Reinventing Discovery - Michael Nielsen [58]
Of course, a human chauvinist might object to my use of the term “intelligence” in “data-driven intelligence,” arguing that there’s nothing very intelligent about a computer searching ten million scientific papers, or searching the SDSS for dwarf galaxies. It’s just routine, mechanical work, albeit on a scale far beyond human ability. But here’s the point: these are problems we humans can’t solve at all. When it comes to making meaning from terabytes or petabytes (thousands of terabytes) of data, we’re not much better than any other animal. We have, at best, a few very specialized abilities in this domain, such as the ability to process visual images, and virtually no general-purpose large-scale data-processing ability. So who are we to judge computers in this domain? An unaided human’s ability to process large data sets is comparable to a dog’s ability to do arithmetic, and not much more valuable. So while these problems perhaps don’t require computers to be very smart, in this problem domain they are a lot smarter than humans. This point of view is captured in the diagram shown on this page.
It’s human nature to focus on the problems on the right of the diagram, the problems where human skill and ingenuity are most valuable. And it’s normal human prejudice to undervalue the problems on the left, the domain where data-driven intelligence really shines. But we’ll put aside this prejudice, and think about the problems on the left. What problems can computers solve that we can’t? And how, when we put that ability together with human intelligence, can we combine the two to do more than either is capable of alone?
As an example of the latter, in 2005 the chess website Playchess.com ran what they called a freestyle chess tournament, meaning a tournament where humans and computers could enter together as hybrid teams. To put it another way, the tournament allowed human intelligence to team up with data-driven intelligence, in the form of chess-playing computers, which rely on enormous opening and endgame databases, and which analyze myriad possible combinations of moves in the midgame. One of the entrants in the tournament was the team behind the Hydra series of chess computers. Hydra, at the time the world’s strongest chess computer, had never lost a game in regular play to any human chess player, and on several occasions had easily defeated top grandmasters. The Hydra team entered two of their computers, one playing entirely on its own, and the other playing with some human assistance. Also entered in the tournament were several teams pairing strong grandmasters with strong chess computers. On their own, neither the grandmasters nor their computers could match the Hydras. But the joint human-computer teams trounced the Hydras. Not only did neither Hydra win the tournament, but in fact neither even made it to the quarterfinals. The grandmasters could beat the Hydras because they knew when to rely on their computers, and when to rely on their own judgment. Even more interesting, the winner of the tournament was a team called ZackS that consisted of two low-ranked amateur players using three off-the-shelf computers, and standard chess-playing software. Not only did they outclass the Hydras, they also outclassed the grandmasters with their strong chess-playing computers. The human operators of ZackS demonstrated exquisite skill in using the data-driven intelligence of their computer algorithms to amplify their chess-playing ability. As one of the observers of the tournament, Garry Kasparov, later remarked, “Weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process.”
Data-driven intelligence has broader goals than artificial intelligence. For the most part, artificial intelligence takes tasks that human beings are good at and aims to mimic or better human performance. Think about computer programs to play human