Machine learning lecture 1  (Page 13/17)

So, for example, it turns out you can prove surprisingly deep theorems on when you can guarantee that a learning algorithm will work, all right? So think about a learning algorithm for reading zip codes. When can you prove a theorem guaranteeing that a learning algorithm will be at least 99.9 percent accurate on reading zip codes? This is actually somewhat surprising. We actually prove theorems showing when you can expect that to hold.

We'll also sort of delve into learning theory to try to understand what algorithms can approximate different functions well and also try to understand things like how much training data do you need? So how many examples of houses do I need in order for your learning algorithm to recognize the pattern between the square footage of a house and its housing price? And this will help us answer questions like if you're trying to design a learning algorithm, should you be spending more time collecting more data or is it a case that you already have enough data; it would be a waste of time to try to collect more. Okay?

So I think learning algorithms are a very powerful tool that as I walk around sort of industry in Silicon Valley or as I work with various businesses in CS and outside CS, I find that there's often a huge difference between how well someone who really understands this stuff can apply a learning algorithm versus someone who sort of gets it but sort of doesn't.

The analogy I like to think of is imagine you were going to a carpentry school instead of a machine learning class, right? If you go to a carpentry school, they can give you the tools of carpentry. They'll give you a hammer, a bunch of nails, a screwdriver or whatever. But a master carpenter will be able to use those tools far better than most of us in this room. I know a carpenter can do things with a hammer and nail that I couldn't possibly. And it's actually a little bit like that in machine learning, too. One thing that's sadly not taught in many courses on machine learning is how to take the tools of machine learning and really, really apply them well.

So in the same way, so the tools of machine learning are I wanna say quite a bit more advanced than the tools of carpentry. Maybe a carpenter will disagree. But a large part of this class will be just giving you the raw tools of machine learning, just the algorithms and so on. But what I plan to do throughout this entire quarter, not just in the segment of learning theory, but actually as a theme running through everything I do this quarter, will be to try to convey to you the skills to really take the learning algorithm ideas and really to get them to work on a problem.

It's sort of hard for me to stand here and say how big a deal that is, but when I walk around companies in Silicon Valley, it's completely not uncommon for me to see someone using some machine learning algorithm and then explain to me what they've been doing for the last six months, and I go, oh, gee, it should have been obvious from the start that the last six months, you've been wasting your time, right?