Okay, so today, we do a little bit of a shifting direction, because so far we have been talking

about the MR Fourier problem all the way from the very beginnings to making it more and

more complex iterative methods going to nonlinear methods. And where I used to end this course

is with using machine learning in particular deep learning methods as an extension of compress

sensing for exactly this type of problem. So what has happened in the past is that at

least the students I taught this last year, they had never heard about machine learning

before. So they didn't get a lot out of the last lecture. So what I changed then was that

I added two lectures that give an introduction to machine learning methods. And then in the

very last one, we're going to tie these together with the image reconstruction problem that

we have been dealing with so far. So this is a heterogeneous audience, right? Because

we have the data science people, the AI people, the medical engineering people, computer science

people. And usually my expectation is that the first couple of lectures are boring for

the medical engineering people because they've heard this before. These particular ones are

now probably boring for the computer science data science AI people. But it is what it

is, right? If you have a heterogeneous audience then... So just a quick show of hands out

of the people who are still here, who is doing medical engineering? Okay. And who is doing

CS, data science, AI? Okay. So it's a bit of a hard... It's fine. But if this is very,

very basic, that means we probably can go away pretty quickly. We can do it more interactively

and go through the content, not just me talking, which is more the same. So what we're going

to do is a little bit of an overview of machine learning, deep learning and AI. And then I

gonna show you in probably at least the second half or the last two-thirds of the lecture,

two examples. One is for using a multi-layer perceptron, the other one convolutional neural

network. Two examples, and those are the two examples that you will work on in the exercises.

So here you get the background and already some insights into the results of the experiments.

And then on Thursday, you're gonna do it yourself and experiment with those. And the topics

that we are mostly gonna touch upon are things like model complexity, overfitting, underfitting,

and we're gonna talk a little bit about architectures, in particular, how convolutional and fully

connected networks relate to these topics. Everybody on board? Okay. So the first thing,

I just wanna get a definition out because this is just something that is mentioned a

lot in this context of artificial intelligence is what these particular terms mean. And in

general, artificial intelligence is a very broad term and covers a wide range of topics.

I'm just gonna give you an example on the next slide. Then machine learning is a subset

of artificial intelligence. And I would say it's the subset of artificial intelligence

that actually works, although there are some people who might disagree with me here. And

deep learning is then another subset of machine learning where we use a certain paradigm of

model training and combined with a certain type of architecture. We're gonna see that

in a couple of years. But let's spend a short amount of time and talk about the general

AI concept. And this is really something that is very, very broad and very, very wide. And

there's one definition of what an artificial intelligence is. And this is the Turing test

that was made famous by Alan Turing. Probably everybody knows him. Anybody? No, I'm sure

everybody, probably everybody, but at least some people here know the Turing test. Anybody

wants to explain it, how it works? When you determine if a system is a real, has real

AI capabilities. Yes. It's a test where in this case, the person can talk to two partners.

And he or she has to find out that one of the partners is an AI or not. And if it's

not possible to find out, then the AI is a general AI. That is correct. That is the Turing

test. And some people these days argue in the field that the Turing test is passed.

I mean, if you think about things like chat PT, anybody played around with this, by the

way? Yes. Yeah. It's a lot of fun, I have to say. I think I asked it who, when I was

at the conference, ISMREM conference last weekend, I asked it what the conclusion of

my talk should be, how MRI is going to be in 20 years. It came up with a couple of interesting