Welcome everybody to our deep learning lecture and today we want to talk a bit about visualization

and attention mechanisms.

Okay so let's start looking into visualization and attention mechanisms.

So we'll first look into the motivation, then we want to discuss network architecture visualizations.

Lastly today we want to look into visualization of training and training parameters that you've

already seen in this lecture and in the next couple of videos we want to talk about actually

the visualization of the parameters, so the inner workings and why we would actually be

interested in doing that.

Lastly we will look into attention mechanisms, so this will be the fifth video of this short

series.

So let's talk a bit about the motivation.

Well why do we want to visualize anything?

Well of course the neural networks they are often treated as black boxes, so you have

some inputs then something happens with them and then there are some outputs.

And today we want to look into how to communicate the inner workings of a network to other people

like other developers, other scientists and you will see that this is an important skill

that you will need for your future career.

So well a couple of reasons why you want to do that.

You want to communicate the architectures, you want to identify issues during the training

like if the training doesn't converge, if you have effects like dying relu's, you want

to identify faulty training or test data and you want to understand how, why and what networks

learn.

So there is three main types of visualization that we want to cover here.

This is the visualization of the architecture, the visualization of the training and the

learned parameters and weights and this is then important of course for visualizing the

representation of the data in the network.

But I don't think anybody believes that layer 150 of the ResNet is a grandmother cell and

you know layer 100 is contours or something like that.

So let's start with the network architecture visualization.

Here we essentially want to communicate effectively what is important about this specific type

of neural network.

The priors that we actually impose by the architecture may be crucial or even the most

important factor for good performance of a specific network.

So mostly this is done graph based structures with different degrees of granularity.

You will see some examples in the following and actually we've already seen this quite

often if you compare to our set of lecture videos on neural network architectures.

So there's essentially three categories.

There is the node link diagrams that work essentially on neuron level where you have

nodes as neurons and then weighted connections as the edges and you've seen them especially

in the early instances of this class where we really go down onto node level and where

really all of the connections matter.

They are for example useful if you want to show the difference between a convolutional

layer or a fully connected layer.

So this is essentially important for small sub networks or building blocks and there's

different variants with explicit weighting, recurrent connections and so on.

If we want to go for larger structures then we use block diagrams and there we have solid

blocks and they then often share only single connections between the layers although actually

all of the neurons are connected with each other.

We have seen plenty of these visualizations and here you can see a visualization for the

block diagrams.