Deep Learning with Tensorflow - Introduction to Convolutional Networks

Hello, and welcome!
In this video, we will provide an overview
of the convolutional neural network model.
Convolutional neural networks, or CNNs, have
gained a lot of attention in the machine learning
community over the last few years.
This is due to the wide range of applications
where CNNs excel, and you can see a partial
list here.
In a later part of this video, we're going
to focus on the object recognition problem,
to get an idea of how a CNN's structure allows
it to extract the elements from an image.
So for example, how could a CNN learn to pick
out a person, a dog, and the chairs from the
image, even if the chairs are only partially
visible?
This is not an easy task for computers, and
it took years of dedicated research to achieve
this.
Historically, the goal of machine learning
was to 'Move Humanity closer to the unreachable
General Artificial Intelligence'.
But not surprisingly, this goal ended up being
lofty and difficult to attain.
So what scientists started doing was developing
a series of models and algorithms that excelled
in specific tasks.
One of these tasks is object recognition,
but even this can be difficult without the
right foundation.
So the original goal of the CNN was to form
the best possible representation of the visual
world in order to support recognition tasks.
The CNN solution needed to have two key features.
It needed to be able to detect the objects
in the image and place them into the appropriate
category.
But it also needed to be robust against differences
in pose, scale, illumination, conformation,
and clutter.
The importance of the second feature can't
be overstated, since this was historically
a huge limitation of hand-coded algorithms.
When faced with a difficult problem like this,
it's sometimes helpful to look at a working
example in nature.
The mammalian brain is a great example of
an object classifier, and as it turns out,
we can mimic the functionality of the brain,
at least partially.
But how does a mammalian brain accomplish
this?
It starts with the photoreceptors in the retina
receiving information from the outside world.
The primary part of the visual cortex, or
V1, uses simple and complex cells to start
processing the input.
V4 identifies different textures, and the
inferotemporal cortex puts everything together
to recognize the object being observed.
The main idea is that the process has multiple
layers, different types of cells, and increasingly
complex functionality.
We can use this as inspiration to develop
a high-level approach to the object recognition
problem.
We start with an image, we extract a few primitive
features, we combine the features together
to form the parts of the object, and then
we combine the parts together to form the
object itself.
The Convolutional Neural Network was developed
from this sequence of steps.
Let's walk through this process with a real-world
example to get a better idea of how this all
works.
Take a look at this image here.
Let's say we wanted a network to be able to
recognize one of the buildings.
It first helps to carefully consider how the
scene is organized.
At the highest level we simply have an image
of the city.
But we can start to extract oriented edges,
we can join them to form doors and windows,
and then put those parts together to form
a building.
It's important now to remember how our biological
system would process this image.
Like we discussed before, we'd start by taking
in the image via the retina, and then allowing
increasingly complex parts of our brain to
extract features and ultimately form an object.
So now that we see the biological system's
method for recognizing objects, we can briefly
discuss how a CNN might use this type of processing.
Consider the building in the lower right hand
corner of this image.
This is the raw image that a CNN would receive
as input.
Some of the best primitive features for a
building would be horizontal lines, and vertical
lines.
By combining these lines together, the net
can start to form the component parts like
windows and the building's external shape.
It can then use these basic parts to form
the complete object.
So at this point, you should have a basic
understanding of the motivation behind a convolutional
neural network, and how they might be used
in an application.
Thank you for watching this video.