We looked at neural networks.

That's about as far as we got.

I would like to give you

a very brief introduction to

a machine learning technique

called support vector machines,

which is essentially linear classification on steroids.

It's one of the most popular machine learning techniques.

So if you have a problem that you want to

attack with machine learning and aren't sure how.

It's not restaurant data,

so you probably don't want to do decision trees.

You have no clue what, then you typically

turn to support vector machines as a first guess.

You pull some support vector machine implementation,

and there are many. There's a couple in MetLab,

and there's a couple in or based on NumPy.

So you would just pull that down from the network,

and just run it and see whether it does something sensible.

So therefore, I would like to just give you

a brief idea of the ideas,

and then I would like to close up

where this machine learning from examples,

from observations chapter and go onto something else.

So the idea with support vector machines is that,

if you look at linear regression,

then you can do better in two aspects.

Say you have a separable dataset,

and then you have the choice, by the way,

actually between multiple separators here.

I have three here.

They're all separators.

So they all solve the problem.

You can ask yourselves, which one is the best?

Is this one good?

The answer is probably, no,

not really because it almost touches these.

So there's a very small margin of error.

The other is good.

Well, this one almost touches here and almost touches there.

So it's a natural idea to say,

well, we're not interested in just any linear separator.

We are in a way interested in the best separator.

One that keeps most distance to all the critical values,

the ones, these, in this case,

three here that come closest to the gap we want to describe.

So this is a good one because it has

an Epsilon environment that gives a stable error,

a region of error.

Just visually, you would think that this here generalizes better