Good…

Okay so welcome back everybody

Today we have the grand finale of our lecture

and we have two very interesting topics today

is weekly supervised learning.

That's the first set of slides,

and then we will talk about integrating known operators

into networks and how we can apply deep learning

onto various signal processing tasks.

So weekly supervised learning addresses the problem

that you may not have a lot of training data available.

And there's several approaches that you can do.

So for example, if you have 3D data,

you may want to use 2D annotations

and still be able to learn.

And there's also methods how to move from labels

really to localization and we will look at those.

So I think these are interesting concepts

and they can reduce the amount of effort

used for labeling tremendously.

So let's look at how we can get 3D annotations

from 2D annotation.

And the idea here is that you have a 3D volume,

but you only annotate a couple of 2D slices

and then you train and apply a 3D unit

and it will still be able then to predict

a 3D generation or a 3D label.

So the idea is you only supply a couple of slices here

and then you want to be able to run it completely in 3D.

So this is the main idea

and one way where you can do that is for example

combine it with interactive segmentation.

So you start with segmenting a couple of slices,

then you train and then you run your net

and then you can correct for them.

So that would be interactive.

So the idea here is that you then train with sparse labels

and you could consider your labels being one-hot encoded.

So you already know this loss function here.

We have your negative logarithm

over the normalized exponential function here.

And now we add one label K

where we simply don't know the reference.

So if we didn't have the observation,

then we just add a weight

and this weight will simply cancel out all the elements

where we don't have an observation.

So essentially this will then cause

in your forward, during the training in the forward pass,

what is generated at this position will be canceled out