So welcome back to our deep learning lecture and today we want to talk a bit

about visualization and attention mechanisms. So we want to figure out, on

the one hand, communicate with other researchers, how we can visualize things

like for example the network architecture,

but we also want to look into different visualization techniques to understand the training better.

We've already seen examples for this,

but we will also iterate through this and see how we can try to understand what's happening during the training.

Then we want to visualize the parameters themselves,

so the state of the network or what the network has been learning,

so we want to figure out what's going on in the network,

and we will start with very simple techniques for parameter visualization,

that is essentially by looking at the parameters themselves,

but we very quickly realize that it's not sufficient,

and we come up with different ideas how to figure out what's happening in the network.

Then there's gradient-based methods that are essentially trying to visualize the gradient in the input domain,

which is also a very interesting technique,

and it also helps you to understand what the network likes to see

and what has a high influence on the decision outcome of the network,

and then we will talk about parameter visualization via optimization,

and this really then allows us to create dedicated inputs that are suitable to figure out what a specific neuron

or a specific output network node likes.

Then in the end of this lecture, because it's related, we will talk about attention mechanisms,

which will then be used in order to focus towards a specific input sequence.

Okay, so why do we need visualization?

There's this common criticism about neural networks that they behave like a black box.

You train them. They do what they're supposed to do on the training data set,

and if you have training and test data set stemming from the same distribution,

then you kind of have a high likelihood that the network will also behave as expected on the test data set.

But generally, it would be really interesting to understand what's happening in this black box.

So we would like to see what's happening in there and to understand why a specific outcome is actually produced.

So this is the main reasons why we want to look into the visualization,

and of course also not just to understand what's happening,

but also how to communicate what's inside the black box to other researchers or to other people.

So an incomplete list, why understanding matters, you want to communicate with others.

If you want to show, look, we have this network architecture.

We already used several visualization techniques when we were talking about the architecture,

so we implicitly also used different techniques here.

Then we want to identify issues during training, like we can't find convergence, we have dying reluces and so on,

so we want to figure out what's happening here.

So this is why we need visualization of the training process.

Then we want to identify problems in the training and the test data,

and you will see the neural networks or the optimization-based algorithms that we have here,

they really like to take the shortcut, like the simplest solution that you can think of to solve a problem,

and sometimes this is actually not what you intended to do.

And of course we want to understand how and why the networks learn,

and we have some ideas on how to understand this a little better.

So there's these three main types of visualization, the architecture, the training, and the learning parameters and weights,

and we want then to visualize the representation of the data in the network.

Okay, so let's start with the first part, that is the network architecture visualization.

So this is of course useful to communicate architectures efficiently,