Hello and welcome to the unit on semi-supervised learning that is part of the seminar advances

in deep learning for time series.

What is this about?

Well

as you know

supervised learning models

they require really large amounts of label

training data to learn properly.

But labeling itself is costly.

It not just takes time, but it also requires knowledge that's sometimes rare and you need

to hire domain experts like medical doctors or lawyers or people that know the topic and

can label the data correctly.

Sometimes our knowledge is just not enough.

Well

generally

there's a large amount of unlabeled data or data that has not yet been

labeled, then labeled data.

So you often as a practitioner in a real world project

you're often faced with this context

with limited labeled data scenarios.

So that means you have just the tiny amount of labeled data

maybe some samples and tons

and tons of unlabeled data.

Now the question is

how can we best leverage this information that is already given but

unlabeled if we already have a little bit of labeled samples?

We are in a scenario such as this.

We have a lot of data which is unlabeled and here shown as these gray dots and we have

some samples that are annotated as a subset.

Here we have three annotated green class triangles and four annotated orange class squares.

If we train a model only on these seven data points

we will receive this decision boundary

has a small margin around it and some of the points are actually on the decision boundary.

Now last week we have used active learning to specifically annotate such samples for

which the model is the least certain or the least confident

for example.

This week we're going to look at what we could do up here.

For example

what we could do using label propagation.

Semisupervised learning assumes that these nearby samples here have the same class as

the labeled samples.

So if you have unlabeled samples

nearby labeled samples

it's likely they have a similar or

the same class.

This is an assumption, of course.

Label propagation assumes this.

Just to recall and to contrast

active learning instead uses the most informative samples

to annotate.