1.8 Machine learning lecture 9

MachineLearning-Lecture09

Instructor (Andrew Ng): All right, good morning. Just one quick announcement, first things for all of your project proposals, I’ve read through all of them and they all look fine. There is one or two that I was trying to email back comments on that had slightly questionable aspects, but if you don’t hear by me from today you can safely assume that your project proposal is fine and you should just go ahead and start working on your proposals. You should just go ahead and start working on your project. Okay, there’s many exciting proposals sent in on Friday and so I think the proposal session at the end of the quarter will be an exciting event.

Okay. So welcome back. What I want to do today is start a new chapter in between now and then. In particular, I want to talk about learning theory. So in the previous, I guess eight lectures so far, you’ve learned about a lot of learning algorithms, and yes, you now I hope understand a little about some of the best and most powerful tools of machine learning in the [inaudible]. And all of you are now sort of well qualified to go into industry and though powerful learning algorithms apply, really the most powerful learning algorithms we know to all sorts of problems, and in fact, I hope you start to do that on your projects right away as well.

You might remember, I think it was in the very first lecture, that I made an analogy to if you’re trying to learn to be a carpenter, so if you imagine you’re going to carpentry school to learn to be a carpenter, then only a small part of what you need to do is to acquire a set of tools. If you learn to be a carpenter you don’t walk in and pick up a tool box and [inaudible], so when you need to cut a piece of wood do you use a rip saw, or a jig saw, or a keyhole saw whatever, is this really mastering the tools there’s also an essential part of becoming a good carpenter. And what I want to do in the next few lectures is actually give you a sense of the mastery of the machine learning tools all of you have. Okay?

And so in particular, in the next few lectures what I want to is to talk more deeply about the properties of different machine learning algorithms so that you can get a sense of when it’s most appropriate to use each one. And it turns out that one of the most common scenarios in machine learning is someday you’ll be doing research or [inaudible] a company. And you’ll apply one of the learning algorithms you learned about, you may apply logistic regression, or support vector machines, or Naïve Bayes or something, and for whatever bizarre reason, it won’t work as well as you were hoping, or it won’t quite do what you were hoping it to.

To me what really separates the people from – what really separates the people that really understand and really get machine learning, compared to people that maybe read the textbook and so they’ll work through the math, will be what you do next. Will be in your decisions of when you apply a support vector machine and it doesn’t quite do what you wanted, do you really understand enough about support vector machines to know what to do next and how to modify the algorithm? And to me that’s often what really separates the great people in machine learning versus the people that like read the text book and so they’ll [inaudible] the math, and so they’ll have just understood that. Okay? So what I want to do today – today’s lecture will mainly be on learning theory and we’ll start to talk about some of the theoretical results of machine learning. The next lecture, later this week, will be on algorithms for sort of [inaudible], or fixing some of the problems that the learning theory will point out to us and help us understand. And then two lectures from now, that lecture will be almost entirely focused on the practical advice for how to apply learning algorithms. Okay? So you have any questions about this before I start? Okay.