De­moc­ra­tiz­ing Deep Learn­ing

HWM (Malaysia) - - THINK -

And then there are star­tups like Ner­vana, which of­fers deep learn­ing on its cus­tom hard­ware as a cloud ser­vice. By mak­ing deep learn­ing ca­pa­bil­i­ties more ac­ces­si­ble, it hopes to spur the de­moc­ra­ti­za­tion of deep learn­ing for an even wider va­ri­ety of ap­pli­ca­tions. It is not alone in this goal, and out­fits like Er­satz Labs, Me­taMind, and Sky­mind all want to en­able a broader adop­tion of the tech­nol­ogy.

Ner­vana CEO and co-founder Naveen Rao be­lieves that medicine could be one of the fields that will see the most ben­e­fit from the im­proved abil­ity of com­put­ers to rec­og­nize pat­terns and im­ages. For all the ad­vances in med­i­cal science, the in­ter­pre­ta­tion of scans like X-Rays and MRIs are still left up to doc­tors them­selves, which means that hu­man er­ror could very well find its way into cru­cial di­ag­noses. Rao thinks that deep learn­ing could make it pos­si­ble to ap­ply decades of ex­per­tise to a com­puter, and help less ex­pe­ri­enced doc­tors min­i­mize their own er­rors.

A more spe­cific ap­pli­ca­tion would be in the case of di­a­betic retinopa­thy (DR), an eye dis­ease that is a lead­ing cause of vi­sion im­pair­ment in di­a­bet­ics. De­tect­ing DR is a te­dious and time­con­sum­ing process that re­quires doc­tors to closely ex­am­ine dig­i­tal pho­to­graphs of pa­tients’ reti­nas. Fur­ther­more, the re­sources and ex­per­tise re­quired for ac­cu­rate di­ag­noses are of­ten lack­ing in ar­eas where di­a­betes is preva­lent, which means a re­li­able method to au­to­mate the process is sorely needed.

Kag­gle, a plat­form for data pre­dic­tion com­pe­ti­tions, ended up host­ing a con­test for quicker and more ac­cu­rate de­tec­tion. The re­sults were promis­ing, as all top five en­tries turned out re­sults that were more ac­cu­rate than hu­man doc­tors, the lat­ter’s ac­cu­racy be­ing de­ter­mined as the rate of agree­ment among three clin­i­cians.

To sum things up, be­fore ar­ti­fi­cial in­tel­li­gence can even sur­pass ours, it must first take a leaf from our book (or rather, our cra­ni­ums). HAL 9000 may be a long way off, but frag­ments of its pre­cur­sors are al­ready driv­ing ad­vances in ar­eas as var­ied as con­sumer tech­nol­ogy and medicine.

Neu­ral path­ways of a macaque brain as sim­u­lated by TrueNorth.

Newspapers in English

Newspapers from Malaysia

© PressReader. All rights reserved.