If you put together the information from the last days, you could say, first of all, we

had this more philosophical insight in artificial intelligence.

Then we had a lot of technical building blocks to get some insight about the behavior of

neural networks.

And yesterday, we have developed a whole puzzle game to put these pieces together to come

to all the features that we need to realize projects with neural networks.

We have done this for regression.

We have done this for classification.

We said we have explained that you can do, let's say, a standard level with three layer

ensemble.

And then you have good arguments that this is giving you a lot of insights into the real

world, for example, discussion with monotonicity and so on.

And then we said, let's assume it's all fine, but it's not good enough.

So the next step then was to say, let's work on this deep, feel-forward neural networks.

And then you have a new chance to finally solve your problem.

Let's assume that even this is not good enough.

Do you have to call it a failure?

Do we have to stop then?

The answer is no.

But the possibilities that you have then afterwards are a little bit more dependent on your application,

not such a general scheme in relationship to what we have seen before.

Now let's think about ideas which showed up mainly because some things were not working

before.

But they are not completely dependent on applications.

So I show you only things which have the possibility as a general tool, as a general insight.

And the first one of these things here is that I want to speak about a situation which

you all observe every day.

In the moment, I will speak about only here about the right side of the picture here.

See in your normal day life, you also have the situation that you have observations.

This might be information from radio, TV, internet, discussion with other people here.

You have a lot of information about this.

And the model which you build up based on this, you could call your world model.

That's the combined model of all the information that you have there.

But hopefully, this world model that you have there has also let us say an internal safety

algorithm so that you not accept everything.

You have your world model and then you go back to the observations, to the data.

And then you have a chance to clean the data so that you do not have to believe on all

data points which you are confronted with.

If you would believe in all the observations that you get over the day, then you also would

have to believe in what you read in newspapers like Bill Cider.

And then I would say forget your world model, your picture of the world.

So the point is you have a bidirectional process in living there.

You have observations.

You have your picture of the world.

And this picture of the world again should allow you to evaluate new incoming data.

So we believe that we are good in doing this two-directional thing here.

But why are we good?

This is because of the normal life.

You do not think about it.

But if you would think about it, then you would see that it's a very difficult question