We were talking about knowledge and learning and after having completed regression-based

learning, essential neural nets and linear regression, I tried to convince you that

knowledge actually might be good in learning because it feels counterintuitive that whenever

you learn you start from zero, from a blank slate. And since we don't really know how to

localize learning or knowledge in say weights in a neural network, a likely way is to change the

learning paradigm to something that can actually look and represent knowledge and that would be

logic. And we've talked a bit about inductive learning and how that actually works with logic.

We can formulate, it turns out, the learning problem easily in logic. The question of course

is can we learn efficiently? How does that work? And the upshot of the whole thing is essentially

that if we can express the stuff we want to learn in terms of logic, which is a big assumption

for motor skills and so on, think about RoboCup or something like that, robot soccer. So most of

the tasks there are motor skills, sensor skills and those kind of things where the big problems

are actually kicking the ball so that it actually goes towards the goal and if you're a humanoid

robot not falling down at the same time. So those are motor skills and it would be very

difficult to express those in logic. So this whole idea of knowledge and learning is really

quite restricted to things that we can express well in logic. But if you can, if you have the

right predicates and so on, then it's very simple. You just basically have to solve an

entailment problem. Essentially if you think of this as an equation, kind of an in equation,

it's directed, but you have a set of examples which we think of as data, we have a set of

say, classifications, which is the stuff we want to learn and then we have hypotheses

and the data plus the hypothesis should entail the classifications. So that's really what

we want. But of course there's a lot that can go wrong here. So that's the main problem

here. So if we go to the restaurant example again, that's easy to express in logic. Indeed

these attributes are very simply logic things as well. And the idea is that with that, with

this framework we can actually do cumulative learning. And we discussed a couple of examples.

One is that when you're in this caveman example, that you actually, when you're trying, when

you're learning, when you're learning that advanced techniques, you're actually building

on the fact that you've already learned that say cooked food is good. The other example

we're going to look at is the kind of Brazilian example where you have, where you meet somebody

at the airport. Fernando, you know that Fernando is Brazilian and he speaks Portuguese. Now

what can we learn from this? Well, there's two obvious things to learn. One is that everybody

in Brazil speaks Portuguese. It's kind of reasonable. And the other thing you could

learn is that everybody's called Fernando. Not so reasonable. So that's a problem for

logic based techniques. And we'll use this example to see how we can get around this

kind of an perceived asymmetry here. And the third example is things where you have a doctor,

she has an intern, and then the intern kind of observes the doctor. We have, we see that

the doctor has a patient, the patient describes the disease, the doctor says, ask a couple

of questions and then says, why don't you take this and that antibiotic? And the intern

is supposed to learn something from this. And probably what the intern learns is that

this particular antibiotic is good in that particular disease. And if the intern sticks

around long enough, she might be able to generalize saying, oh, whenever there's a cough involved,

take this or something like this. And the theme that we're going to use is that we're

going to use background knowledge in this kind of an iterative way, building up a body

of knowledge that actually makes subsequent learning easier. And probably that rings a

bell. Probably that's something you feel familiar with. Usually you take AI1 before AI2 because

the knowledge builds up, hopefully.