1.11 Machine learning lecture 12

MachineLearning-Lecture12

Instructor (Andrew Ng) :Okay. Good morning. I just have one quick announcement of sorts. So many of you know that it was about two years ago that Stanford submitted an entry to the DARPA Grand Challenge which was the competition to build a car to drive itself across the desert. So some of you may know that this weekend will be the next DARPA Grand Challenge phase, and so Stanford – the team that – one of my colleagues Sebastian Thrun has a team down in OA now and so they’ll be racing another autonomous car.

So this is a car that incorporates many tools and AI machines and everything and so on, and it’ll try to drive itself in midst of traffic and avoid other cars and carry out the sort of mission. So if you’re free this weekend – if you’re free on Saturday, watch TV or search online for Urban Challenge, which is the name of the competition. It should be a fun thing to watch, and it’ll hopefully be a cool demo or instance of AI and machines in action.

Let’s see. My laptop died a few seconds before class started so let me see if I can get that going again. If not, I’ll show you the things I have on the blackboard instead. Okay. So good morning and welcome back. What I want to do today is actually begin a new chapter in 229 in which I’m gonna start to talk about [inaudible]. So [inaudible]today is I’m gonna just very briefly talk about clustering’s, [inaudible] algorithm. [Inaudible]and a special case of the EM, Expectation Maximization, algorithm with a mixture of [inaudible] model to describe something called Jensen and Equality and then we’ll use that derive a general form of something called the EM or the Expectation Maximization algorithm, which is a very useful algorithm. We sort of use it all over the place and different unsupervised machine or any application. So the cartoons that I used to draw for supervised learning was you’d be given the data set like this, right, and you’d use [inaudible]between the positive and negative crosses and we’d call it the supervised learning because you’re sort of told what the right cross label is for every training example, and that was the supervision. In unsupervised learning, we’ll study a different problem. You’re given a data set that maybe just comprises a set of points. You’re just given a data set with no labels and no indication of what the “right answers” or where the supervision is and it’s the job of the algorithm to discover structure in the data.

So in this lecture and the next couple of weeks we’ll talk about a variety of unsupervised learning algorithms that can look at data sets like these and discover there’s different types of structure in it. In this particular cartoon that I’ve drawn – one has the structure that you and I can probably see as is that this data lives in two different crosses and so the first unsupervised learning algorithm that I’m just gonna talk about will be a clustering algorithm. It’ll be an algorithm that looks for a data set like this and automatically breaks the data set into different smaller clusters. So let’s see. When my laptop comes back up, I’ll show you an example. So clustering algorithms like these have a variety of applications. Just to rattle off a few of the better-known ones I guess in biology application you often cross the different things here. You have [inaudible] genes and they cluster the different genes together in order to examine them and understand the biological function better. Another common application of clustering is market research. So imagine you have a customer database of how your different customers behave. It’s a very common practice to apply clustering algorithms to break your database of customers into different market segments so that you can target your products towards different market segments and target your sales pitches specifically to different market segments.