Press "Enter" to skip to content

Learning from the brain for the technologies of the future | Bart ter Haar Romeny | TEDxEindhoven


[Music]
[Music]
like golf moments that’s what it’s all
about and a light bill of course at this
city of Philips as a medical meeting and
a light belt means a bright idea and
ideas where do they come from they come
from the brain that the brain is today
actually being discovered in a way that
you won’t believe we can look inside the
brain at incredible resolution we can
see what the brain is doing we have deep
learning etc and that’s what I will talk
a little bit about today what is in this
brain happening I’m working at the
University and my field is pattern
recognition
I designed with my whole team automatic
software to find tumors patterns in lots
of medical images at the best pattern
recognition machine is this guy
that’s an amazing pattern recognition
how do you design all this and I said
okay let’s be inspired by this brain
let’s see what’s in there
so insight is this magic thing Eureka
and I want to take you on a journey what
do we know how can we learn from this
how can it be how can it inspire us so
let’s start with some facts of our brain
it’s incredibly small it’s only about a
kilo 2% the far away that used a lot of
energy 20% of our energy and it has an
incredible amount of neurons 86 billion
at the nerve fibers they’re all
connected to each other all these nerves
spiders wire a thin micron thin all the
wires together is bordered to the moon
and back the connections between these
neurons are called synapses we have
amazing amount of shine obsessed and
it’s an computer outperforming all our
other computers they have calculated
it’s about one petaflop
floating-point operations per seconds
so let’s put a little bit more in
perspective if you see the trees in the
Amazon forest and the neurons in your
brain
it’s about the same number if you look
at the neuron it has connections one
neuron is connected to 50,000 other
neurons so that’s quite an broad spread
out and if you look at these sign lapses
and we have trillions of these synapses
how do we learn if you use a synapse
many times it gets bigger and if you
forget it gets smaller so by repetitive
action you learn so we can also now see
how this learning can be done and one of
the amazing feat of our brain is that
this guy is only using 25 watt and
that’s quite stunning and if we look at
our computers today we look at these
Google data centers and so on and all
these computing and Amazon computing
centers they use megawatts they have
cooling towers on top and that’s a waste
of energy they’re just incredible so we
should learn so much more from this this
brain so how is it wired
how’s it work what do these functional
circuits actually do and we have a
community now in the world that’s really
focusing on can we mimic this brain use
this incredible computing algorithms and
then apply them in our clever technology
I go a little bit deeper in that let’s
first tell it the brain how can we
measure this well the first of all is
and that’s me and one of the biggest
companies in the world making all these
fantastic machines is right here
it’s Philips it’s in the top three they
make MRI machine CT machines etc we can
now see the brain and incredible
resolution of resolutions now about a
quarter of a millimeter so this is
called an MRI scan and this is my own
brain slice by slice you can see every
quarter of a millimeter in this brain
you see in the middle the rested spinal
cords connecting all these fibers to my
muscles to my heart so everything but
this is not only we can see we could
also see function and how can we see
function if a brain cell is active
suppose I do like this then my motor
court
is active and these cells they generate
action potentials they need oxygen they
need food so they give a signal to the
small capillaries to the small blood
vessels open and bring me oxygen an
oxygen can be seen on an MRI scan so you
first do nothing you make a scan and
then you do this and if you take the
difference of these kinds you see a tiny
difference and now you see this boy is
looking at the screen he’s using his
visual system and the back of our brain
is for vision and you see that’s quite a
lot this for vision a quarter of our
brain is for vision and if you see how
many nerve cells we have so how many
filters how many calculations are done
vision is an incredible thing we are
purely visual machines because images
came on the internet it exploded we do
nothing else than look at images because
our brain is made for images we are
images machines so we have functions we
can do more we also look at connections
how can you measure connections in the
brain well they found some years ago
that you can measure in an MRI scan the
very tiny motion of water molecules they
make the magnetic field to the left a
little bit different than to the right
and if you move a little bit they change
the magnetic field will change in
frequency and if water is just
surrounded by water it goes in every
direction you see a little bit like a
sphere but if water is in a nerve fiber
it can only move along the nerve fiber
and not so much the other way so if you
look for some time you see you kind of
an ellipsoid well the long axis of the
ellipsoid is of course the direction of
the fiber and what you can do in every
voxel you measure these millions of
foxholes you connect them and then you
get the wires in my group we studied
this for a long time because we really
wanted to see how is this whole brain
Wyatt
losses passed before neurosurgeons they
need to see what are the major tracks if
you remove a tumor you don’t want to
make the patient blind or paralyzed you
need to prepare all this so what we did
is we made fantastic
visualizations we program this in the
GPU you see the fibers here connect
the left and the rights in the brain the
colors are X Y Z is red green blue so we
have some color coding as well this
cannot be done for all the areas in the
brain how we here how we see how we move
so we have now the the wires of power
electronic electronic schematic and the
visual system is the most study Teufel
and I’m going to go a little bit deeper
in the visual system they made all these
measurements and they looked how does it
work and it turns out they’re in stages
about 11 stages these stages have been
named the primary visual cortex a stage
number one that’s in the back of our
head it’s called v1 then v2 v3 v4
they’re connected you see the
connections and these layers we can move
these layers by looking at that’s model
one neuron and then model thousands of
neurons and millions of neurons and the
model egg goes by modeling these
synapses and these synapses are
connected by what we call weights in a
neural network everybody has heard about
neural networks artificial intelligence
is exactly what we do what we actually
do is we mimic all these connections
what happens in the brain and people
made these neural networks you see a
three layer neural network and what you
put in this network has to be a
classifier you put in lots of images and
the network has to say what is it it’s
the cat this the dog is it the horse but
these little networks were actually for
years not working so well it was
terrible it was 75 maybe 80 percent and
that’s not good enough
so we struggles and we have major
competitions in 2010 we had these big
yeah gruel images coming up we had
access to millions of images and people
started to make competitions they said
ok sending your program we have here 1.2
million images 1,000 classes is that the
horses at the beach is this building and
people were in about 2010 75% correct
today we are 96% because what happens it
was about 2012 people had a major
invention and actually that invention
was you should not take three layers in
a neural network but you shake should
take many many more layers more not only
10 maybe go to 20 you maybe go to 100
and that’s called deep so that’s what we
call deep learning we can train these
networks and this network was yeah what
is the network like this doing stage one
is doing very simple things if there’s
only filters to find some lines some
edges very primitive things the second
stage is doing some more complex things
so these lines are now triangles so it
finds noses eyes and so on so parts the
third one finds objects finds a face and
so on so you get more contextual areas
if you go deeper in the brain
what made it especially happen well we
have now really amount of enormous
amounts of data we don’t talk about
millions of images we talk about
trillions of images so that’s what we
call big data we also have GPUs graphics
processing units that’s actually game
cards people wanted to play games and if
you play games you have to generate all
kinds of 3d artificial environments and
the computers were not fast enough so
they said ok let’s make many small
parallel computers one computer per
pixel and then we can generate thousand
pixels at a time and that’s exactly what
everybody wanted to have so it costs
nothing the GPU is in every computer now
and it’s actually a small computer with
1000 processors in there and it costs 52
to a few hundred euros it’s made here by
a SML they make the machines who make it
so that’s now available and the third
step is of course we realized that we
can have deep neural networks and
there’s a huge community now it’s
exploding and the trick is of course to
the upper
you see the visual system in the middle
of you see these all these layers these
layers and we see an increasing context
but let’s go a little bit deeper into
the brain what did they find in vision
if you cut the brain you see that this
actually it has layers it has lots it’s
incredibly well organized not a single
neuron is on the wrong place that every
neuron is used if you don’t use in their
own it generates so people did lots of
experiments they stuck a needle in v1 v2
v3 and what is the cell seeing and there
so indeed if you stick a needle in these
very primitive first layers v1 you saw
these very simple filters coming up it’s
exactly what’s happening there so in the
first layer we do these primitive things
and in mathematics we say ok we take
derivatives the first derivative is
actually the change in intensity
these are edges second derivative how
corrupt is it so we can do mathematics
and the brain is actually doing a lot of
mathematics here we can also look at how
is this formed in the brain how do these
receptive fields as we call them these
filters how did they originate and it
turns out you are not born with these
filters if you’re born every neuron is
connected to every other neuron in its
neighborhood so how can it get organized
it organizes itself the first three
months after birth you make them so
that’s quite exciting so we said let’s
look at an image and take a monkey and a
million years ago or did the monkey see
a forest so we took the first image you
take some very tiny patches from the
forest and you do some statistical
mathematics what is the best basis for
all these patches and you find exactly
the same filters so you can learn the
filters from the data and that’s exactly
what we do so I said suppose I’m an
animal and I live in a restricted world
I have only vertical lines it turns out
that I’ll you got very specific filters
restricted filters so nothing happens if
I go up done but the lot happens in this
direction so I
the right filter so the lesson is that
handcrafted filters are out for many
many years me and all my colleagues we
worked with let’s design a lot of
filters and do pattern recognition and
so on that’s over we learn them from the
data we regulate the visual environment
by keeping kittens in complete darkness
from birth this is an experimental kit
for periods of controlled visual
experience inside special special
cylindrical chambers this kitten
they did the experiments on how was it
formed after birth so the kitten was
raised in the dark and for three months
the only thing it could see was the same
as I did with the trees horizontal lines
can you turn off the sound from the next
slide because let’s now see what the
kitten has seen if the kitten sees a
horizontal stick it perfectly sees the
horizontal stick and of course you
expect it if there’s a vertical stick it
has no filters in its brain to see a
vertical stick it cannot see it anymore
so that’s quite striking this is what
biology does this is what’s called brain
plasticity adaptivity we learn from what
we see well of course this kitten went
to the family again and after that it’s
a fully learnt it it’s all these things
the images that we see most are faces
and you know you look the whole day at
faces Facebook the way it even turns out
that they have now found it in all kinds
of experiments that we have a dedicated
area in the brain that is just
specifically meant for faces it only
reacts on faces you give it anything
else it doesn’t react but there’s a face
but there it goes I wish he faces now
everywhere even in the most stupid
things we see faces so and why is this
you look at Facebook the whole time but
it’s one thing you don’t see and just
just like the kitten the face upside
down you have never seen so you don’t
recognize it you have the same as the
kitten even if you see a picture of a
well-known person your mother whoever
it’s upside down you have to look a
second longer because we have no trains
regions for faces upside down okay let’s
go to some applications of all this what
we learned especially with this deep
learning one of the big breakthroughs is
of course the self-driving car until it
is very active in this area and why is
it actually now becoming possible we
have already cars that drove a million
kilometers because they need to see
passengers traffic signs all the
different things and this deep learning
this recognition
you remember 96% is okay is that really
working so it’s feasible it’s coming
right there we have in Holland the
largest flower production in the world I
think we make about 40% of all the
flowers nine billion roses every year
and it’s the we have 450 roses some are
very expensive is the right throws on
the pallets does the farmer say okay
that’s the expensive rose or is it an
cheap one so we have now computer
systems who check everything
and my own field is medical imaging
that’s a huge area it big
80% of all the diagnosis is done on
images and if you make images you make a
CT scan you easily make 1,000 slices so
which doctor is going to look at 1,000
slices the doctor needs to be helped by
these deep learning pattern recognition
systems and they work they do an
incredible job he I’m project Leda now
of a large screening program that’s
preventive medicine for diabetes and
with diabetes your blood vessels start
to leak you can’t see them you can’t put
everyone in an MRI scan many people may
have diabetes but they don’t know but
your retina your eye you can see that
very well and there you can see with the
incredible resolution you can see the
blood vessels and in China they have an
enormous problem which that I be this
11.6% so our university is oval
university is asked can you help with
this huge problem and finds in our
population who has doubled the amount of
diabetes from elsewhere to find them so
we’re now running a program finally we
hope to screen 24 million people and our
software can now automatically detect
the beginnings of small leakages and so
on and all things of course prevent
blindness solace concludes the brain
it’s our next revolution actually we
have just started we know nothing yet
about the brain we begin to see it but
we had a major leap a couple of years
ago the disruptive change because of
these deep neural networks we will see
this everywhere now it’s coming up
radiologists are going to use it
everyone who use computer inspection
will use it and so on it will be one of
the major descriptives so you saw some
examples and actually now we have a
little bit better understanding them
were the creativity were it’s all
emerges it’s in our brain our light bulb
thank you very much
you
[Applause]
[Music]
Please follow and like us: