1.2 Machine learning lecture 3  (Page 4/14)

Student: Your Y is exponentially [inaudible]?

Instructor (Andrew Ng) :Yeah. Let’s see. So it turns out there are many other weighting functions you can use. It turns out that there are definitely different communities of researchers that tend to choose different choices by default. There is somewhat of a literature on debating what point – exactly what function to use. This, sort of, exponential decay function is – this happens to be a reasonably common one that seems to be a more reasonable choice on many problems, but you can actually plug in other functions as well. Did I mention what [inaudible] is it at? For those of you that are familiar with the normal distribution, or the Gaussian distribution, say this – what this formula I’ve written out here, it cosmetically looks a bit like a Gaussian distribution. Okay? But this actually has absolutely nothing to do with Gaussian distribution. So this is not that a problem with XI is Gaussian or whatever. This is no such interpretation. This is just a convenient function that happens to be a bell-shaped function, but don’t endow this of any Gaussian semantics. Okay?

So, in fact – well, if you remember the familiar bell-shaped Gaussian, again, it’s just the ways of associating with these points is that if you imagine putting this on a bell-shaped bump, centered around the position of where you want to value your hypothesis H, then there’s a saying this point here I’ll give a weight that’s proportional to the height of the Gaussian – excuse me, to the height of the bell-shaped function evaluated at this point. And the way to get to this point will be, to this training example, will be proportionate to that height and so on. Okay? And so training examples that are really far away get a very small weight.

One last small generalization to this is that normally there’s one other parameter to this algorithm, which I’ll denote as tow. Again, this looks suspiciously like the variants of a Gaussian, but this is not a Gaussian. This is a convenient form or function. This parameter tow is called the bandwidth parameter and informally it controls how fast the weights fall of with distance. Okay? So just copy my diagram from the other side, I guess. So if tow is very small, if that’s a query X, then you end up choosing a fairly narrow Gaussian – excuse me, a fairly narrow bell shape, so that the weights of the points are far away fall off rapidly. Whereas if tow is large then you’d end up choosing a weighting function that falls of relatively slowly with distance from your query. Okay?

So I hope you can, therefore, see that if you apply locally weighted linear regression to a data set that looks like this, then to ask what your hypothesis output is at a point like this you end up having a straight line making that prediction. To ask what kind of class this [inaudible] at that value you put a straight line there and you predict that value. It turns out that every time you try to vary your hypothesis, every time you ask your learning algorithm to make a prediction for how much a new house costs or whatever, you need to run a new fitting procedure and then evaluate this line that you fit just at the position of the value of X. So the position of the query where you’re trying to make a prediction. Okay? But if you do this for every point along the X-axis then you find that locally weighted regression is able to trace on this, sort of, very non-linear curve for a data set like this. Okay?