through this process of breeding.

So since then we've you know our research
going to split off into lots of different

directions one direction is improving
the three D. printing because one thing

the first created me a lot without process
is that we could only print the body

the white plastic but equally couldn't
print the wires and the batteries and

everything that makes the robot robot so
since then we've been working a lot on

printing electronics and we can now print
wires and actuators even batteries and

things like that we haven't quite
pointed the complete robot but

we're getting close and
there's a lot of people catching up with.

Technology so
you probably heard about three D.

printing and sort of the next phase
is printing electronics and so for

this is a printed loudspeaker hundred
percent printed with wires and magnets and

everything and you can play music on
it and a lot of by products happened

when we did this exploration like printing
biomaterials this is a printed miniscule

cartlidge of the need for
implants this is the printed cookie so

food printing another new area that's
that's happening unfolding as we speak

very excited about going to talk about all
this three D. printing stuff even though

there's a lot of interesting things
happening I want to get back to our story

of evolution so since doing the war

we've been spending a lot of time evolving
in breeding robots of various shapes and

sizes and we switch from from evolving
rigid structures to soft structures so

what you see here are a couple
of evolved soft robots and

it's a skipping because of the video not
because of the robots but you can see sort

of interesting evolve structures that
happen when you let evolution breed robots

made out of soft materials much
more difficult to simulate but

much more exciting and much more rich in
the space of design little bit closer to

biology but let's go back to our story
back in two thousand so we printed

these robots here both of one of them and
that's the point where I was looking for

my faculty position I just
finished my postdoc and and

the question is you know what what what
where can I go with this kind of idea so.

You know it made headlines appear on
the front page of The New York Times robot

making robots and
end of the world that kind of stuff but.

I got my fact of the position at Cornell
based on this list of that work so

that was a very exciting moment but
I knew.

I'm not going to get tenure
making plastic robots.

I needed to make robots made of titanium.

And we the robots to
look much more complex

that have much more much more
sophisticated structure that could impress

my colleagues in
the control theory group so

we build this incredible
robot it's actually.

Made of aluminum by the looks like a team
and it has this paint ball canister in

the center it's twelve pneumatic activator
is very powerful This is a powerful

machine in the sense that it can flip
up in the air it can do loops and

it's a very large power density
this machine very rich.

Potential space of
behavior the question is

how can we control the can we make
this thing gallop through the fields.

So so the idea was again to use this kind
of evolution because it was very difficult

to simulate we said OK we're going
to put it in the big cage and

we're going to see if we can breed
controllers for a fix morphology so

we have a fix morphology here
that we had this line but

we're going to let the controller
evolve so we're going to have initially

maybe a hundred different controllers
that's a neural networks they're all going

to take turns riding this robot the ones
that in there's a camera at the top left

the watch is how far this robot moves
Here's the camera view it actually watch

of this red dart and the farther the red
dot moves the better the controller is and

against a cross breed with other
controllers and so forth there's a very

fairly automatic process you leave it
overnight it just keeps on running.

Even a real here that pulls
back the robot when it's done.

And this is the only and not automatic
thing here is underneath it here

behind the curtain there's a student and
the student once in a while comes out and

pulls the robot out of the corner
when it when it's stuck but

apart from that it's completely automatic
you just leave it overnight and

you let it run so we left it over night.

And we got some interesting results
here are some of the first.

Behaviors and
if the video will start playing

we'll see it.

OK so
I'm going to just skip the slide apologize

so.

OK so.

In that video you would say you
would have seen this amazing robot.

Galloping in the field you know.

So the truth is that that robot
did not go anywhere all right

it stumbled along and.

It made a lot of noise and.

The air pressure made a lot of dust also
some of those it was amazing experience

but it didn't go very fast anywhere and
we're going to conquer the world and

neither was I going to get
tenure All right so so

this is a this is where the sort
of the story could have ended but

but we sort of we said OK we were doing
something wrong so let's go back and

look at these two project the first
project we had a big physics simulator and

we bred everything inside the simulator
and then we sort of after a week

we took the best of all of that we had
then we built it in physical reality and

it happened it so happened that
those robot work pretty nicely but

they only work nicely because the robust
a very simple very quasi static kind

of behavior but in general we have what
we call the simulation reality gap and

if you're doing any kind of robotics
you know that will work the fact

that works and simulation doesn't mean
it's going to work in reality and

that gap grows the more complex
the robot the second project

that we looked at had a there was no

simulation at all we bypass the whole
simulation thing you all happen.

In the big cage but the problem there was
that it took too many physical trials and

there is a limit to how many times you can
run your robot in a big cage over night

and we can do it fifty times a hundred
times five hundred times is not

enough to do a lot of learning if you're
just sort of blindly reading controllers

through trial and error so.

Again the more complex the robot the more
difficult it is going to be to learn

things in physical reality so we start
with when these two approaches simulation

doesn't transfer reality to expensive
What do we do so the third approach

which you're going to sort of think about
is a hybrid approach it looks at it in

different ways so they there is very
simple we start off with a simulator and

I start putting a similar in quotes
because it's a simulator that are not very

good it's a it's a sort of crude simulator
we take what we have and we use it to do

breed robots then we take the best robot
that we get we build it in reality but

because that robot the simulator isn't
very good the robot is not going to work

in reality very well either Nevertheless
we take that robot is that experience and

we collect a lot of data about it so
we collect specifically actuation and

sensation data we can we collect more
commands let's say an acceleration

of actor and sensory information
we take all that data and

we use it to breed better simulators so

just like we were breeding robots before
to accomplish a particular task we now

also breeding simulators and

the figure of merit of a simulator is
how well it does compared to reality so

how well it predicts reality as
compared to what reality is for real so

if you think about it there's a sort of
core evolutionary process happening here

this sort of two things the optimizer the
same time robots are being optimized for

a task based on the simulator
simulators are optimized for their sin.

Really simulation capacity based
on performance of the robots and

these two things are sort of feeding off
of each other in an interesting way so

there's an interesting relationship
between them and maybe a sort of

competitive cliff illusion like
predator prey or maybe it's a.

Symbiotic relationship where they help
each other maybe it's a little bit

mix a bit like Professor and
student right so

there's sort of up teacher a learner kind
of relationship between us is interesting

to think about that analogy but that helps
us set things up so this is a third robot

that we built this way in my fifty
years of the clock is ticking and

this is an interesting robot it has four
legs as you can see it has eight mortars.

Two on each leg one of the hip and
one of the knee and

it also has to feel sensors that
tells how still to the left and

right for them backwards a very
sort of sparse machine and

what he needs to do is to learn how
to walk it's a very simple task but

the challenge is that this robot
has no clue what it looks like so

you look at this robot and you can
understand how four legs maybe can start

thinking about how to make it move forward
even though it's very tricky because these

legs and only go up and down they can go
sideways so only eight more but this robot

has no clue what it looks like doesn't
know if it's a snake if it's a spider if

it's a tree doesn't know how these pieces
are connected and how long they are in all

that kind of stuff so it needs to figure
out how to move so what it does so

to understand what that means Imagine
yourself a moment seeing in the black box

with no windows and all you have are eight
norms and as you turn these knobs you

can feel the box tilting left and
right forward and backwards and

your goal is to turn the knob in such a
way that makes you cover the most distance

how do you do that so of course you can do
trial and error by the you know this is

different very difficult so maybe maybe
this is what the brain of a newborn

child feels it has all these actors
has these sensors and needs to do so.

But can't figure out yet
what it is and how to do this so

this is what this robot is trying to
do and that's how we set it up so

we set it up in a sort of iterative
process as I described before

where we have a robot in the middle and
it's going to start off by doing a random

babbling it's just going to move
its Motors in a random way and

it's going to then collect all that
data and it's going to generate

hypotheses about what it is the top
left box there the yellow box all has

of hypotheses about what it might be again
it doesn't know if it's a snake a spider

is just putting these parts together and
they all see a lot of these models in

the moment in the movie hopefully of the
movie works and you'll be able to see that

they're all wrong because you happen to
know what the rover really looks like but

insofar as the robot knows these
are going to be valid hypotheses

about what it is right the next thing is
that in this is a really tricky step is

the green box is going to figure out
how to activate the model the motors

to create the most disagreement
between predictions

of these different models so why the most
disagreement it's a little bit like

a scientist designing an experiment and
the best experiment to do is one that

causes disagreement between predictions
of two competing theories because

then the results from that experiment will
refute at least one of those theories so

it looks for them for the maximum variance
in predictions of all these models and

then it takes that next and
various variants activation in actually

perform as this only it's a second
physical trial it does that

that new data eliminates some of the
models and then it goes again looks for

disagreeing with and so forth that
will do fairly few trials like that

fairly few physical trials a lot of
computation he remains grounded and so for

the until it cannot make it small disagree

disagree more at which point it will look
for a pattern that makes the model moved.

Forward and whatever because the model is
pretty close to reality will make a small

move forward will also make
the robot in reality move forward so

we started that off a while ago and

I hope this video works
because I have a lot of those.

There you go.

Thank you so OK so

here the robot is finally moving
this is one of our first experiments

if you do robotics you know the camera
is always rolling and the robot died and

that was it so you know it so this is the
lesson here is you don't explore the robot

to explore too far beyond what's safe for
the robot this is an interesting sort of

whole area of thinking about
the future of robot safety but

here's one of the more
successful runs this is.

Now the robot exploring itself.

So here it's going to.

Hopefully.

Start and experiment with itself.

All right so if this with video would
have worked I'm tempted to try it out

the every every experiment is expensive
right so you know I think about it so.

So what this robot does it begins by sort
of doing a few trials here a few trials

there and gradually it begins to form
a model of itself wish you could see this

and and it takes it about four days and
after four days it has a pretty good

image of itself with
this it looks a bit like.

A cartoon image of itself made of blocks
very similar to the kind of early

rovers that I showed you earlier and
it's pretty close to reality and

using that self image it can
learn how to walk in at the end.

It does this sort of interesting crawling
behavior and you know we were hoping

to get an evil spidery walk out of this
robot but instead we got a kind of

lame way of moving forward this
robot kind of crawls encroaches and

sort of moves in a way that no human would
program it but when you look at this

that behavior have to remember that
it was not programmed by humans.

It was there was no model that was
you used to try things in simulation

in order to do trials of walking
before that behavior again so

did it only uses these brief brief
trials with brief experiments and

figure out what it is and how to move
forward All right so after that we decided

to something very cruel we chopped off a
leg of the sort of to see what happens so

we move the way we put it back
on again but no need to worry.

By that and we watched what happened and
within about a day the robots self image

also lost the leg was very interesting
the robot sort of moved around and

it's in self model began to adapt and
one of its legs became shorter and

shorter and till it sort of more or
less resembled this broken leg and

then using that self image it
moved it learn how to move but

it moved in different ways sort of limped
around a qualitatively different way of

moving around that took of that sort of
compensated for the fact that it only

three and a half legs so again if you
looked at that robot you might say

OK you have to remember that there
wasn't let's say a a sensor that said

came off switch to plan B.
There was an engineer that sat down and

thought OK if the leg comes
off here's plan B. And

let me program of it know what happened
to lead came off the dynamics change

therefore the self model change therefore
the behavior change show it so remember we

talked about adaptation and that sort of
is one what was behind the scenes where in

the given idea of how
this thing works as that.

Mystically these are thirty independent
runs they all start at the origin

there is a sort of the red dots
are what the random controller does.

And the.

Black dots are where the robot thinks is
going to get to using its self image and

the blue dots are where it actually gets
to physical reality so that you can see

here is like anything else in academia
this robot has a self inflated self image

about what it thinks it can do about
twice as much as it really can but

nevertheless that self image allows it
to make the right kind of decisions and

this is a very interesting question we
talk about about human self image and

this is good to have an inflated self
image of or not an optimism and pessimism

of these kinds of questions this is
an interesting window into some of that so

you can see here the robot whether it's
self image kind of looks like all right so

there's a lot of interesting questions and
things to do on that one of the projects

we're doing now are robots that
create a model not of themselves

the robots the create models of other
robots purely by observing them so

not by communicating with
a human engineered protocol but

just observing them like
the way you can create in your

mind a model of a cat by looking at
a cat there's no bluetooth communication

protocol between you and the cat it's
all you are observing in Mali and

we're trying to do that which robots that
can model not just their own mechanics but

their own behavior so that's
equivalent to sort of introspection or

self-awareness of this all these
interesting robot psychology experiments

to do there but I'm not going to talk
about any of that I'm going to talk about

go back to my story and tell you what
happened after we did this all right so

we came out with this came out in
science lab people were excited and

a student came along it's a let's take
that algorithm out of the of the robot.

And try them all general dynamical systems
so you think about of this robot inside

of it so this robot is sort of a dynamical
systems with mortars and sensors and

things but inside of it there's an
algorithm that just is looking at a black

box with inputs and outputs and is trying
to maul what's going on inside it's a very

sort of generic kind of thing and
switch back as well to have videos for

a while but
have tons of cool any Maisha it's.

So.

They hope those work.

So we.

So we.

So they there was that maybe we
can generally take this algorithm

out of their wild and try to maul
various general that ample system so

now if you think about
modeling their systems.

It's basically a sort of machine learning
task in we call it system a fixation but

all these different problems all have
to do with observing a inputs and

outputs the system understanding what's
inside and what's interesting about this

particular situation is that you're
not just observing a system inputs and

outputs and trying to figure out
what's inside you can actually choose

what is going into the box so you have
control over how you perturb this black

box you can interrogate the black box
by perturbing it in various ways and

watching what happens and then comes
the question how do most efficiently

perturb it how do you ask the least amount
of questions to get out the most amount of

information so so that's sort of where
where we are with this system so

I'm going what I'm going to do is I'm
going to replace this this is the same

slide as I had before with a robot that
instead of a robot I have a sort of.

Some experimental system with lots of
inputs and outputs instead of models

made out of building blocks
of components of robot.

Points are going to have building blocks
I'm going to have models made of building

blocks of equations so
lots of different types of the plus minus

divide multiply and so forth so
literally building blocks of equations and

I'm going to have instead of patterns
of actuation of the motors as my tests

I'm going to have various knobs of the
experiment that I can turns I'm going to.

Basically change or the initial conditions
of the experiment and so forth so it's

the same idea but I have a very different
sort of more general set up of how I

interrogated an ample system now what I'm
showing here this is where the experiment

is and you can have a real physical
experimental system with lots of knobs and

things like that and measurements you can
have a human doing experiment in service

of the robot that is coming up with
a path assist apostasies that's a sort of

reversal of roles most usually we think
about robots as doing the experiments and

the human coming up with their policies
but this is the other way around or

in this it turns out to be the most
practical way of doing it and

the way most people do it is they do lots
of experiments randomly put them all in

a big database and
then when this entire process asked for

an experiment there's
a separate process they go so

this big database pulls out a snippet
from a previous experiment that

is as close as possible to what
was requested and uses that So

these are all different ways
of implementing this but

this is probably the most practical OK So

the first time we implemented this
idea outside of robot was on this

bridge this is a bridge across every
morning when I walk into Cornell

and as you can see Cornell is
very easy to defend So this is

the this is the one of the advantages
we have over Harvard for example but.

It's also when you walk into across
this bridge this is a foot narrow

footbridge you have a long time to think
what would happen if something here wasn't

as strong as it needs to be and so
what we decided to do exactly the same

thing we have vibrators on the bridges on
simulation we didn't do this you couldn't

get permission to do it on the real bridge
but this small Vibrators the vibrate

the bridge in different ways sensors that
can measure how the bridge right breaks

in from this relationship between
activation and sensation it can for

models see how to activate again to
make them all disagree the same idea and

we could find a broken link or

a weakened member of this bridge
a lot faster and more accurately

than conventional civil engineering
approaches against a very simple idea and

applied here to a bridge so err Here's an
example applying it to a biological system

started to a circuit in a black
box we can perturb it in different

ways observe the outputs evolve
the test to make the model disagree and

we can figure out what's going on in
the circular not only inner circle.

Pretty nicely so
lots of examples like that so.

So that brings me to.

This so
I can show lots of examples of famous but

I want to kind of focus on
mathematical modeling so

brings me back to this idea that I have to
take rest for a second talk about symbolic

aggression embodied you know what
symbolic regression is nobody knows and

nobody admits All right so
symbolic regression is this idea.

So you all know what a linear regression
is known or aggression where you fit

the equations to occur to you and fit
curves to the data and you let you think

about the curve but it looks like
symbolically and then the computer fine to

use the coefficients I left that's what
regression is very easy to understand.

But what's embark aggression is.

When you don't even have you don't even
know what the form of the mole is so for

example look at these dots over here so

there's a couple of dots what's
the function the generated the fits this

dot right so because there's no any
mation of the answers right here but

if there were any Masons I would give
you a second to think about it or

maybe a minute and somebody would say
this is exponent times sign right

there if you were able to defer trying
to do that in your head you were sort of

trying to do symbolically version because
you don't know what the function is

you don't care about fitting coefficients
you're taking basic mathematical

alphabet functions that you know through
the metric functions exponential functions

here trying to see how to put them
together to qualitatively get that data so

that was embargoed in his in arguably

that's what a lot of scientific Malling
is we look at data we see patterns and

we try to figure out what's the what's
you know what's the how they describe

that mathematically using building
blocks where where so I can get more and

more complicated that was a simple example
to get multi-dimensional to get very

complex so you know what's the function
that fits this data center this is a lot

more complicated and so forth I have
this in case somebody would have solved

the other Want to quickly
show that all right so so.

So the way we thought of doing this
this so this is sort of the old

symbolically version idea from ninety
two is just like we breed robots

to walk we can breed equations to fit
data very simple idea so we have lots of

equations represented this trees this is
a tree that represents the equation X.

one minus three times sign of X.
two plus minus seven so

this is a mathematical expression and
it's a two dimensional expression has X.

one X. two and we can use it and
measure how well it fits there this is.

This is a function here and
you know we can start off with just.

We started with random piles of junk
we could start with random piles of

equations don't make sense known for
the data and over time they get better and

better right so that's the that's called
symbolic regressions technique invented

back in the early ninety's.

By John Koza from Stanford is a great
idea there's only one problem it doesn't

work why why apart from that is that it's
a great idea and published papers and

everything but that was so but why doesn't
the worry has the two problems as do

most or many machine learning
approaches one over fits the data.

You get crazy expressions have nothing do
with reality and the second problem is

that it's too slow why because you have to
test a lot of functions against a lot of

data points takes forever and
it just doesn't scale well so

so if this technique has been sort of
mostly use that condemn a CLI But but

it was interesting so what we did
is basically a small change and

we the only thing we did is we bring in
this idea of cold evolution so not so

just like we didn't just evolve
robots will be also clearly of all

the simulators if you remember what we're
doing here is now we're going to evolve

models to fit the data but we're not going
to fit the models to the entire dataset

we're only going to fit them all
to a small subset of points but

those subset of points are themselves
evolving Sometimes it's these ten

points sometimes it's an different set of
ten points they're going to be a different

subsampling of the tire did a set and that
subsampling is going to be rewarded for

how much disagreement it can
create between models so

a point that all the models can
predict correctly is useless for

breeding better models but

a point that some models think is going
to be value five in other models.

Addicted to be a different value when you
actually measure it will tell you which

model is right and which maul is wrong so
it turns out that you can use much much

fewer points than the entire It is said
to breed models so what happens here is

two fold it runs faster and
it of voids all of the overfitting So

two of these problems go away so here's
some examples this is trying to fit.

A general this is a symbolic aggression
problem of the that is written here

if you use the entire data set even
the random subsampling it sort of

over time begins to find it but if you
use the core evolved there the evolved to

defeat to create the most disagreement
between your models you can find

it takes a long time in the beginning but
eventually it.

It can find them all more accurately and
sooner than other approaches this is

another chart that shows that if you
use the core of all data said even for

the same amount of C.P.U. you get much
simpler solutions for the same accuracy so

it involves a lot of the overfitting
why because of the many ways to overfit

the data and so there's a lot of
disagreement the bad mauls disagree how

to overfit and then they get thrown out so
interesting sort of.

Perspective to me this was very
interesting because it suggests that

more data is not necessarily better and

this flies in the face of all the big data
movement in all that when we have more

data it doesn't necessarily get better
we have to have the right data and

this sort of coalition area idea helps
you find the right data All right so

with that tool in mind what do we do first
thing we try to do is the fit the stock

market and why not I'm still here so

obviously that didn't work but by they
did work for some and there's a couple of

blogs that use our software and
then the blogs were taken down so

that's the ultimate proof
that it worked so OK we said.

Try to find the equation for
all the prime numbers

small simple tasks we threw in the prime
numbers we hit Enter we waited waited for

a week and we did not find a formula in
fact we found some approximation some well

known approximations but nothing useful so
think that we're not going to be rich or

famous we said let's write some
proposals since here we can do so.

So we published this and.

Immediately I got an email from.

CERN from a particle accelerator.

And Geneva and they said we have
the perfect problem for you we have this.

We have this.

Data set that describes
the we have a gap between

the binding energy of atomic nuclei doing
exactly understand that means that but

they have this gap they can't explain and

I thought maybe our software can
actually find what's going on so.

We ran it and we got this top expression
here we're very excited it's a very simple

expression explains everything to three
digits so we're very excited we wrote

back and they merely wrote back to us and
said we were ready know about this thing

it's called the Y. It's like a formula
in our formulas good to four digits.

So there were disappointed but
actually we were excited because it means

our software can find
named formula right so

that means that if we get the right data
we might actually be able to generate

something that nobody knew before so
we decided to kick it up a notch and

try to model dynamical system
instead of static systems so

whatever would what do you do in order
to model the emphasis is actually simple

in principle you take time series data and
you computed the riveted

you create a static model of
the derivative and that's a dynamical.

So we can we take our time series data
compute the route of in fit the derivative

Now if I have my so there's lots of
things that you can try with this this

is one of our first experiments we took
seven equations that describe the glycol

as the cycle of the cell I don't know
exactly what that means either but

the looked like pretty complicated
differential equations to me.

We simulated them to create data
we sprinkle ten percent noise and

we fed it back to the system without about
four within about four hours able to

generate the seven differential
equations that describe this data and

they matched almost perfectly for this
storm over here which is wrong but but

but small so we have this incredible
hammer that can look at this crazy

the medical system with all
kinds of interdependency and

write down non-linear equations and
again it's not as fitting parameters

the only thing we gave it here is plus
minus divide multiply and open and

close brackets that say there's
no biology background nothing and

is able to come with it so we started
running this on all kinds of systems if

the video is working this is
the You'd see this flapping.

It doesn't even want to
skip through the videos so

we ran it on.

On this flapping So

there's a flapping having robots very
cool flab it collapse its wings and

hovers in places what weighs about three
grams was the world record holder for

a couple of months is very stable
you can you can punch out it and

it stays in place it hovers untethered for
ninety seconds.

And it's passively stable so
there's no control system in it but

it turns out the mauling this flexion
of the wing was very very difficult

people spent careers mauling
the elastic behavior wings and

we said OK let's dump it in here and
see what happens so.

And we created models that
were more accurate and simpler

than state of the art models in air the
NAM explore for flapping wings so again

it shows you that this is sort of them
only a tool that if you're a scientist you

want to pay attention to because it might
show use a little bit of where things

are happening is another interesting
example of this is a mother a guy.

Came and wanted to create models for tire
road interaction this was important for

the SE formula team but
it's also important for gaming and

lots of the vehicle safety and
again so they created got

all the day they bought the data barely
the formula as the team have paid for

all this data it plugged it in and
you could find models that are simpler

and more accurate then state of
the art Malls used my industry and

again I'm emphasizing these two things
as these are usually the two dimensions

we care about in science accuracy and

complexity usually there's a tradeoff you
can have models that are more accurate but

also more complex models are simpler but
less accurate and there's room for

Also people lots of papers in between but
what you usually don't want is for

somebody to come up with a model
that is simpler and more accurate

than what you've done because that
really means they don't mean it you

in terms of scientific accomplishment and
that sort of what this thing is doing so

I have lots of examples of that kind of
thing so I just want to although the last.

Piece of the story is that we.

Ended up.

Talking about this other conference and
asking for data and somebody came up and

said I have this perfect data
set this about logical data set.

Tracking cells over time but these
are individual cells individual cells.

That can come and go and
tracking various ingredients in them so

he could track those cells and
collect this data he said

I'll give you the entire data set and
maybe you can find what the equation is so

we plugged in the data we hit enter.

And we waited for a week and

we ended up coming up with the system
came up with these models and

here the top so either of this is true or
not I email this is girl soil who

did this to give us the data and who looks
at the small and he says this is told crap

doesn't even does not even biological
there's nothing you know it's wrong and

you know we were again
slightly insulted by that but

the but we also thought OK we're
probably doing something wrong and

you know we have this awesome algorithm
but it's coming up with garbage results of

what's what's wrong with it so
we thought about this for a while and

realized we were giving it the wrong
building blocks of first of all we gave it

signing call sign as building blocks and
that's totally relevant for biology and

that's an engineering thing physics so
we threw those away but

there was another building block that we
were missing but was that so we thought

about eventually realized the missing
time delays so biology the differential

equations new quiet time to his nothing
happens exactly the right time right so

so the biology he's the Slackware's
you know physics the pendulum can't

delay right so he put in
the time delay and rewrite it and

BAM we've got these two
beautiful equations

that describe the entire system different
cells might have different professions but

they all full exactly this behavior so
we said this to grow and

he said we have a problem
this is the established

mall in the field and
your model is simpler and more accurate.

All right so what do we do.

So.

So that's exciting but
there's only one problem and

that is that we have
no clue what it means.

This one was the rifle first principles
and we understand everything here this one

we don't know what it means so
since then we released

this software there's forty thousand
peoples right now using it lots and

lots of papers whose main result was
generated by the software called Eureka

misspelt with a Q So
you can find it easily.

I think the student who did
this has created a start up but

the software still free I think for
academia and a lot of people using

it doing lots of interesting things often.

You know half of them are trying to
predict the stock market but others

are using it to do scientific discoveries
because you can give it a lot of data and

look at sort of what's happening in there
trying to understand how things are going

on so will this replace human
scientist probably not.

As I was very careful to show you we
humans are still needed to choose what

variables to feed the system to decide
what building blocks to allow the system

to work with and if we give the wrong
building blocks we get garbage results and

finally to assign meaning to the results
but I do think this is the future

of science THANK YOU THANK YOU I.

Thank you.

Yes So you're asking the question of
scale ability and you can sleep through

a talk about AI and then you wake up in
the end he said how there's a scale and

that's always a good question to ask.

And so so this particular process
sort of scales almost linearly with

the number of equations but exponentially
with the complexity of each equation so

the hope is that we can
model things that have

involved interaction of many things
that each of them is fairly simple but

if you're trying to model something
that involves a few monolithic monsters

it's going to be very difficult to maul
it with this or with anything else but

to address the question of scale
ability what there have been shown here

to have time is one thing we did is
that we analyzed many systems and

began to look for
motifs in that begin to reappear and

once you take those materials and
you capsulated them as new building blocks

you can start accelerating the search you
can actually discover the alphabet of the

domain that you're interested in I think
that's the real key to scaling That's how

science Those are we encapsulate
concepts and use them as building blocks

that's what this system that's the only
way really to scale to the high

level.

So it's all odd with them that try to
create models out of they die if you

give them the wrong data there will
not work the question is will they

tell you they're not working or will they
give you some wrong answer which is worse

at least I can say here when we when
we feed it let's say random data

then the only the only mall
it generates is a constant

This is this is the only thing I can
predict about this David is it ever so

that kind of can give you a sense
that you have the wrong data set but

apart from that it's very difficult.

If you don't cover the entire decide
that's wrong data is the same thing.

Yeah.

OK.

Thank you thank you.