Market segmentation.

3D Structure Segmentation

Let's build a probabilistic word model of segmentation.

What one really wants is fully automatic semantic segmentation.

By definition, the best segmentation, which we'll call S , is equal to the one which maximizes the joint probability of the segmentation.

Sometimes people call this segmentation. Sometimes people call it grouping.

And we saw special effects like segmentation and changing background.

That is the segmentation that comes out now is the time.

But what we really need is one for the interactive segmentation approach.

So reorganization means segmentation. So bottom up. Assemblies of super pixels, regions.

And, from bottom up segmentation we can come up with candidate objects.

Because, examples of objects against the natural background really challenge segmentation systems.

Both in the medical example as well as in the image segmentation example.

So let's look at each of those in turn. First, the segmentation model.

So here are some examples, segmentations from the work of segmentation data set.

First, a segmentation model of how do we break up the German into phrases?

And for these more bottom up segmentation approaches, we need them to include semantics.

Another process is the one of reorganization or sometimes its called grouping and segmentation.

A very different application where PGMs have also been used is that of image segmentation.

The best segmentation of the rest of the words doesn't depend on the first word.

Now I want to give you an idea of how well the segmentation program performs.

And here are some results of the semantic segmentation task on the Pascal visual object challenge.

A semantic segmentation really means division of the image into regions corresponding to different objects, fully automatic.

We'll have a lot of segmentation problems for example if we see that the string y o r k dash any w as part as New York New Haven, how do we know, the correct segmentation is New York?

We'd like some way of searching through them efficiently without having to consider the probability of every possible segmentation.

This word here is not a word like in dash law. We have to solve the segmentation problem correctly.

And in the context of the image segmentation, it's the labels for different pixels or these larger regions called superpixels.

We would also consider the probability that they're considered a phrase together, and come up with a high probability segmentation.

One is market segmentation, where you may have a database of customers and want to group them into different market segments.

There's a certain state of the art in reorganization which is captured, for example, by research on the monthly segmentation data set.

We can say that the best segmentation is equal to the argmax over all possible segmentations of the string into a first word and the rest of the words of the probability of that first word times the probability of the best segmentation of the rest of the words.

So this is a very simple model. It takes only into account segmentation, translation between phrases, and just the simplest model of distortion.

And when you do a bottom up segmentation you are not making use of drop down knowledge such as that or object categories.

And that will enable us to do semantic segmentation. We should be able to mark all the pixels that correspond to a cat.

Now most media companies television, radio, publishing, games, you name it they use very rigid segmentation methods in order to understand their audiences.

And this is in fact kind of an example of market segmentation where you're using K Means to separate your market into 3 different segments.

Here it seems like it really has erred that small and insignificant seems like a much better segmentation than the one my program came up with.

So I think it is perfectly appropriate to call this process of segmentation and grouping a reorganization of the pixels into entities which are more semantically meaningful.

So we're going to segment the text into a sequence of words word 1 through word n and find that segmentation into words that maximize the joint probability.

Number two thing that Howard did is he made us realize it's another very critical point he made us realize the importance of what he likes to call horizontal segmentation.

So if we had a Markov model where the probability of insignificant depended on the probability of small, then we could catch that, and we could get this segmentation correct.

Here I've trained it on a corpus of 4 billion words not just the Shakespeare corpus but a larger corpus, and then I give it some test cases to try to find the best segmentation.

The market segmentation example I talked a little bit earlier, you can do that as an Unsupervised Learning problem because I am just gonna get my algorithm data and ask it to discover market segments automatically.

Then if we take the rest and multiply out this whole term, the best segmentation of the rest times the probability of the first that comes from this column, then that column will give us about 10 to 19 for the segmentation that starts with n, 10 to 13 for the one that starts with no, 10 t the 10 for the one that starts with now, and 10 to 18 for the one starts with nowi.

We can say that the best segmentation is approximately equal to the one that maximizes, and what we could do here is we could make the Markov assumption and say we're only going to be considering the few previous words.