So, welcome everybody to

our deep learning lecture and today,

we want to talk a little about

unsupervised deep-learning methods and related methods.

So, why are we talking about unsupervised methods?

So far, we had

data sets and for classification,

we had really huge datasets.

You remember we are training up to millions of parameters in those deep networks,

and this also means that we have a lot of unknowns,

and we have to estimate a lot of parameters.

So because of these huge datasets,

we need many objects,

and typically we have few modalities.

If you think about specific tasks,

so for computer vision tasks we have the internet and ImageNet variable.

Oh!

What's happening?

And I need to start over.

Ah, there we go. Excellent.

So, welcome back to deep learning,

and today we want to talk about unsupervised methods.

You could argue that listening to a talk without slides is also kind of unsupervised.

No.

Unsupervised methods are very useful,

and the great thing about unsupervised methods is that we don't need any labels,

so we can apply those methods without the need of any specific training data,

and we also put in a related method,

the generative adversarial networks that are able to generate data,

and they are not absolutely unsupervised,

but they are to some degree related,

because you can use a different training label in order to generate data.

So these are also quite popular right now.

So the problem that we are facing so far,

we have those millions of parameters,

and we need a lot of labels,

so we need a lot of training data.

And if you consider general computer vision tasks,

they have databases like ImageNet and so on,

then you are able to use data from the internet,

and maybe you can also use the crowd to annotate your data,

so there is really ways of generating this,

and typically we have many objects and not so many modalities.

If you think of other fields like medical images,

then you run into a problem,

because a medical data is typically very, very scarce,

and if you have 30 to 100 patients,

then you are already super happy,

because you have obtained such a large data set.

Typically, also medical images are very focused,