Obiter dicta by Professor Gavin Brown AO

Ocular revolution

13 September 2002

There is a chapter in an autobiographical work by a famous Australian scientist with a heading like "Positions which I have been offered". It is devoted to jobs for which the hero was head-hunted but which he declined. Feeling that the act is in doubtful taste and being rather short on examples, I have never felt tempted to kiss and tell in this way. There are, however, exceptions to every rule.

The distinguished Hellenist, Sir Kenneth Dover, tried very hard to persuade me to become his PhD student in the early '60s. This was an unusual compliment as I had no Greek. It is true that I was a fair Latinist but my real allure was as a precocious mathematical statistician. Dover, who soon returned from my native heath to his beloved Oxford, saw me as sculptor's clay, a renaissance post-adolescent, and, given the times, I speculated fleetingly over a putative link to substance experimentation.

In fact, he was fascinated by the work of an amateur Scottish clergyman who had painstakingly applied the simple device of frequency of word count in an attempt to identify authorship of biblical passages. Sir Kenneth had the notion of developing more sophisticated statistical techniques and applying those to ancient classical texts.

All of this brings me to Dasher. I am indebted to The Economist for bringing to my attention some work by two Cambridge scientists, David MacKay and David Ward, which has recently been reported in Nature. Dasher is a text-input program which allows eye-typing. By the last descriptor I mean quite literally typing by using one's eyes rather than one's fingers.

Motivated by the needs of the severely disabled, the new technology measures light reflected from the eyes to determine which letter they have chosen on a computer screen. The beauty and effectiveness of the program is that different letters are displayed with varying prominence depending on a simple piece of learning software.

The underlying mathematical model was frequency distributions to predict from each string of five characters what, in order of likelihood should be the next. It is a learning algorithm because the frequencies are constantly and automatically updated according to the user's style and vocabulary.

This is a nice example of good science because it is so primitive. By isolating one component of a profoundly complex process it exploits to the full the capacity of a dumb machine which needs to know nothing of what is really going on.

There is obvious potential beyond servicing the needs of quadriplegic communicators and further speculation is irresistible. Might one have the beginnings of a dramatically simple scanning test for plagiarism? The underlying mathematical program could easily be adapted to flag a passage where the frequencies were being re-adjusted beyond an individual's established norm. How much of a fingerprint does a user leave? I do not purport to know the range of existing work in that field. Apparently Dr MacKay calibrates the program by using Jane Austen's Emma as a trial text. Does a machine trained on Billy Connolly show a discernibly different profile? Can one extend to words – even phrases – with benefit? Will my next column be written by a machine?