BIOTECH

THE THINK­ING RO­BOT: YOUR NEW BEST FRIEND?

Guru Magazine - - IN THE NEWS - Si­mon Makin

Sci­en­tists at the Univer­sity of Water­loo, Canada, led by Pro­fes­sor Chris Elia­smith, have built the most so­phis­ti­cated sim­ula

tion of a work­ing brain ever con­structed. Al­though much smaller than the hu­man brain it­self, con­sist­ing of only 2.5 mil­lion brain cells (com­pared to 100 bil­lion) and many fewer than some pre­vi­ous sim­ula

tions – it dis­plays an im­pres­sive range of dif­fer­ent be­hav­iours. The ar­ti­fi­cial brain can recog­nise im­ages, re­mem­ber se­quences, and even com­plete the kind of com­plex task you might find in an IQ test. The su­per­com­puter pro­gram, called SPAUN (Se­man­tic Pointer Ar­chi­tec­ture Uni­fied Net­work), uses a 28 x 28 pixel dig­i­tal cam­era ‘eye’ to gather in­put from its sur­round­ings and then gives its re­sponses with a ro­bot arm. For in­stance, when shown the se­quence 1 2 3 - 5 6 7 - 3 4 ?, to­gether with an in­struc­tion, SPAUN will scrawl the digit ‘5’ on a piece of pa­per. Un­like IBM’s Wat­son su­per­com­puter, which was built in 2011 to do one thing (play Jeop­ardy!) and do it well, but made

no at­tempt to copy how the brain works, SPAUN repli­cates ac­tual brain cell ac­tiv­ity and wiring. More im­por­tantly, though, it turns this ac­tiv­ity into be­hav­iour. This is in con­trast to larger, more de­tailed brain mod­els, such as the Blue Brain Pro­ject, which pro­duce de­tailed sim­u­la­tions of neu­ral ac­tiv­ity, but don’t nec­es­sar­ily do any­thing. SPAUN has two soft­ware sys­tems that

work in har­mony: a ‘work­ing mem­ory’ sys­tem that is mod­elled on the ‘higher’ think­ing part of the brain (called the pre­frontal cor­tex – where we make our de­ci­sions), and an ‘ac­tion se­lec­tion’ sys­tem, which is based on other parts of the brain called the basal gan­glia and thal­a­mus (more prim­i­tive, in­stinc­tual re­gions). The ‘ac­tion

se­lec­tion’ sys­tem routes data to the part of the ‘work­ing mem­ory’ sys­tem ap­pro­pri­ate for the task, which then stores that data, and per­forms the nec­es­sary ‘think­ing’ func­tions. All this is ac­com­plished by soft­ware run­ning on a su­per­com­puter, but, as it stands, is still very limited com­pared to a real brain: SPAUN can only tackle eight pre­de­fined tasks, and is far slower than the real thing, tak­ing around two hours of com­put­ing time to sim­u­late one sec­ond of brain ac­tiv­ity. It does how­ever demon­strate a range of cog­ni­tive skills and even makes some of the same mis­takes we do, such as re­mem­ber­ing the first and last items in a list bet­ter than the oth­ers (known to psy­chol­o­gists as pri­macy and re­cency). The team also looked at the ef­fects of ‘killing off’ neu­rons to sim­u­late age­ing, and have seen pat­terns of de­cline sim­i­lar to what hap­pens in old age. Cru­cially though, SPAUN lacks adaptivity – the abil­ity to learn new tasks. This is a short­com­ing the team hopes to tackle in

fu­ture. Even so, there’s no rea­son to pre­sume that build­ing on this sim­u­la­tion will at some point pro­duce the more elu­sive qual­i­ties of liv­ing brains, such as aware­ness or in­ten­tions: SPAUN doesn’t do any of its im­pres­sive tricks be­cause it wants to – it is ex­plic­itly pro­grammed and fed in­struc­tions, like any other com­puter sys­tem. So with­out free will, there’s prob­a­bly not much need to worry about a fu­ture ver­sion tak­ing con­trol of our mis­sile de­fence sys­tems just yet…

Newspapers in English

Newspapers from International

© PressReader. All rights reserved.