Interacting intelligence

Recent work (1) by researchers at MIT and ABB show a more productive path for machine intelligence. Learning machines have been the holy grail of computing but applications have been trivial, thus far. Systems have been mostly limited to mechanical automatons fancifully called, robots. Machine learning has been a broad area – ranging from statistical models to artificial vision. But applications have been heavily dominated by rules – humans unable to shed the limitations of their own wiring. A more practical interaction between humans and machines in which machines are allowed to learn from the rather fickle humans may be better than prescriptive teaching by humans.

Prescriptive languages – that dominate today’s computing – are ill-equipped to implement learning robustly. Early attempts at descriptive languages such as PROLOG were promising but could not get adoption at scale. Every time something new appeared, programmers quickly went back to the environment that is rules based. One of the reasons for this is that engineering education is rules based and this is replicated up and down the whole computing stack.

To get out of this catch 22, two important things need to happen. First, a new generation of computing technologies are needed that fundamentally break away from prescriptive programming. Scientists and programmers have to get comfortable with the notion of a descriptive architecture that does not allow outcomes that are complete and deterministic. And, second, the education systems need to change – with computing making a break from engineering and moving toward the studies of more complex systems such as in Biology and Physics.

Without a major conceptual change in the computing architecture, machine learning will remain a hobby for the conceivable future.

Robotic assistants may adapt to humans in the factory. Published: Wednesday, June 13, 2012 - 11:35 in Mathematics & Economics Source: Massachusetts Institute of Technology