Convolutional Neural Networks [Part 1] - Introduction Video Lecture | Machine Learning with Java - AI & ML

hi explorers and welcome to another
video and today we're going to talk
about convolutional neural networks and
convolutional neural network is
something that we use to identify things
and it's often used in image recognition
but could also be used to find patterns
in really large data sets but as an
exercise it's easier to explain it with
images so that's what I'm going to do
and to start off a convolutional neural
network could be seen as something
modeled on our own brain because we have
a lot of neurons they take stimuli from
I know hearing or our sight or taste and
so on and by these different signals we
will decide on a specific view on the
world and we will understand the world
from the different senses that we get
and in a neural network you can take an
image like this one it's a very
simplified image I try to draw on e so
you've seen the ones here is in a
pattern of an E and then you have all
the zeros non-activated areas around it
and the neural network will look in this
image for patents so let's say that we
look at the little bit smaller piece of
the image and then it look at this part
here then we will see some kind of an
arch or a curb or perhaps even a two
straight lines that was found in this
part and if we then take and stride one
over so we will look at
part from here so we look at this
rectangle then we will see a longer part
but a much fewer and zeros around it
so we will identify this part it's
something and then we take one piece
over more and we'll see that the line
ends over here so here you have just a
line record all of these results that
you see and then we move down and and
look at this part over here and so on
we go through all this and that what I
explain now we take a stride of one we
say and that will give us a lot of
information about this image that we
could put into a new network so let's
say that we found this here and we
call this thing that we found a one
let's say that that's the one figure and
when we come to this box here we see
this here and that's a two figure and
then we have this line over here that's
a new figure 3 and then when we come to
this this is something new so that's a 4
and we go on so and identify different
figures and just give them a number or a
specific kind of filter that could
recognize this type of thing that is so
we don't know what it is yet but we can
recognize that kind of a thing again or
that shape again that's the important
part and so when we have this network
that could recognize specific shapes we
could use this for it for instance to to
recognize a little bit more different
shape so let's say we have a circle and
we put it into four parts then we can
say that with all these strange patterns
that we found we can say that
okay we have a lot of patterns here that
looks somewhat like an arch with this
curvatures and then we have this arch
over here that looks something like this
with this kind of an angle and we have
an arch over here with this kind of an
angle and so on so we will identify
these arches as as different entities as
well and let's say that this image is
much more a very more a lot more
intricate so I will try to draw this
strange shape here I wonder what the
computer will say about this strange
shape here we have the circus we talked
about earlier that we had it identified
and we will probably identify something
up here as some kind of shape with a
line in this area and so on so we will
have a lot of filters that find
different objects in an image so let's
say that with the first part of our
neural network we had a lot of neurons
that found ones and zeros in one part
and then we have another filter that
found them in another one and these are
interconnected to new neurons that well
that they will give weight to and these
weights will tell it that it's either a
circle or two straight line or it's a
it's some kind of this curve over here
and this will tell new neurons that well
this kind of curve over here is a part
of the circle and this curve over here
that's a part of a roof or whatever and
then we go even further to another layer
over here a few more neurons and those
will say that while this is a back part
of something and this is front part of
something and so on and then we
last in this neural network we will
have some neurons that are fully
connected so every one of them have
weighted pairs between them because the
first of these is a lot of filters a lot
of filters to find things and then we
slim our network down so it will be
lighter and lighter can make even more
neurons over here to illustrate this bit
more that it's a lot of things in this
end and then we'd slim it down and slim
it down until we get to a point where we
have a fully connected network where the
output that we know that we have let's
say four outputs it could either be a
car a dog a cat or a zebra let's say and
we know over here that it's not a dog
probably not a dog it doesn't look at
like a dog zebra now doesn't look like a
zebra I can say is it the cat now
probably not the cat either so it must
be a car and in our training set we
have a lot of pictures of cars so we can
definitely say that this is a car and
when we have found that this is a car we
do as we did in our previous examples
with the networks that we go back from
this point here and say this is car this
is not a dog it's not a zebra it's
not a cat and then we change the weight
backwards so this fully connected
network will change all the weights to
weight more at the car for this specific
image and then we will go back and do
the same for all the other networks in
our series until we get to the point
where we have the weights from the first
neural network that found shapes and
often
when you do image recognition you have
some neural networks that have already
been trained with a lot of shapes and a
lot of things so often these layers are
so good at finding circles finding
arches and so on so these layers
seldomly change but the later layers
will change because your problem domain
is a bit different than anybody else and
then we have a little bit of a twist
here we we could get a network that is
great at finding cars it's so good at
finding cars that when it sees a
specific new zebra that it's never seen
before it says that it's probably could
be a car but I can't say that it's a
zebra least and we can't find a way
to get it to a zebra result because
it's weights too much to a car result
then this is probably over fitted it's
too good to find cars and it could be
bad at finding other things so then you
can have a neuron in here somewhere
that's also connected with weights and
you say that this in neuron it has no weight
so from this neuron it's all
always 0 and in this net that we have
over here that a dropout net that will
drop some of the neurons out at random
just to remove some of this overfitting
results and so the neural network be
good at finding things but not too good
so it can't recognize new images that
you gave it so that's something that you
always have to keep in mind that you
could train a network to be really good
at finding things but if it gets too
good at something it will not figure out
a result
for things that it haven't seen before
so that's pretty much a simple
explanation now convolutional neural
network will work i will go into more
detail on different part in later videos
but this is a high-level overview for
introducing the concepts if you have any
suggestions of things that we'll talk
about when it comes to convolutional
neural networks please tell me and in
the comments and I hope you found some
interesting concepts in what I talked
about and that you learned something and
I really hope to see you in the next
video