SMALL NEU­RAL NET­WORKS, SMARTER PHONES

HWM (Malaysia) - - IMPACT - Ko­hWanzi

by As it turns out, it’s not just the hu­man brain that is worth mod­el­ing. Re­searchers at IBM have cre­ated a compact sys­tem mod­eled af­ter the 48 mil­lion nerve cells in the brain of a small-sized ro­dent. By mim­ick­ing the net­work of neu­rons and synapses in a liv­ing brain, com­puter sys­tems and al­go­rithms can be fed data and learn to per­form spe­cific tasks.

While deep learn­ing tech­nol­ogy is al­ready avail­able in the form of Face­book’s face recog­ni­tion fea­ture and Skype’s real-time trans­la­tion ser­vice, IBM’s chip could al­low smaller devices like smart­phones, hear­ing aids, and even watches to har­ness the tech­nol­ogy and re­lated al­go­rithms more ef­fec­tively. In the case of Google Now, most of the com­put­ing takes place on Google’s servers, which means they’re in­ac­ces­si­ble if you hap­pen to be work­ing off the grid.

How­ever, IBM chips, which have been dubbed TrueNorth, open up the pos­si­bil­ity of mov­ing at least some of the pro­cess­ing heft onto phones and other per­sonal devices. A sin­gle TrueNorth chip con­tains 5.4 bil­lion tran­sis­tors, but draws only 70 mil­li­watts of power.

So while a com­pany like Google could train neu­ral net­works in its own data cen­ters to rec­og­nize say, pic­tures of cats, the smaller-scale neu­ral net­work on your phone would build off this ca­pa­bil­ity and sim­ply help to ex­e­cute it.

Newspapers in English

Newspapers from Malaysia

© PressReader. All rights reserved.