An overview of CNNs

In this lesson we will look into convolutional neural networks.

You’re already familiar with the M.A.

data set.

So it will be easier to describe a typical CNN in our approach which was simply feedforward neural networks.

We flattened the images into a vector of length 784 that way however we lost the spatial information

of every pixels neighborhood of pixels.

For example the twenty eighth pixel on the 29 pixel in the vector we use are next to each other in the

vector right.

Well in the picture they are actually quite far away.

Furthermore since we always start by taking a linear combination of the input it will search for particular

digits in particular places.

For instance a seven written in the top left corner is very different from a seven written in the bottom

right corner of convolutional neural networks solve this problem by dealing with the original 28 by

28 photos without flattening them.

Instead it applies Tiny say five by five so-called kernels to every possible position of the image.

Kernels are like wait’s.

So if we start from the top and continue down we can count a total of twenty four five by five squares.

The same applies if we go from the left to the right side.

Therefore the next layer is 24 by 24 which is the number of 5 by 5 matrices.

This is called convolution layer that we get is called the convolutional layer.

The number of Purnell’s you choose is a hyper parameter and you are not restricted to a single kernel

apart from convolution.

There is another main step in CNN’s Poulenc.

Most commonly we would divide these 24 by 24 squares into multiple two by two squares without overlapping.

And will take the largest number and the two by two Matrix because we assume it is the strongest detail.

That’s how we reduce the dimensionality of the problem.

What makes the issue slightly more complicated though is that most photos have colors which implies

that images have height and width but also color or depth.

This is a third dimension with three variables according to the G-B scheme.

Red green and blue.

Thus our original photo was not of shaped 28 by 28 but rather 28 by 28 by 3.

That’s also where the whole tensor approach fits perfectly.

So the initial 28 by 28 Matrix was not reduced to 24 by 24 but was actually converted in a 28 by 28

by three tensor which was then reduced to a 24 by 24 by three plus tensor depending on the number of

kernels we chose to use.

If we convolute and pool for long enough we can reduce the dimensions to a vector containing one hot

encoded categories like dog cat horse and so on.

In order to truly get CNN’s We need much more time than a couple of minutes.

In general though to get the gist of it CNN’s are mainly used in image recognition due to their two

major advantages spatial proximity are preserved.

It matters where in the photo we find a certain detail and certain details such as a human eye is looked

for everywhere in the photo.

Thus a face can be recognized everywhere on the image.

These two advantages make CNN’s predictive power much higher than that of an end especially when it

comes to image related problems.

Instances a robot version self-driving cars Facebook tagging and Apple’s face recognition for unlocking

the iPhone.

A recent application of CNN’s for a non-image related problem was the Google assistant technology developed

by deep mind.

This is a competitor product to Siri Alexa and Cortana.

Allegedly they managed to apply CNN’s In a way that converts artificial sound in a more human like shape

than ever.

You can find the original announcement by deep mine in the course resources.

All right.

How likely is it that you’ll use CNN’s.

Well currently image related problems are mostly used by companies such as Google Tesla Apple Microsoft

Amazon and tech startups.

The applications of CNN so far seem more or less out of reach for most people.

That’s the reason why we decided to focus our efforts on creating a comprehensive course with a focus

on deep neural networks.

That’s a great value you can add to any company you work for which CNN don’t really deliver as they

have a different focus.

OK.

And the next lesson we will make a quick overview of our own ends.

Thanks for watching.