OK, hello, good evening everyone, welcome back to the lecture series.

I thought maybe I start by giving you a kind of review of where we have come during the

course of the lecture series.

So this was the mountain where we, so to speak, learned how to implement a neural network

and how to do backpropagation in order to train it.

And since then, we've looked at a series of applications.

So the first was just to show that we can represent,

basically, arbitrary functions, including an arbitrary image

in terms of a neural network.

And we said we can use neural networks for classification.

So you give me an input like an image.

I tell you what it means if I have

seen many training examples.

Then we said that for the particular purpose of images,

it's actually good to exploit translational invariance.

And you can build a neural network

that uses many, many fewer weights in its construction.

These are the so-called convolutional neural networks.

Then we went to something that is no longer supervised

learning, but which just tries to encode as well as possible

given input examples.

So these are the autoencoders.

Then we had another brief remark about how

to do unsupervised learning in the sense of trying

to visualize high-dimensional data in a meaningful way

that would give you, for example,

clusters that you can recognize of data points that

are similar.

And so that was dimensional reduction, the TSME method.

And then we went to look at more sophisticated networks.

And the first one was recurrent networks, that

is networks with memory.

So that is when neurons also depend

on what has come before in terms of the input

and in terms of their internal memory state.

Then we briefly mentioned word vectors.

So that's just useful for a particular application

of recurrent neural networks when

you want to analyze a full text, which can

be seen as a string of words.

But how do you represent these words?

You can just label them in a big dictionary.

That would be the very basic version.

But you can also invent these so-called word vectors

where different components of the vector

have different semantic meanings.

And then you can even do calculations

with such word vectors.

And then very recently now, we started

by looking into the big subject of reinforcement learning.