A less-technical description of Waleed Kadous' PhD

What is your thesis really about?

A fair enough question -- I mean what does "Extending Classification models to Temporal Domains" actually mean?

One very interesting area of research right now is machine learning. This area is really abou trying to make machines learn from their experience -- if a computer observes something new, it tries to take the new observation into account. It can also be viewed as trying to make computers that are more like people: the more experience we have of something, the better we get at it. Computers aren't like that right now.

One interesting part of machine learning is concept learning, sometimes known as classification. Concept learning works something like this: you are given examples (called instances) of different types of things (called classes), and you have to come up with a way to tell them apart (this is called a classifier). Sometimes, you are given hints as well about how to tell them apart (this is called background knowledge).

Sounds complicated? Well, people do it every day. For example, most people learn to tell fruit apart by seeing lots of different examples of fruit; not from a dictionary.

I might give you lots of different fruit, and tell you what kind of fruit it is. I might give you a spherical, orange, pitted object and tell you it's an orange; a smooth-surfaced, spherical red object about hand-sized and tell you it's an apple; a small, green, smooth-skinned spherical object, which is an apple; and a long yellow, smooth-skinned object called a banana.

After seeing a few more examples of fruit, you'll be able to guess the kind of fruit without me telling you what kind of fruit it is. Nobody quite understands how we do this - there's a lot of speculation about it. Some people say that we simply remember every single fruit we've seen and find what looks the most similar; some people think we try to make rules (like: if it's long, yellow and bent then it's probably a banana).

Researchers have figured out ways to do this kind of learning - telling different kinds of flowers apart, deciding whether you should get a loan or not, all that sort of stuff.

Telling apart oranges and apples is one thing, but the world's a little more complex than that. People seem to have the ability to recognise patterns that occur over time. Imagine now that I'm not asking you to classify oranges and apples, but trying to get you to recognise something more complex - say, something that varies over time - for example, different melodies. Even if a melody is played on a different instrument, slower or faster, or in a different style, we can tell that it might be the same or different from another tune.

This is what I am interested in: How do we learn to classify things that vary over time? And how can we use existing techniques to solve these kind of learning problems? In particular, is it possible to find a way to make a computer recognise temporal variations the same way we do in a general manner that people can?

Note that I said in a general manner. It might be easy to solve one particular learning problem, but that really does not tell us how to solve another one.

To test out my theories, I'm looking at three "testbed" applications, three different learning problems:

Auslan sign recognition

Auslan is the language of the Australian Deaf community. Like most sign languages, it involves movements of the hands, as well as facial expressions. By capturing information using a pair of instrumented gloves, we can try to learn a small subset of Auslan signs

Robots

A lot of what is used to control robots today uses a very simple analysis of sensors. Perhaps by adding the ability to recognise complex patterns, more interesting behaviours could be developed.

ECG Analysis

Electrocardiographs (ECG's) are used in diagnosing heart problems. Doctors already have rules about what particular patterns in ECG's mean; what would a computer make of the data? Would it come up with the same rules as doctors use?