Neural nets make Chicago blues see red

Rooting out rogue cops from amongst Chicago's finest is the job of an artificial intelligence program that mimics the neural networks of the brain. The city's police department is the first to use BrainMaker Professional, a program developed by California Scientific Software, to anticipate misconduct by law-enforcement officers.

The Mayor of Chicago, Richard Daley, directed the police to employ the software program when the Democrat Party announced it would hold its convention in the city this August for the first time since what a presidential commission called a "police riot" occurred at the 1968 convention.

The software simulates a grid of interconnected processors that correspond roughly to neurons and synapses in the brain. The ability of neural network software to predict an outcome from the input of disparate data has resulted in a range of uses from predicting recidivism by criminals on probation to recognising mosquitoes from the sound of their wings.

Using the performance records of nearly 200 police officers who had been dismissed or resigned under investigation as a control, the program is able to match behavioural characteristics against 12,500 officers on the force and forecast which are "at risk".

BrainMaker was "trained" using information such as age, race, sex, number of traffic accidents, reports of lost badges or weapons, frequency of sick leave, insubordination or criminal misconduct. The program has turned up 91 officers, of whom half were already enrolled in counselling programs.

"We're very pleased with the outcome," says Raymond Risley, the assistant deputy superintendent in charge of the Internal Affairs division. "We consider it efficient at identifying at-risk personnel sooner than command officers might be able to.The old method just can't compete."

The use of BrainMaker is strongly opposed by Chicago's Fraternal Order of Police, who view the system as Orwellian and unethical. "It's ludicrous," says Bill Nolan. "I told them if this thing is so good, we should give it to all the detectives so they can solve all the murders and robberies."

For statisticians and computer experts, the main objection to neural networks is that they do not show how they arrive at a conclusion. Unlike other kinds of artificial-intelligence techniques that produce results according to a set of rules, BrainMaker operates by complex non-linear processes. In theory, it should constantly correct itself by weighing the records of dismissed officers against those with a clean record. If it has guessed incorrectly - as it often does initially - a mathematical formula makes a correction.

"Voodoo," says Zenon Pylyshyn, a professor of cognitive science at Rutgers University. "People are fascinated by the prospect of getting intelligence by mysterious Frankenstein-like means - by voodoo! And few attempts to do this have been as successful as neural nets."

Following police union protests, Internal Affairs announced the network would only be used to complement division supervisors and stressed that an officer who came up on BrainMaker would not be subject to disciplinary action. It would, the department said, merely provide an opportunity for an officer moving in the wrong direction to rehabilitate him or herself.