pdorg_site/talks/svm-lightning/svm_lightning.txt

Support Vector Machines

(lightning talk)

(LPW '07) (john melesky)

---

Presupposing:

<ul>
<li>You have a bunch of something.</li>
<li>You can transform relevant attributes of those things into numbers.</li>
<li>You can connect those numbers into vectors (think coordinates in an attribute space).</li>
<li>You want to categorise them base on those numbers.</li>
</ul>

---

The problem: find a line that separates these two categories of thing

<img style="float: right;" src="basedata.png" />

---

For humans, this is easy.

<img style="float: right;" src="cleansep.png" />

For mathematicians, it's actually not too hard.

---

For humans, this is easy.

<img style="float: right;" src="cleansep.png" />

For <del>mathematicians</del> computers, it's actually not too hard.

---

There are two problems, though.

---

Problem, the first:

<img style="float: right;" src="badline1.png" />

---

Problem, the first:

<img style="float: right;" src="badline2.png" />

---

Problem, the second:

<img style="float: right;" src="badset1.png" />

---

Problem, the second:

<img style="float: right;" src="badset2.png" />

---

Problem, the second:

<img style="float: right;" src="badset3.png" />

---

Conveniently, Support Vector Machines address both of the problems i've identified.


---

Solution, the first:


---

Solution, the first:

<img style="float: right;" src="bordervectors.png" />

<ul>
<li>Create "border" vectors, parallel to eachother, touching the outermost edge of each category dataset.</li>
</ul>

---

Solution, the first:

<img style="float: right;" src="bordervectors.png" />

<ul>
<li>Create "border" vectors, parallel to eachother, touching the outermost edge of each category dataset.</li>
<li>As you add new items, ensure these "borders" stay parallel.</li>
</ul>

---

Solution, the first:

<img style="float: right;" src="supportvectors.png" />

<ul>
<li>Create "border" vectors, parallel to eachother, touching the outermost edge of each category dataset.</li>
<li>As you add new items, ensure these "borders" stay parallel.</li>
<li>Create your categorizing vector equidistant from your two "borders".</li>
</ul>

---

Solution, the first:

<img style="float: right;" src="supportvectors.png" />

<ul>
<li>Create "border" vectors, parallel to eachother, touching the outermost edge of each category dataset.</li>
<li>As you add new items, ensure these "borders" stay parallel.</li>
<li>Create your categorizing vector equidistant from your two "borders".</li>
<li>These "borders" are called "support vectors".</li>
</ul>

---

A joke:

Q: How many mathematicians does it take to change a lightbulb?


---

A joke:

Q: How many mathematicians does it take to change a lightbulb?

A: One, who hands it to 127 Londoners, thus reducing it to an earlier joke.


---

A question:

Q: How do mathematicians categorize non-linearly-separable data?


---

A question:

Q: How do mathematicians categorize non-linearly-separable data?

A: Munge the data until it's linearly separable, thus reducing it to an earlier slide.


---

A question:

Q: How do mathematicians categorize non-linearly-separable data?

A: Munge the data until it's linearly separable, thus reducing it to an earlier slide.

Seriously. The munging is done using what are known as "kernel methods".

---

Kernel Methods

<ul>
<li>Functions that munge data</li>
<li>Very faintly magical (because i have no idea how they were derived)</li>
<li>Require some skill to choose the right one for the problem</li>
<ul>

---

Kernel Methods + Support Vectors = Support Vector Machines

---

In Perl:

Algorithm::SVM - bindings to libsvm

(Also wrapped by AI::Categorizer)