Press "Enter" to skip to content

AE + AI: the way towards autonomous computers | Pierre Collet | TEDxUTTroyes


right so the topic of tonight is how to

shape the future and because I’m a

computer scientist so something that we

are working on on our research team

right now in Strasbourg University is to

think on how it is possible to make

computers to make them autonomous to

have some creative autonomous computers

so in order to understand how this can

happen it can happen through the mix of

two things which is ai ai is artificial

intelligence but also artificial

evolution most people don’t know that in

parallel with the development of AI

artificial evolution has also developed

so if you look at a timeline on how

computers have and research on these

domains have been now going on

well 1953 is more or less the first

business computer the IBM 650 okay and

you see that artificial intelligence

developed in 1955 1960 56 and so on so

it was what some of the very very first

algorithms that have been developed on

computers were of artificial

intelligence but of course the very same

thing that has been puzzling people is

about evolution so if in 1953 you’ve got

first trials in Australia USA in Europe

in 1957 chains of bits have been evolved

by Frazer and you go on like this

and those two technologies have come up

come of age and interestingly enough it

has been sooner in artificial revolution

than in artificial intelligence

ever since the year 2000 now we have

human competitive computers algorithms

that are capable of creating of

innovating creating get imagining new

solutions that had never been thought

about before by human beings

and since 2009 there’s a there has been

another revolution which is the advent

of graphic processing units these are

massively parallel chips on which you

have several thousands of course that

can work in parallel all right but the

only problem is that you need to have

the right algorithms in order to be able

to exploit them and artificial

revolution is one of them and in

parallel artificial intelligence is also

one of the type of algorithms that can

exploit these GPUs with in 2012 the the

the advent of deep learning so for those

of you who not who don’t know about deep

learning yet and who don’t know what

deep learning is capable of doing here

are a couple of examples so the

state-of-the-art here is deep learning

has been developed by young account and

here is typically what comes up if the

video is coming yeah so this is what a

computer is capable of doing right now

yes what is the man doing typing is the

baby sitting on his lap yes are they

smiling yes is there a baby in the photo

yes where is the baby standing bathroom

what game is being played

soccer is someone kicking the ball yes

what color is the ball yellow what game

is the dog playing frisbee what color is

the dog black is the dog wearing the

color yes

what is the cat sniffing banana where is

the cat

I’m bed what color is the cat black and

white what color are the bananas green

okay all this is being done in real-time

it is that’s the current state of the

art of deep learning as it is called

which is what has been developed by

Facebook through Yemen account this is

one of his achievements but also the

problem is that you can have some big

problems too because now you have Google

photos that is using same technology

with sometimes some bad labeling on some

photos but then so not only is it is

this working on photos but it’s also

working on videos here you’re just

looking for people all right the

computer is automatically detecting the

people but not only it is is it capable

of detecting people it’s also capable of

detecting any kind of other things so

you can take a computer for a walk you

know in town and so this is the analysis

of the computer a man in a suit and a

tie

hold standing in front of the building

of course it’s not a tie so computer

makes some mistakes but you can see that

the zip could really be mistaken for a

time here a sign on the wall that says

it’s trying to read the sign a man is

walking down the street with a cell

phone in front of a building of course

the thing makes mistakes but roughly you

know it

it gets an idea on what is happening and

this is again current technology so this

is what AI is currently capable of doing

and not only does it happen to don’t

does it apply to images does it apply to

videos but also it applies to

text-to-speech and roughly everything

that is around here so for those of you

who have been looking old movies here is

a movie that was considered as being

science fiction complete science fiction

back in the 90s

some of you will recognize it that’s the

time Terminator

okay just arrived on earth and he’s

looking for some things to some clothing

so he’s trying he’s analyzing all the

people that he sees because he you know

needs to find someone who that matches

his gait and apparently this guy is

matching it you know so probably this

this guy has clothes that he could put

on I need your clothes your boots and

your motorcycle so that’s the

you forgot to say please you forgot to

say please that was a wrong answer from

a human because if you look at what

happens to him next this is exactly is

not what he should vegetative sunset so

here

the Terminator was a autonomous computer

and what made it autonomous first of all

it was capable of analyzing what was

around it understanding what was

happening and what you see is that what

was total science fiction at this time

in the domain of recognizing what is

happening around it this is exactly what

we have right now all right now what is

missing in order to get an autonomous

computer it is the ability for a

computer to be able to find solutions on

so what is autonomy one of the

definitions autonomy is could be the

ability to solve a problem on your own

and it’s here that artificial evolution

comes in before speaking of artificial

evolution what is evolution first we

could define it in the following way in

a dynamic environment where multiple

coexisting species and individuals

evolve thanks to random mutations the

selection of the best creates some new

species that are emergent and that are

able to solve the problem of surviving

adapting to the current conditions so if

you look at what is happening in nature

nature is extremely inventive you have

some very improvable animals around here

one of those that I really like is the

opposite or proxy day it’s a fish that

is a deep-sea fish and what you may

think are its eyes are not its eyes is

the gills this is where it breathes is

two black spots in front and in fact its

eyes this is not a drawing by the way

it’s a picture

okay so living anymore it’s taken from a

video the eyes are the semi spheres the

green semiosphere

that are looking up okay and the fish is

looking up through its transparent skull

because it’s a deep-sea fish that is

hunting for other for for other fish and

typically it looks up it keeps looking

up in order to find through to see if

there’s a shadow of another fish that is

swimming just over it and if there’s a

shadow the fish would swim up the ice

swivel in front in order to be able to

catch its prey so you see that nature

can be extremely inventive in order to

find ways to adapt and and and for for

these animals to survive so the question

is can we use evolution for our own

purposes and if can could be could we be

implementing evolution into computers in

order to create artificial evolution so

this is what has been happening ever

since the 1950s the second timeline that

I described you start with a population

that you initialize and as soon as you

have a population of potential solutions

to the problem you just evaluate them

and this is what we call parents parents

are solutions that have been evaluated

you know whether they are good or not

now once you have this population of

parents well first of all you have a you

you try to see if one of them is solving

the problem if it’s the case then you

stop there and you pick up the best of

the population – that would be the best

solution to your problem but if not then

you select n parents and you have these

different parents which are potential

solutions to your problem you cross them

over you mutate them you have variation

operators and through those variation

operators you create a population of

children and once you have this

population of children you’ll evaluate

them then you get parents and children

that you reduce through a replacement

phase and well here you are you have a

tip you have evolution inside a computer

and once you get this well you get

creative computers this is these are

creatures that have been evolved by

calcium’s back in 1994 and the idea was

to find creatures that could

swim so the computer has blocks you can

assemble the blocks can move the blocks

around and the idea is is to have a

computer find a way to swim as

efficiently as possible now some of them

you see some some gates that that you

can find in nature that the computer had

been able to find on its own like these

you know looking like sperm or and then

this would be how to how you would run

on a surface that would not be on the

fluid anymore but on a solid surface but

then of course you these are very old

works you can see how the evolution

creates for instance a marching cube so

you start with a cube which has muscles

which has bones which has fat cells and

this is the result one of the resulting

gates that’s another way for a cube to

walk that has been evolved here again

the programmer didn’t program any of

these behavior it is the computer that

came up with the behavior but through

evolving steps so these are type protein

problems but what could how could we be

using them for real this is an example

with the st5 project at nasa by jason

long the idea now is that you have

satellites that are micro satellites

really small but because they’re really

small they don’t have much batteries

they are very they’re not powerful at

all and they have to transmit back their

data to the earth so either you have an

antenna that is a single wire in which

case you have an omni omnidirectional

antenna that emits through the whole

universe and just you catch a very small

part of the signal from the earth or

you’re trying to focus the signal

towards a radio telescope and if you’re

trying to focus the signal to a radio

telescope you you had rather focus it so

that you’ve got the exact surface of the

radio telescope that is listening to the

satellite so that you don’t waste any

part of the signal typically have such

little energy to transmit with the

problem is that if you use

the antennas hoody you gay models to

model this well you could need an

antenna that would be several meters

long or if you were using a parabola the

parabola would be much too large for the

very tiny satellites that you’re

creating so at some point a way to do is

to give artificial revolution the

Maxwell equations to give a an antenna

simulator to the computer and say okay

please can you evolve an antenna that

would be small enough to be compatible

the size of satellite but in that valid

nevertheless would be able to focus

enough the data towards the radio

telescope and after I think it was three

weeks of massively power computing using

artificial evolution this is the result

that was obtained by computer scientists

not antenna engineers all right

because we can’t see anything that was

what was found by the computer and

what’s very interesting is that it is

the computer that found this shape very

odd very strange shape very complex

shape and when this was shown to the

antenna engineers they said what is this

computer scientists that well we don’t

know the computer found this with your

equations with your simulators right and

what they did there then this of course

tests the antenna into an anechoic

chamber it worked perfectly and now they

constraints and the computer came up

with another antenna that was different

or maybe a third antenna and the

computer keeps coming up with antennas

that are totally improbable and when you

just plug them on they’re working

perfectly well in the end so NASA had

made a very small movie to deploy the

est five mission and I think that they

had at some point they said it was the

first time a nonhuman technology was

sent into space why is it non-human

hasn’t been designed by a human has been

designed by a computer and you’re not

able to reproduce it it’s only computers

that can do it

so the idea now is that if you imagine

that artificial intelligence allows

computers to spot out problems because

it knows what is happening around it

then if it spots out a problem has to

solve in order to complete this mission

then computers could use artificial

evolution to find a way to complete the

mission to be inventive to find a new

solution and so the the next revolution

to come in computer science could be to

a couple artificial intelligence and

artificial evolution in order to have

computers that are autonomous now of

course there could be unexpected

problems and that was one expected

problem it should have been zooming on

the on the bottom part of it and because

now we need to think of a bit of

philosophy and I have three sentences

that were supposed to be printing just

on the bottom of the screen the first

sentence sentence comes from a French

writer who said that scientists who

should really rethink about what they’re

doing and the real sentence I just

forgot it because it was just written

there well I’m not going to be

remembering it it is possibly a French

writer who said that science is but the

rim of the soul if you’re not looking at

what you’re doing now the second

sentence was a very interesting one

because it came from a very renowned

scientist who said well if we knew what

we were doing we wouldn’t be we wouldn’t

be calling this research would we and so

who has been saying this sentence if we

knew what we were doing we wouldn’t be

calling it research it’s just

Einstein who said that and finally there

is another philosopher lavished horse

who said that scientist is not here to

give answers the scientist is here to

ask the right questions so finally we

are now in a stage where we can create

autonomous computers this is what we are

doing in my team in Strasbourg

University but it is we need to be very

wary of what we are doing because these

computers can make mistakes and if they

do make you make mistakes then they they

are autonomous they can

think of new solutions in order to

circumvent the problems and they will do

it and the the last there was a very

small video which was taken from 2001

Space Odyssey where you had Hal the

computer that was flying the ship who

computed that the in order to fulfill

its mission there would be more

probability for the computer to fulfill

its mission without the human crew than

with the human crew so at some point in

the movie the computer decides to do

away with the human crew because it

computed that the probability for its to

completes mission would be better

without the humans so it proceeds into

eliminating the humans so you see that

computers can become very inventive and

so now the very last sentence by a great

philosopher who is sharpen our is that

the task is not so much to see what not

what no one has yet seen but to conceive

what no one has yet imagined but that

everyone sees so I let you ponder on

this thank

[Applause]

[Music] merci beaucoup ejaculate

Please follow and like us: