Okay, that gives me five minutes for inductive logic programming, which is essentially again

this idea of cheating at the same, at the right time.

Okay, so I would like to make an example here, is when you kind of learn new predicates or new relations from facts.

And to make this accessible, we're using family ties. All of you understand that, right?

And we have kind of this kind of relations, right? We have, we know the genders of the objects involved, right?

Male and female. We have the married to relation, possibly with a little bit of background.

And we have the parentship or a childship relations.

And this tree gives you a knowledge base of, I haven't counted, but I would say something like 50 facts or something like this.

And we have a couple of, we have facts of the form fatherhood, motherhood, married, and the genders.

And you can then go and give examples about other relations.

For instance, and that's I'm going to, I'm going to use the grandparent relation as an example here.

And then you can give examples like Queen Mum is the grandmother of Charles and, but not of Harry.

Right. So you can then say you have a couple of examples here, say something like 12 positives and 400 negatives.

Now you can find a set of sentences that give you a hypothesis so that actually the entailment constraint,

which I had on this slide, is actually satisfied. So find a hypothesis that solves that thing.

And then if you do that, you come up with a nice set of hypotheses.

And one of that is there is a Z such that some Z is married to X and also to Y.

And there's fatherships things.

And this will typically be a huge thing, which is, which doesn't generalize very well.

OK. This is an artifact of the fact that marriage is symmetric. Yes.

I believe that's the sense of motherhood.

Hmm?

I believe that it's the sense of power.

Oh, yes, yes, I'm sorry. Yes.

I got I had to abbreviate and now I got sidetracked. Yes.

Mother and father. Yes, yes, yes. I'm sorry.

So what can we do? We could try attribute based learning, which is something we've learned.

And that happens to not actually work very well for relations because we would have to do it on pairs.

So things just becomes very awkward because logic becomes very awkward.

You want to you want to have things like mom is actually the mother of who is the mother of Elizabeth.

Because the mother relation is actually an interesting thing here.

So that doesn't you can't really make it work.

OK. So the first thing I would like to do is to show you that.

Changing the syntax or changing the set of context or changing the background knowledge in this case,

having background knowledge about parents, having the parent concept actually helps quite a lot.

Instead of having to synthesize this monster here.

You only have to synthesize this little thing.

Right. X is a grandparent of Y. If there is a Z such that X is a parent of Z and Z is a parent of Y.

Only that our data actually had this kind of unnatural feeling of having mother and father where we're thinking about grandparent.

Makes this other thing quite so big just we can factor essentially via the parent predicate and that actually makes.

Makes things much easier.

So. Where does parent come from?

One way that's the way I showed you here is by the problem designer.

And we humans when we describe problems are typically quite good at having the necessary concepts around.

Why? Because many of us have either wondered about or explain to our children what these grandparents are actually.

So we had the problem and. Imagine your three year old asks and you say, well, it's either the father of the mother or the mother of the father or the father of the father and the mother of the mother and they'll say.

That's not what we want.

So we will naturally invent these.

These intermediate things that make us make it easier for us to explain and if we have it easier to explain them, things like explanation based learning works much better.

and so on.