One person who shows up at the last minute for some reason.

I'm wondering whether that's like an artifact of the system or whether that's an actual

person.

But always at like 24 it goes up by one.

Anyway, right before we start one reminder again that there is going to be this teaching

assessment thingy the day after tomorrow.

I put all the details on Stutt on so you can read up on them there.

Someone else suggested that since that basically removes half of the lecture on Thursday anyway

we might use the remainder of the session for something like a Q&A.

I would be happy to do that.

I will prepare slides anyway because I'm not sure whether we're going to manage filling

an entire remaining half of the lecture with questions alone.

I'd be delighted to be proven wrong.

But just in case if you have any questions if there's anything you would want me to

repeat with respect to the subject matter of the semester.

Prepare questions think about something or whatever we can gladly use the rest of the

session to like try to sort those things out before we continue with new stuff.

So last time we started with talking about machine learning in particular supervised

learning.

The general set up in the case of supervised learning is that we assume we have some kind

of function f from some domain A to some codomain B which we don't know.

And we have a bunch of hypotheses all of which are basically functions from A to B and we

have a set of examples which you can think of as a function on a subset of A to B or

IE a function on the subset of the domain.

We call those the examples or the training set and the goal is to derive some hypothesis

in H that is ideally consistent with all of the examples that we have with the goal of

that approximating the true function f which we don't actually know as well as by any means

possible.

In practice it's not always the case that our training set actually does represent a

function there is usually noise in there, there might be conflicting classification

these kinds of things.

But in order to like derive all of the like general techniques for machine learning we

basically just assume it is because that makes the problem easier and then some training

methods are more robust against the case where there are conflicting labels than others and

all these kinds of things.

If our codomain is just some finite set we call it a classification problem, if the codomain

is the reals we call it a regression problem.

Obviously we should to some extent try to make sure that the functions that we're trying

to learn or our hypothesis base is as simple as possible and of course as complicated as

necessary.

If we don't we get these kinds of artifacts which we're going to call over fitting later

where it turns out that we find some hypothesis that actually perfectly matches all of the

data that we have all of our training examples but it's just overly complicated with the

result that it's probably not going to generalize as well as simpler hypotheses tend to do.

So in generally keep the hypothesis base as small as you can without losing realizability

i.e. the possibility that our hypothesis base contains the function that we're trying to

learn in the first place.

In general we assume that all of the data that we have is IID independent and identically

distributed.

That is mostly an assumption that we use in order to derive all of these machine learning