Welcome back to deep learning. So today we want to talk about unsupervised methods and in particular we will focus on auto encoders and GANs in the next couple of videos.

And we will start today with the basics, the motivation and we will look into one of the rather historical methods that is the restricted Boltzmann machines.

We still mention them here because they are kind of important in terms of the developments towards unsupervised learning.

So let's see what I have here for you. So the main topic as I said is unsupervised learning and of course we start with our motivation.

So you see that the datasets we've seen so far they are huge. They had up to millions of different training observations, many objects and in particular few modalities.

Most of the things we've looked at were essentially camera images. There may have been different cameras that have been used but technically typically only one or two modalities that were in one single dataset.

However, this is not generally the case. For example in medical imaging you have typically very small datasets, maybe 30 to 100 patients.

You have only one complex object that is the human body and many different modalities from MR to X-ray to ultrasound and all of them have very different appearance,

which means that they also have different requirements in terms of their processing. So why is this the case?

In Germany we actually have 65 CT scans per 1000 inhabitants and this means that in 2014 alone we had 5 million CT scans in Germany.

So there should be plenty of data. Why can't we use all of this data? Well, these data are of course sensitive and they contain patient health information.

For example if you have a scan that contains the head in CT then you can render the surface of the face and you can even use an automatic system to determine the identity of this person.

There are also non-obvious cues. So for example if you have the surface of the brain, the surface is actually characteristic for a certain person and you can identify persons by the shape of their brain to an accuracy of up to 99%.

So you see that this is indeed highly sensitive data and if you share whole volumes people may be able to identify the person, although you may argue that it's difficult to identify a person from a single slice image.

So there's some trends to make data like this available, but still you have the problem even if you have the data you need labels.

So you need experts who look at the data and tell you what kind of disease is present, which anatomical structure is where and so on.

And this is also very expensive to obtain. So it would be great if we had methods that could work with very few annotations or even no annotations.

So I have some examples here that go in this direction. One trend is weekly supervised learning. So here you have a label for related task and the example that we show here is the localization from the class label.

So let's say you have images and you have classes like brushing teeth or cutting trees, then you can use these plus the associated gradient information like using attention mechanisms and you can localize the class then in that particular image.

So this is a way how you can get a very cheap label, for example, for bounding boxes.

There's also semi supervised techniques where you have partial or very little data of labels and you try to apply it to a larger data set.

So a typical approach here would be things like bootstrapping. You create a weak classifier from few labeled data, then you apply it to a large data set and then you try to estimate which of the data points in that large data set have been classified reliably.

And then you take the reliable ones into a new training set. And with the new training set, you then start over again, train a new system and you iterate until you have a better system.

Of course, there's also unsupervised techniques where you don't need any label data and this will be the main topic of the next couple of videos.

So let's have a look at label free learning and one typical application here is dimensionality reduction.

So here you have an example where data is on a high dimensional space. We have a 3D space actually, but we're just showing you one slice through this 3D space and you see that the data is rolled up and we identify similar points by similar color in this image.

And you can see that this is actually a slice through a 3D space and the manifold that you see here is often called the Swiss roll.

Now, the Swiss roll actually doesn't need a 3D representation. So what you would like to do is automatically unroll it.

And you see that here on the right hand side, there the dimensionality is reduced. So you only have two dimensions here and this has been done automatically using a manifold learning technique or dimensionality reduction technique that is nonlinear.

And with these nonlinear methods, you can break down data sets into smaller dimensionality.

Now, this is useful because the smaller dimensionality is supposed to carry all the information that you need and you can now use this as a kind of representation.

And what we'll also see in the next couple of videos is that you can use this, for example, as network initialization.

You already see the first auto encoder structure here that you train such a network with a bottleneck where you have a low dimensional representation.

And later then you take this low dimensional representation and repurpose it.

This means then that you essentially remove the right hand part of the network and replace it with a different one.

Here we use it for a classification and again our example is classifying cats and dogs.

So you can already see that if we are able to do such a dimensionality reduction and preserve lots of the information in a low dimensional space,

then we potentially have fewer weights that we have to train in order to approach a classification task.

And by the way, this is very similar to what we'd already discussed when talking about transfer learning techniques.

You can also use this for clustering and you already seen that we are using this technique here in the chapter on visualization,

where we had this very nice dimensionality reduction and we zoomed in and looked over the different places here.

And you've seen that if you have a good learning method that extracts a good representation,

then you can also use it to identify similar images in such a low dimensional space.

Well, this can also be used for generative models.

So here then the task is to generate realistic images.

So you can tackle, for example, missing data problems with this and this then leads into semi supervised learning where you can also use this, for example, for augmentation.

You can also use it for image to image translation, which is also a very cool application.

We will later see the so-called cycle gun where you can really do a domain translation and you can also simulate possible futures.

Let's say you use this in reinforcement learning.

So we would have all kinds of interesting domains where we could apply these unsupervised techniques as well.