It's the paradox of human-computer interaction. Computers can process huge numbers quickly and without bias, but programming them to detect faces, trees and puppies is incredibly difficult. Determining beautiful, pristine or cute is impossible.
People, on the other hand, are adept at recognizing patterns. Even newborn humans show a tendency to prefer human faces, demonstrating that the pattern-recognition part of us is deep and innate.
Now researchers are readying a new suite of tools that marry those two complementary skills, using software to enhance and refine human intuition. "The computers do the computer stuff, and the humans do the human stuff," says Eric Bonabeau, founder of Icosystem.
Bonabeau, a former researcher at the Santa Fe Institute, calls his innovation "the hunch engine." Presented to a general audience for the first time at the O'Reilly Emerging Technology Conference here, the engine is a technological implementation of the "obscenity principle" -- a user of the hunch engine may not know what they are looking for, but they will "know it when they see it," the test Supreme Court Justice Potter Stewart famously offered as a metric to define obscenity.
When the user starts the hunch engine he is presented with a seed -- a starting point -- and a set of mutations. The user selects mutations that look promising in his eyes, and the application uses that selection to generate another set of mutations, continuing in that fashion until the user is satisfied with what he sees.
Call it guided natural selection, where the selector for fitness is what looks good to the human in front of the monitor.
Bonabeau wants to see the technology move into the consumer sphere, but acknowledges difficulties. Using any application built on the hunch engine requires some dedication of time and attention, to look at and select favorable versions of whatever the user is evaluating.
"People don't want to go through (even) five versions," Bonabeau says.
To that end, Bonabeau is trying to build applications compelling enough that people will take the time and get the results they want. One of his first applications, demonstrated at the O'Reilly conference, is a filter for images that allows a naive user to improve digital photos without understanding complex tools like Adobe Photoshop, by choosing from mutations of the picture to make it better. "My grandmother doesn't know anything about improving pictures," says Bonabeau, "but she knows which pictures of her grandchildren she likes."
Wired News tried out the photo selector. After loading the photograph you want to improve, the application shows you nine mutant versions. In the case of a dark photo, it's easy enough to pick a lighter version and move it to the seed area so that it becomes the foundation of the next crop of mutation pictures. You can keep selecting and mutating indefinitely. When you find the version you like, you save it. In a photo of a dark house and a moon, seven iterations were enough to lighten the photograph adequately.
Other planned offerings from Icosystem include an interior-decorating aid and name-selecting software that checks whois information as well as patent and trademark data to make sure company names are available as .coms and are not trademarked.
In France, the hunch engine has already been used by postal workers to build optimal carrier routes that fulfill management's requirements, while allowing workers to optimize for intangible benefits. "There could be a nice family or a cafe they liked," says Bonabeau.
None of these are the killer apps of genetic algorithms that Bonabeau hopes for -- but Bonabeau says they're a start.
Coalesix, an Icosystem spinoff company, is working with pharmaceutical industry partners to develop and roll out Mobius, a hunch-engine application for drug discovery. Using Mobius, an experienced pharmaceutical chemist searching for a new drug can examine 12 molecules, choose three that look promising and then watch them recombine and mutate into 12 more mutant molecules, one of which might be closer to a drug that can be developed and tested on people.
Jim Wikel, chief technology officer of Coalesix, says the feedback from the companies and chemists testing the software has been promising. "The ideas that our engines are generating are reasonable ... and ideas (chemists) wouldn't have had if left to their own devices," he says.
Dave Weinberger, author of Small Pieces Loosely Joined, says the hunch engine might be useful when people are aiming for a result that they cannot predict, but are able to recognize. "We already design cars and eyeglasses ... but our hunches tend to be repetitive and predictable," while mutation can take us in novel directions, he says.
"Computers only deal with explicit data, but so much of what we care about is implicit," says Weinberger. "It provides a way for a human to guide a computer through the inarticulate and implicit direction she wants to go."