AI allows machines to car­ry out all sorts of tasks that used to be the domain of humans alone. Need to run qual­i­ty con­trol on a fac­to­ry pro­duc­tion line? Set up an AI-pow­ered cam­era to spot defects. How about inter­pret­ing med­ical data? Machine learn­ing can iden­ti­fy poten­tial tumors from scans and flag them to a doc­tor.

But appli­ca­tions like this are use­ful only so long as they’re fast and secure. An AI cam­era that takes min­utes to process images isn’t much use in a fac­to­ry, and no patient wants to risk the expo­sure of their med­ical data if it’s sent to the cloud for analy­sis.

These are the sorts of prob­lems Google is try­ing to solve through a lit­tle-known ini­tia­tive called Coral.

“Tra­di­tion­al­ly, data from [AI] devices was sent to large com­pute instances, housed in cen­tral­ized data cen­ters where machine learn­ing mod­els could oper­ate at speed,” Vikram Tank, prod­uct man­ag­er at Coral, explained to The Verge over email. “Coral is a plat­form of hard­ware and soft­ware com­po­nents from Google that help you build devices with local AI — pro­vid­ing hard­ware accel­er­a­tion for neur­al net­works … right on the edge device.”

Coral’s prod­ucts, like the dev board (above), can be used to pro­to­type new AI devices. Image: Google

You might not have heard of Coral before (it only “grad­u­at­ed” out of beta last Octo­ber), but it’s part of a fast-grow­ing AI sec­tor. Mar­ket ana­lysts pre­dict that more than 750 mil­lion edge AI chips and com­put­ers will be sold in 2020, ris­ing to 1.5 bil­lion by 2024. And while most of these will be installed in con­sumer devices like phones, a great deal are des­tined for enter­prise cus­tomers in indus­tries like auto­mo­tive and health care.

To meet cus­tomers’ needs Coral offers two main types of prod­ucts: accel­er­a­tors and dev boards meant for pro­to­typ­ing new ideas, and mod­ules that are des­tined to pow­er the AI brains of pro­duc­tion devices like smart cam­eras and sen­sors. In both cas­es, the heart of the hard­ware is Google’s Edge TPU, an ASIC chip opti­mized to run light­weight machine learn­ing algo­rithms — a (very) lit­tle broth­er to the water-cooled TPU used in Google’s cloud servers.

Build an AI marsh­mal­low-sort­ing machine or a smart bird feed­er

While its hard­ware can be used by lone engi­neers to cre­ate fun projects (Coral offers guides on how to build an AI marsh­mal­low-sort­ing machine and smart bird feed­er, for exam­ple), the long-term focus, says Tank, is on enter­prise cus­tomers in indus­tries like the auto­mo­tive world and health care.

As an exam­ple of the type of prob­lem Coral is tar­get­ing, Tank gives the sce­nario of a self-dri­ving car that’s using machine vision to iden­ti­fy objects on the street.

“A car mov­ing at 65 mph would tra­verse almost 10 feet in 100 mil­lisec­onds,” he says, so any “delays in pro­cess­ing” — caused by a slow mobile con­nec­tion, for exam­ple — “add risk to crit­i­cal use cas­es.” It’s much safer to do that analy­sis on-device rather than wait­ing on a slow con­nec­tion to find out whether that’s a stop sign or a street light up ahead.

Tank says sim­i­lar ben­e­fits exist with regard to improved pri­va­cy. “Con­sid­er a med­ical device man­u­fac­tur­er that wants to do real time analy­sis of ultra­sound images using image recog­ni­tion,” he says. Send­ing those images to the cloud cre­ates a poten­tial weak link for hack­ers to tar­get, but ana­lyz­ing images on-device allows patients and doc­tors to “have con­fi­dence that data processed on the device doesn’t go out of their con­trol.”

Google’s Edge TPU, a tiny pro­cess­ing chip opti­mized for AI that sits at the heart of most Coral prod­ucts. Image: Google

Although Coral is tar­get­ing the world of enter­prise, the project actu­al­ly has its roots in Google’s “AIY” range of do-it-your­self machine learn­ing kits, says Tank. Launched in 2017 and pow­ered by Rasp­ber­ry Pi com­put­ers, AIY kits let any­one build their own smart speak­ers and smart cam­eras, and they were a big suc­cess in the STEM toys and mak­er mar­kets.

Tank says the AIY team quick­ly noticed that while some cus­tomers just want­ed to fol­low the instruc­tions and build the toys, oth­ers want­ed to can­ni­bal­ize the hard­ware to pro­to­type their own devices. Coral was cre­at­ed to cater to these cus­tomers.

The prob­lem for Google is that there are dozens of com­pa­nies with sim­i­lar pitch­es to Coral. These run the gamut from star­tups like Seat­tle-based Xnor, which makes AI cam­eras effi­cient enough to run on solar pow­er, to pow­er­ful incum­bents like Intel, which unveiled one of the first USB accel­er­a­tors for enter­prise in 2017 and paid $2 bil­lion last Decem­ber for the chip­mak­er Habana Labs to improve its edge AI offer­ings (among oth­er things).

Giv­en the large num­ber of com­peti­tors out there, the Coral team says it dif­fer­en­ti­ates itself by tight­ly inte­grat­ing its hard­ware with Google’s ecosys­tem of AI ser­vices.

This stack of prod­ucts — which cov­ers chips, cloud train­ing, dev tools, and more — has long been a key strength of Google’s AI work. In Coral’s case, there’s a library of AI mod­els specif­i­cal­ly com­piled for its hard­ware, as well as AI ser­vices on Google Cloud that inte­grate direct­ly with indi­vid­ual Coral mod­ules like its envi­ron­ment sen­sors.

In fact, Coral is so tight­ly inte­grat­ed with Google’s AI ecosys­tem that its Edge TPU-pow­ered hard­ware only works with Google’s machine learn­ing frame­work, Ten­sor­Flow, a fact that rivals in the AI edge mar­ket The Verge spoke to said was poten­tial­ly a lim­it­ing fac­tor.

Coral’s com­mit­ment to Google prod­ucts might put off some cus­tomers

“Coral prod­ucts process specif­i­cal­ly for their plat­form [while] our prod­ucts sup­port all the major AI frame­works and mod­els in the mar­ket,” a spokesper­son for AI edge firm Kneron told The Verge. (Kneron said there was “no neg­a­tiv­i­ty” in its assess­ment and that Google’s entry into the mar­ket was wel­come as it “val­i­dates and dri­ves inno­va­tion in the space.”)

But exact­ly how much busi­ness Coral is doing right now is impos­si­ble to say. Google is cer­tain­ly not push­ing Coral with any­where near as much inten­si­ty as its cloud AI ser­vices, and the com­pa­ny wouldn’t share any sales fig­ures or tar­gets for the group. A source famil­iar with the mat­ter, though, did tell The Verge that the major­i­ty of Coral’s orders are for sin­gle units (e.g. AI accel­er­a­tors and dev boards), while only a few cus­tomers are mak­ing enter­prise pur­chas­es on the order of 10,000 units.

For Google, the attrac­tion of Coral may not nec­es­sar­i­ly be rev­enue, but sim­ply learn­ing more about how its AI is being applied in the places that mat­ter. In the world of prac­ti­cal machine learn­ing right now, all roads lead, inex­orably, to the edge.

Some of Coral’s hard­ware prod­ucts, includ­ing an AI accel­er­a­tor (far right) and dev board (cen­ter). Image: Coral

Source link