The Power of Networks

March 2017

I got it wrong. In my 1984 PhD dissertation I imagined a future in which we'd have a handheld device that would have encyclopedic information. I imagined we'd be able to interact with it just by talking. But I also imagined that the information would be stored on the device.

That was wrong. Today we have the handheld devices that we can talk to and that can deliver up just about anything we want to know. But that information is stored on computers all around the world. The internet is simply a network of millions of computers.

In the wonderful new book Homo Deus: A Brief History of Tomorrow, Yuval Harari explains that we can thank networks for the ascent of humanity. We had smarts, and we had language. But what was really necessary for progress was to connect people into networks. Writing and money did that. It facilitated trade networks, and began the march toward globalization, for better or worse.

At first the network was slow — about as fast as a horse could run. But the telegraph, railroads, telephone, satellite communications, and computer networks changed that. Today the network is nearly as fast as the speed of light, as lasers pulse through the fiber-optic cables that circle the planet, delivering gigabytes of knowledge, information, and data for just about everything.

Networks bring power. And computer scientists have realized since the 1940s that the brain is essentially a network of 100 billion neurons. Those neurons communicate with each other, retaining memories, responding to stimuli, generating ideas, fomenting desires.

Computer scientists in the 1950s began to imagine that they might be able to create computer programs that would work like brains: they would learn in the same way connected neurons learn. And they would create a sort of artificial intelligence.

It didn't work. And a different approach to artificial intelligence became predominant: one in which you programmed all the rules, and the computer then operated on the basis of those rules.

That worked. Computers learned chess and a computer eventually beat the world's reigning chess champion. But it was limited. They realized that there was much a two-year-old child could do, such as recognize a kitty, that a computer couldn't. They were stumped.

How could a computer be programmed to recognize "catness" in a photograph? Imagine all the different sizes and shapes and positions a cat could have in a photo. And how could a computer distinguish a cat from a squirrel?

Fortunately, despite the fact that the neural network approach to artificial intelligence was initially deprecated, a few diehards kept working in this area. And in the last five years, they've been successful, thanks in part to beefier computers and a welter of data. Google has now invested heavily in this area, hiring some of the world's leading experts (paying them starting salaries at over $1 million).

The results have been extraordinary. An infant's neurons learn to recognize a kitty based on patterns. Computers are now being programmed with neural network techniques, with nodes in the software programs simulating neurons in a brain. Show a neural net computer 10 million images of cats from YouTube videos (as Google did), and the software learns – by trial and error – to recognize catness. No need to tell it anything about eyes or tails or size. It learns the patterns on its own.

Google's artificial intelligence project, called Google Brain, began five years ago, and it's gone far beyond the ability to recognize cats in videos. It's now the intelligence behind Android's speech recognition on smartphones, which works well. Google uses it in their photo search technology and when giving viewing recommendations in YouTube.

But its biggest breakthrough has been with Google Translate. Until Google Brain, the technology for Google Translate was rules-based: dictionaries, grammars, etc. Once Google Brain was working well, they decided to try it for translation. Instead of dictionaries and grammars, they used neural net learning to learn language patterns. And last November, without telling anyone, they began using Google Brain for all the translations. People around the world immediately noticed the huge improvement in quality.

Four days later Google announced their new Translate service, and in a meeting with the press used this example from Borges: ""Uno no es lo que es por lo que escribe, sino por lo que ha leído."

Translation the old way: "One is not what is for what he writes, but for what he has read." With Google Brain: "You are not what you write, but what you have read."

Artificial intelligence has finally arrived, and now every company is scrambling to do what Google has done. And we are the beneficiaries, as our computers and smartphones can now be intelligent in ways not possible before.

© 2017 by Jim Karpen, Ph.D.

E-mail Jim Karpen