Hello and welcome to our class on active learning.

This is part of our seminar on advances in deep learning for time series.

Today's topic, as already mentioned, is on active learning.

So we are going beyond supervised learning.

The sources for this lecture are essentially Lillian Wang's blog posts.

She is now working at OpenAI and she has a really recommendable blog for different topics

among them also active learning and a publication from myself

Rasmus and Jan

Learning with

Limited Labels Data.

If you want to dive a little bit deeper

I recommend these three papers.

In this lecture

we are going to review Jordan Ash's paper and we are going to see Senna

I.

Zavarez's paper.

The last one, by Settles, we won't review because it's a literature review which is

very, very thorough.

But it is the standard literature on algorithms and the topic of active learning.

If you are interested

you can use this paper for a deep dive.

So what is active learning?

Our motivation is first

what are the downsides of supervised learning?

So supervised machine learning models, they require large amounts of labelled training

especially if you are working with deep learning models.

What's the problem with it, you might wonder?

They ask how many data to download.

Yes

but labelling

the annotating process of your large datasets is costly.

It not only takes time but also requires knowledge.

Often you have to involve domain experts.

Now domain experts

they are not like run of the mill experts

no they are your domain

experts like a medical doctor or an engineer that has specific knowledge on your time series.

So as you can imagine

these people

they have limited amount of time they can spend

in labelling data.

If you ever work with medical data and the University Hospital of Erlangen for example,

it's really hard to get experts to annotate data for you.

They have better things to do basically.

Doctors can heal a person or they can annotate your data

so what would they prefer to do?

So generally

there is a large amount of unlabeled data in the world that has not yet been annotated

by domain experts.

If you have ever worked in a real world project