The Far Limits of Artificial IntelligenceBy
The year 2017 was hailed by the World Economic Forum as the Year of AI (artificial intelligence), and early signs verified the claim with the debut of a flurry of new products running AI and with established products being retooled to add AI.
Along with the marketplace, the dam was breached on worldwide investment. BI Intelligence waded in to measure the depth of the deluge, plotting the numbers on global quarterly artificial intelligence spending: First quarter 2012 was $79 million; first quarter 2017, a remarkable $1.73 billion. The leading investor, for now, is the United States, with 66% of the total, but China is second and expanding. And the BI group’s analyses foresee AI technologies contributing up to $16 trillion to the global economy by 2030.
Over the course of the year, the already loud discussion drifted to focus on an older AI question. Could a machine that has sufficient memory to manage Big Data, along with the ability to learn on its own within a neural network, also be creative?
Humans are proud of our ability to produce works of the imagination. We might have conceded math processing, complicated prediction chores, film animation, and even, to an extent, medical diagnoses and some surgeries to computers and robotics, but we draw the line, with a smirk, when asked whether we would like to read a story or listen to a musical composition created by a computer.
Time magazine’s Lev Grossman got to the heart of the problem with this explanation: “Creating a work of art is one of those activities we reserve for humans and humans only. It’s an act of self-expression; you’re not supposed to do that if you don’t have a self.” We have ceded much to our expanding population of “thinking machines,” but few are ready to accept androids or networks that are self-aware.
For any discussion of creative computers, you have to begin with a working definition of creative. What does a computer, or human, need to create works of art or inventions, and how would we evaluate the results? In Luke Dormehl’s Thinking Machines: The Quest for Artificial Intelligence—And Where It’s Taking Us Next, the author suggests beginning with this definition: “Creativity is the act of creating something that is new to society as a whole.”
Some computers have been taught the structural rules for sonnets and string quartets, and they can reassemble works in the proper formats. And some programs are writing stories covering sporting events for newspapers and financial reports for companies. In these cases, the programs organize factual content that isn’t “new to society as a whole.”
But then there’s the Beatles program created by Lior Shamir, a computer science professor at Lawrence Technological University. Shamir wanted to create an algorithm that would chart the evolution of the band’s sound as it changed over the years. His algorithm analyzed the 13 Beatles albums and found 2,883 “unique numerical content descriptors.” Asked to sort the albums in chronological order, it did it without an error. To be sure, Shamir then tested his algorithm on ABBA, U2, and Queen, and the computer once again sorted the groups’ albums correctly in order, despite having no information other than the music itself.
Shamir now believes it’s possible to create new songs that could have been on past specific albums. “The computer will be able to compose songs on the heuristics that you would find on that album. That may mean using the same scales, time signatures, musical instruments, or whatever.” Whether Samir’s computer, with its 2,883 descriptors, will be able to compose new Beatles songs from specific periods is still an open question awaiting future improvements to his algorithms.
Dormehl describes an antenna that was invented by AI for a NASA Explorer rocket. Electrical engineer Jason Lohn was asked to create a small but high-bandwidth antenna for a lunar space mission. He explains, “I wanted to use AI to improve the actual hardware that was being used in space missions.” After being provided all the constraints, an evolutionary algorithm optimized a solution. It took several hundred generations, and the resulting antenna design looked like a mistake. In fact, it resembled a “bent paper clip.”
Lohn made the prototype, and when he tested it, it worked extremely well even though he couldn’t explain why. On September 6, 2013, NASA’s Lunar Atmosphere and Dust Environment Explorer was launched with the only antennae being three of Lohn’s computer-invented rigs aboard. The mission was a success.
Next month in “The Far Limits of AI, Part 2,” we will look at a way of testing computer imagination and IBM’s Chef Watson.