KEY POINTS:
One of the most popular storylines in science fiction is the one about the computer that eventually outsmarts its creator. The machine's maker either comes to a sticky end at the hands of his invention or puts it to work in a plot to dominate the world.
Professor Nik Kasabov, head of the Knowledge Engineering Discovery Research Institute (KEDRI) at AUT University, has no fears of falling victim to a silicon-based killing machine - not for another 40 or 50 years, at any rate.
And rather than plotting world domination with a super-intelligent but inanimate ally, Kasabov wants to help solve some of medicine's biggest challenges using computers that learn.
Kasabov, president-elect of the International Neural Network Society, is a world leader in developing artificial neural networks - computer systems that consist of small processing units that are connected like neurons.
Like the brain, Kasabov's processing units adapt on the basis of the inputted data. But he doesn't pretend they can match human intelligence yet.
"The brain is so complex, it cannot be simply emulated fully, but we do use principles of brain information processing to create new information technologies."
Kasabov uses the term "connectionist system" to describe the brain structure and the design of the systems being built at KEDRI.
As the brain processes information, it learns by giving different weights to the synapses, or connections, between neurons. Synaptic weights are determined by the brain's ability to generalise, to deal with abstraction and to recognise clusters of information.
Kasabov's processing units use mathematics to mimic those abilities.
"The system creates its own structure, just like the brain. We don't tell the brain which neuron to use for recognising smell and recognising images, and how to connect them. The brain does it in a self-organised way."
If that sounds well on the way to artificial intelligence, Kasabov thinks not. What his systems actually display is computational intelligence, which they do by storing large amounts of data from which generalisations can be made.
They can make associations between features of an object and the class of object to which it belongs.
After exposure to lots of information about chairs, one of his systems would be able to tell you that a four-legged object on which someone sits is called a chair.
Where neural systems have an advantage is in sheer processing power. So, for example, KEDRI is collaborating with Dunedin company Pacific Edge Biotechnology on tools for predicting cancer patients' chances of survival based on genetic data.
Inevitably, some patients live and some die, and when that data is entered into the system, it adjusts accordingly.
"If we present a new person's data, the system will give their probability of surviving based on the previous data."
But Kasabov says that kind of system is old hat and a long way from a computer that is able to think for itself. KEDRI is edging towards such capabilities. "We have created so-called evolving connectionist systems. They are still neural networks but they are more intelligently self-organising in evolving their structure.
"You can add new data, new variables to the system and it will adapt incrementally on-the-fly. It improves over time."
But this is "just the tip of the iceberg" - even more exciting is the development of computational models that burrow still deeper into the brain.
With processing units on that scale, the vast number of connections where information can be stored creates a correspondingly powerful system.
It's at the atomic, or quantum, level where Kasabov thinks computers might be clever enough to pass the Turing test.
The test was proposed in 1950 by mathematician Alan Turing, whereby a human judge has a typewritten conversation with another human and a machine and tries to tell them apart.
It remains the benchmark for determining whether a machine can think. It's not under immediate threat, Kasabov believes, but will eventually be passed.
The subtlety of information quantum computing might produce could take predictions of disease outcomes to another level. Kasabov has applied for government funding to take KEDRI's quantum computing work further.
His guess is that the first pure quantum computer is 25 years away and, in 50 years, computer systems will be dramatically different from those we know now.
"Computers can do quite a lot of intelligent information processing. But in terms of human-like thinking, that is still a long way to go." For now, then, there is nothing to be afraid of.
- NZ HERALD
