Cutting Edge Computing

No AI Cannot Solve the World’s Biggest Problems February 23, 2016

‘AI can solve world’s biggest problems’ – Google Brain engineer.

The headline screamed for a response. The type of work being done with Google Brain is impressive.  It relies on neural networks.  Neural networks are cool things.  I was recently reminded by a new LinkedIn connection of the work and writing that I did on neural networks in the 1990s.

Neural networks are exceptionally good algorithms. They use multiple levels of nonlinear equations, running in parallel and serially, that can approximate a given response to a set of data inputs, then adjust the variables so that each successive pass can result in a slightly better approximation (or it may not; that’s just the way nonlinear algorithms work sometimes).

That may sound like just so much gobbledygook, but it’s not. In the 1990s, I used neural networks to design a set of algorithms to power an electronic wind sensor.  In my aborted doctoral research, I was using neural networks to help dynamically determine the most efficient routings through large scale networks (such as the nascent Internet at the time).

Let’s be clear. What I did had nothing to do with the world’s biggest problems.  The world’s biggest problems don’t already have mountains of data that point to the correct answer.  In fact, they rarely, if ever, have correct answers.  What they have is a combination of analysis, guesswork, and the many compromises made by dozens of competing interests.  And they will never result in even close to a best answer.  But sometimes they produce a workable one.

From my own experiences, I came to believe that having an instinctive feel for the nature and behavior of your data gave you a leg up in defining your neural network. And they tended to work best when you could carefully bound the problem domain.  That is hardly the stuff of something that can solve the world’s most difficult problems.

Neural networks can seemingly work magic, but only when you have plenty of data, and already know the right answers to different combinations of that data. And even though, you are unlikely to get the best possible solution.