Small Numbers, Huge Changes

In a recent inter­view, Sun­dar Pichai of Google dis­cuss­es improve­ments in the accu­ra­cy of their voice recog­ni­tion:

Just in the last three years, we have tak­en things like error in word recog­ni­tion down from about 23 per­cent to 8 per­cent.

That’s the dif­fer­ence between mis­un­der­stand­ing one word in four, to one word in twelve; the dif­fer­ence between com­plete­ly unus­able, and annoy­ing.

Andew Ng, for­mer­ly of Google and now of Baidu, expands on this:

Most peo­ple don’t under­stand the dif­fer­ence between 95 and 99 per­cent accu­rate. Nine­ty-five per­cent means you get one-in-20 words wrong. That’s just annoy­ing, it’s painful to go back and cor­rect it on your cell phone.

Nine­ty-nine per­cent is game chang­ing. If there’s 99 per­cent, it becomes reli­able. It just works and you use it all the time. So this is not just a four per­cent incre­men­tal improve­ment, this is the dif­fer­ence between peo­ple rarely using it and peo­ple using it all the time.

It’s fas­ci­nat­ing to see how these small num­bers make a huge dif­fer­ence; you might think Google’s 92% accu­rate is only a lit­tle less than Baidu’s 95% accu­rate, but in prac­ti­cal terms there’s a big gulf. And it gives me pause to think about the mon­ey, human resource and com­put­ing pow­er spent on try­ing to make those small huge increas­es in accu­ra­cy.

An interview with Andrew Ng

The Huff­in­g­ton Post series, Sophia, brings ‘life lessons from fas­ci­nat­ing peo­ple’. Their lat­est inter­view is with Andrew Ng, a Stan­ford Uni­ver­si­ty pro­fes­sor, co-founder of Cours­era, and key mem­ber of the deep learn­ing teams at first Google and now Baidu. I real­ly like some of the insights in his inter­view, the prac­ti­cal­i­ty with which he treats inno­va­tion and the easy way he explains hard con­cepts.

For exam­ple, here he talks about career advice:

I think that “fol­low your pas­sion” is not good career advice. It’s actu­al­ly one of the most ter­ri­ble pieces of career advice we give peo­ple. Often, you first become good at some­thing, and then you become pas­sion­ate about it. And I think most peo­ple can become good at almost any­thing.

On retrain­ing the work­force for a more heav­i­ly auto­mat­ed future:

The chal­lenge that faces us is to find a way to scal­ably teach peo­ple to do non-rou­tine non-repet­i­tive work. Our edu­ca­tion sys­tem, his­tor­i­cal­ly, has not been good at doing that at scale. The top uni­ver­si­ties are good at doing that for a rel­a­tive­ly mod­est frac­tion of the pop­u­la­tion. But a lot of our pop­u­la­tion ends up doing work that is impor­tant but also rou­tine and repet­i­tive. That’s a chal­lenge that faces our edu­ca­tion­al sys­tem.

On why machine learn­ing is sud­den­ly more pop­u­lar, despite the tech­nol­o­gy being around for decades in some cas­es:

I often make an anal­o­gy to build­ing a rock­et ship. A rock­et ship is a giant engine togeth­er with a ton of fuel. Both need to be real­ly big. If you have a lot of fuel and a tiny engine, you won’t get off the ground. If you have a huge engine and a tiny amount of fuel, you can lift up, but you prob­a­bly won’t make it to orbit. So you need a big engine and a lot of fuel. We final­ly have the tools to build the big rock­et engine – that is giant com­put­ers, that’s our rock­et engine. And the fuel is the data. We final­ly are get­ting the data that we need.

And the chal­lenges of talk­ing to com­put­ers:

Most peo­ple don’t under­stand the dif­fer­ence between 95 and 99 per­cent accu­rate. Nine­ty-five per­cent means you get one-in-20 words wrong. That’s just annoy­ing, it’s painful to go back and cor­rect it on your cell phone. Nine­ty-nine per­cent is game chang­ing. If there’s 99 per­cent, it becomes reli­able. It just works and you use it all the time. So this is not just a four per­cent incre­men­tal improve­ment, this is the dif­fer­ence between peo­ple rarely using it and peo­ple using it all the time.

It’s a real­ly inter­est­ing inter­view, I encour­age you to read it in full.

The Future Is Coming Faster Than We Think

Today I read a fas­ci­nat­ing arti­cle in the Lon­don Review of Books. The Robots Are Com­ing, by John Lan­ches­ter, is about the rise of cheap automa­tion and the effect it’s going to have on the work­force and soci­ety at large. In his intro­duc­tion he talks about the Accel­er­at­ed Strate­gic Com­put­ing Ini­tia­tive’s com­put­er, Red, launched in 1996 and even­tu­al­ly capa­ble of pro­cess­ing 1.8 ter­aflops — that is, 1.8 tril­lion cal­cu­la­tions per sec­ond. It was the most pow­er­ful com­put­er in the world until about 2000. Six years lat­er, the PS3 launched, also capa­ble of pro­cess­ing 1.8 ter­aflops.

Red was only a lit­tle small­er than a ten­nis court, used as much elec­tric­i­ty as eight hun­dred hous­es, and cost $55 mil­lion. The PS3 fits under­neath a tele­vi­sion, runs off a nor­mal pow­er sock­et, and you can buy one for under two hun­dred quid. With­in a decade, a com­put­er able to process 1.8 ter­aflops went from being some­thing that could only be made by the world’s rich­est gov­ern­ment for pur­pos­es at the fur­thest reach­es of com­pu­ta­tion­al pos­si­bil­i­ty, to some­thing a teenag­er could rea­son­ably expect to find under the Christ­mas tree.

This makes me think of IBM’s Wat­son, a deep learn­ing sys­tem, ten years in the mak­ing at a cost in excess of $1 bil­lion, with hard­ware esti­mat­ed at $3 mil­lion pow­er­ing it, and com­ing soon to children’s toys.