The per­ils of treat­ing ba­bies like ro­bots

If we could un­der­stand how the in­fant mind de­vel­ops, all chil­dren might reach their full po­ten­tial. But see­ing them as learn­ing ma­chines is not the an­swer, says Alex Beard

The Guardian Weekly - - Weekly Review - Nat­u­ral Born Learn­ers by Alex Beard is pub­lished by Wei­den­feld & Ni­col­son

Deb Roy and Ru­pal Pa­tel pulled into their driveway on a fine July day in 2005 with the beam­ing smiles and sleep-de­prived glow com­mon to all first-time par­ents. Paus­ing in the hall­way of their Bos­ton home for Grandpa to snap a photo, they chat­tered hap­pily over the pre­cious new­born son swad­dled be­tween them. This nor­mal-look­ing sub­ur­ban cou­ple weren’t ex­actly like other par­ents. Roy was an AI and robotics ex­pert at MIT, Pa­tel an em­i­nent speech and lan­guage spe­cial­ist at nearby North­east­ern Univer­sity. For years, they had been plan­ning to amass the most ex­ten­sive home-video col­lec­tion ever.

From the ceil­ing in the hall­way blinked two dis­creet black dots, each the size of a coin. Fur­ther dots were lo­cated over the open-plan liv­ing area and the din­ing room. There were 25 in to­tal through­out the house – 14 mi­cro­phones and 11 fish-eye cam­eras, part of a sys­tem primed to launch on their re­turn from hos­pi­tal, in­tended to record the new­born’s ev­ery move.

It had be­gun a decade ear­lier in Canada – but in fact Roy had built his first ro­bots when he was just six years old, back in Win­nipeg in the 1970s, and he’d never re­ally stopped. As his in­ter­est turned into a ca­reer, he won­dered about an­droid brains. What would it take for the ma­chines he made to think and talk? “I thought I could just read the lit­er­a­ture on how kids do it, and that would give me a blueprint for build­ing my lan­guage and learn­ing ro­bots,” Roy told me.

Over din­ner one night, he boasted to Pa­tel, who was then com­plet­ing her PhD in hu­man speech pathol­ogy, that he had al­ready cre­ated a ro­bot that was learn­ing the same way kids learn. He was con­vinced that if it got the sort of in­put chil­dren get, the ro­bot could learn from it.

Toco was lit­tle more than a cam­era and mi­cro­phone mounted on a Mec­cano frame, given char­ac­ter with ping- pong- ball eyes, a red feather quiff and crooked yel­low bill. But it was smart. Us­ing voice recog­ni­tion and pat­tern-analysing al­go­rithms, Roy had painstak­ingly taught Toco to dis­tin­guish words and con­cepts within the mael­strom of ev­ery­day speech. Where pre­vi­ously com­put­ers learned lan­guage dig­i­tally, un­der­stand­ing words in re­la­tion to other words, Roy’s break­through was to cre­ate a ma­chine that un­der­stood their re­la­tion­ship to ob­jects. Asked to pick out the red ball among a range of phys­i­cal items, Toco could do it.

Pa­tel ran an in­fant lab in Toronto and Roy flew up there to see what he could learn. Ob­serv­ing the moth­ers and ba­bies at play, he re­alised he’d been teach­ing Toco badly. “I hadn’t struc­tured my learn­ing al­go­rithm cor­rectly,” he ex­plained to Wired mag­a­zine in 2007. “Ev­ery par­ent knows that when you’re talk­ing to an 11-month-old, you stay on a very tight sub­ject. If you’re talk­ing about a cup, you stick to a cup and you in­ter­act with the cup un­til the baby gets bored and then the cup goes away.”

His ro­bot had been search­ing through ev­ery pho­neme it had ever heard when it was learn­ing a new ob­ject, but Roy tweaked its al­go­rithm to give ex­tra weight to its most re­cent ex­pe­ri­ences, and be­gan to feed it au­dio from Pa­tel’s baby lab record­ings. Sud­denly Toco be­gan to build a ba­sic vo­cab­u­lary at a rate never seen be­fore in AI re­search. His dream of “a ro­bot that can learn by lis­ten­ing and see­ing ob­jects” felt closer than ever. But it needed to feed on record­ings, and these were hard to find.

No one had ever truly stud­ied “in the wild” what hap­pens to a child in those first cru­cial years. The norm for re­searchers were weekly hour-long ob­ser­va­tion ses­sions – that was how Pa­tel stud­ied moth­ers and in­fants in her lab. If you were go­ing to study the way a baby learned to talk, you’d need some­one ec­cen­tric enough to rig up a house with hid­den record­ing de­vices.

I first heard about Pa­tel and Roy’s ex­per­i­ment while work­ing as a teacher at a Lon­don com­pre­hen­sive. Most of the chil­dren I taught ar­rived at school aged 11 far be­hind where they we were ex­pected to be with their lan­guage, and as a novice I strug­gled to help them catch up. Whereas ev­ery­thing I tried seemed out­dated, Roy’s ap­proach was sci­en­tific. I hoped his find­ings would un­lock a secret that could help kids to re­alise their full po­ten­tial. If we could cre­ate ma­chines that learned like hu­mans, could we also de­velop ones that could help us per­fect hu­man learn­ing?

Be­fore press­ing record, Roy and Pa­tel agreed some ground rules. The record­ings would be avail­able only to their most trusted in­ner cir­cle of re­searchers. If at any time they felt un­com­fort­able with the film­ing, they would junk the footage. When pri­vacy was re­quired, the sys­tem could be tem­porar­ily shut down. It was a leap of faith, but they agreed it was worth it. Their ex­per­i­ment had the power to un­lock new in­sight into the work­ings of the in­fant mind. Toco was Pinoc­chio to Roy’s Gep­petto. But whereas he was won­der­ing what real kids could teach ro­bots, I wanted to know if those home videos might hint at how to en­hance learn­ing for the youngest hu­mans.

In 1995, two re­searchers, Betty Hart and Todd Ris­ley, pub­lished the re­sults of a study in which they trailed 42 Kansas City fam­i­lies to com­pare the ex­pe­ri­ences of preschool­ers from poor fam­i­lies with their richer peers. Start­ing when the in­fants were nine months old, they ob­served them reg­u­larly over a two-and-a-half-year pe­riod, record­ing and tran­scrib­ing all par­ent-and-child speech dur­ing their hour-long vis­its. The find­ings were stark. The num­ber of words a child heard by their third birth­day strongly pre­dicted aca­demic suc­cess aged nine. The dif­fer­ence was barely fath­omable. They es­ti­mated that, at the age of four, the rich­est kids had heard 30m more words than the poor­est.

“The prob­lem of skill dif­fer­ences among chil­dren at the time of school en­try is big­ger, more in­tractable and more im­por­tant than we thought,” Hart and Ris­ley said. Their re­search showed it was worth in­ter­ven­ing as early as pos­si­ble. “The longer the ef­fort is put off, the less pos­si­ble change be­comes.”

If the prob­lem was stark, the so­lu­tion seemed sim­ple. There was a gap, and it had to be filled with words. Hart and Ris­ley’s find­ings fu­elled a word-rush that en­dures to­day. Across the English­s­peak­ing world, par­ents flocked to buy flash­cards and brain train­ers for their tots.

But my ex­pe­ri­ence in the class­room sug­gested that the in­ter­pre­ta­tion was a lit­tle sim­plis­tic, equat­ing the de­vel­op­ment of the hu­man mind with the in­puts and out­puts of com­put­ers. I sus­pected that there was more to in­fant learn­ing than the quan­tity of words you heard.

A pro­fes­sor of early child­hood de­vel­op­ment at Tem­ple Univer­sity in

‘The learn­ing in­dus­try has con­vinced many that the mem­o­ri­sa­tion of con­tent is all that’s needed’

Penn­syl­va­nia, Kathy Hirsh-Pasek, seemed to agree. She had writ­ten that “just as the fast food in­dus­try fills us with empty calo­ries, what we call the ‘learn­ing in­dus­try’ has con­vinced many among us that the mem­o­ri­sa­tion of con­tent is all that is needed for learn­ing suc­cess and joy­ful lives”. She had also writ­ten an in­flu­en­tial book that laid out her reser­va­tions about the word-rush: Ein­stein Never Used Flash­cards: How Chil­dren Re­ally Learn and Why They Need to Play More and Mem­o­rize Less. I thought she might have some an­swers.

Hirsh-Pasek is leg­endary in the field of early child de­vel­op­ment. The au­thor of 12 books and hun­dreds of aca­demic ar­ti­cles, she is a dis­tin­guished fac­ulty fel­low who runs Tem­ple’s In­fant and Child Lab­o­ra­tory, whose slo­gan is “Where Chil­dren Teach Adults”.

At the lab, sci­en­tists were putting tiny hu­mans through their paces. Re­searchers had de­vel­oped in­ge­nious ex­per­i­ments that mea­sured changes in heart rate to show some of the things that eight­month-olds al­ready knew. “They know the mo­bile won’t fall on them,” said Hirsh-Pasek. “They know that if I drop this plate on the ta­ble, the plate won’t go through the ta­ble. That’s amazing. They know that if I’m sit­ting across from you, and you can’t see the bot­tom part of my body, I still have one.”

Un­til re­cently, sci­en­tists had tended to think of in­fants as ir­ra­tional, il­log­i­cal and ego­cen­tric. In his Prin­ci­ples of Psy­chol­ogy in 1890, Wil­liam James had de­scribed ba­bies’ ex­pe­ri­ence of sen­sory over­load: “The baby, as­sailed by eyes, ears, nose, skin, and en­trails at once, feels it all as one great bloom­ing, buzzing con­fu­sion.” This un­der­stand­ing had con­trib­uted to a mech­a­nis­tic view of learn­ing, and the idea that the sheer rep­e­ti­tion of words was what mat­tered most. But it wasn’t true.

Even in utero, ba­bies are learn­ing. At that stage, they pick up sounds. One-hour-olds can dis­tin­guish their mother’s voice from an­other per­son’s. They ar­rive in the world with a brain primed to learn through sen­sory stim­u­la­tion. We are nat­u­ral-born explorers, ready made for sci­en­tific in­quiry. We have to un­der­stand this if we were to re­alise our learn­ing po­ten­tial.

“We enter the world ready to ‘read the per­fect cues out of the en­vi­ron­ment’,” said Hirsh-Pasek. I thought back to Toco. He read the en­vi­ron­ment, too – or at least what his eye cam­eras saw and ear mi­cro­phones heard. But ro­bots can only reach out in ways they have been pro­grammed to, can only learn from stim­uli they were in­structed to pay at­ten­tion to. It lim­its them to a small range of ex­pe­ri­ences that would shape their be­hav­iours. There is no mean­ing in their meth­ods. Ba­bies, on the other hand, are so­cial learn­ers.

“We ar­rive ready to in­ter­act with other hu­mans and our cul­ture,” said Hirsh-Pasek. The real ge­nius of hu­man ba­bies is not sim­ply that they learn from the en­vi­ron­ment – other an­i­mals can do that. Hu­man ba­bies can un­der­stand the peo­ple around them and, specif­i­cally, in­ter­pret their in­ten­tions.

As we evolved, so­cial and cul­tural trans­mis­sion be­came pos­si­ble. Lan­guage was our start­ing point – the pos­si­bil­ity of two be­ings as­crib­ing a shared mean­ing to an oth­er­wise ab­stract con­cept or sym­bol.

Couldn’t we see the be­gin­nings of this in ba­bies’ be­hav­iour? In­fants under a year en­gaged in proto-con­ver­sa­tions with car­ers. They bab­bled away, held eye con­tact, ex­changed things, mim­icked their ex­pres­sions or ac­tions. They also ex­per­i­mented with tools, stick­ing them in their mouths, bash­ing them on things.

At the Max Planck In­sti­tute for Evo­lu­tion­ary An­thro­pol­ogy in Leipzig, Prof Michael To­masello wrote that our young learn “in an en­vi­ron­ment of ever-new arte­facts and so­cial prac­tices, which, at any one time, rep­re­sent some­thing re­sem­bling the en­tire col­lec­tive wis­dom of the en­tire so­cial group through­out its en­tire cul­tural his­tory”.

If all of us are to achieve our po­ten­tial as learn­ers, the ques­tion we have to an­swer is how we ought to shape this en­vi­ron­ment. Hu­man brains have spe­cially adapted to learn. Our long pe­riod of im­ma­tu­rity is a risky evo­lu­tion­ary strat­egy, mak­ing us vul­ner­a­ble early on to preda­tors or sick­ness, and de­lay­ing for many years our ca­pac­ity to re­pro­duce, but the pay­off is im­mense. We can ac­tively in­cor­po­rate enor­mous amounts of the lat­est in­for­ma­tion from our en­vi­ron­ment and so­cial group into our cog­ni­tive de­vel­op­ment.

Sci­en­tists have long recog­nised the na­ture-v-nur­ture de­bate as fal­lacy. A huge amount of our brain de­vel­op­ment takes place in the first three years. In those years, the brain grows in re­la­tion to the en­vi­ron­ment, form­ing it­self in in­ter­ac­tion with sen­sory ex­pe­ri­ence. As Hart and Ris­ley showed in their study of the word gap, that ex­pe­ri­ence can have a huge ef­fect on who that per­son be­comes.

We have evolved to be a species of teach­ers and learn­ers. Our abil­ity to un­der­stand other peo­ple ar­rives around the ninth month, at a mo­ment in their de­vel­op­ment at which ba­bies be­gin to check the at­ten­tion of oth­ers by hold­ing or point­ing at ob­jects. At a year, they can fol­low an­other’s at­ten­tion, gaz­ing at, touch­ing or lis­ten­ing to the same thing. At 15 months they can di­rect it. Lis­ten to that! Look over there! Shared at­ten­tion is the start­ing point of con­scious hu­man learn­ing. It is why in­fants don’t learn to talk from video, au­dio or over­hear­ing parental con­ver­sa­tions. We haven’t evolved to. That’s why it mat­ters that we talk to our chil­dren. It’s also why we can’t learn from ro­bots – yet.

The im­pli­ca­tion for un­der­stand­ing how we learn sounds like com­mon sense: each gen­er­a­tion ought to en­sure the next is steeped in their ear­li­est years in the tools, sym­bols and so­cial prac­tices of the cur­rent cul­ture.

In search of the kind of learn­ing en­vi­ron­ment that might best cul­ti­vate our nat­u­ral abil­i­ties, I vis­ited Pen Green Early Child­hood Cen­tre, a spe­cial­ist cen­tre in early child de­vel­op­ment in the Northamp­ton­shire town of Corby in the UK. The out­door space was cold and over­cast, but that wasn’t de­ter­ring the chil­dren. By a bam­boo bush, two small boys splashed at an ever-run­ning tap. “Don’t get me wet!” they squeaked with de­light. A teacher bent down to com­fort a tod­dler in a “Be Fast or Come Last” T-shirt. Four small girls were deep in a se­ri­ous con­ver­sa­tion while ab­sent-mind­edly dig­ging sand into colour­ful buck­ets.

Pen Green had a global rep­u­ta­tion for ex­cel­lence in early child de­vel­op­ment and fam­ily sup­port, a pro­to­type that had in­spired suc­ces­sive early-years in­ter­ven­tions by gov­ern­ment, in­clud­ing Sure Start and Early Ex­cel­lence. I spoke to the di­rec­tor, An­gela Prodger. She had just taken over from the leg­endary Margy Whalley, who set up the cen­tre in 1983. In the 1980s, Corby was among the UK’s poor­est towns, its pop­u­la­tion of Scot­tish mi­grant work­ers un­moored by the clo­sure of the steel­works for which they had moved south – 11,000 peo­ple had been made re­dun­dant. The cen­tre was in­tended as a lifeline for the next gen­er­a­tion. To­day it serves 1,400 of the UK’s least well-off house­holds.

I asked her about lan­guage learn­ing. We knew words mat­tered, but I’d not heard much talk at play­time. “If we’re not ad­dress­ing per­sonal, so­cial, emo­tional de­vel­op­ment first, you’re not ready to

Ro­bots learn from stim­uli they are in­structed to pay at­ten­tion to – it lim­its their range of ex­pe­ri­ences

learn,” said Prodger. She ex­plained that be­fore chil­dren could ac­quire the tools of speech and lan­guage, you had to en­sure they felt a sense of “be­ing and be­long­ing”. Too fre­quently, she thought, our ap­proaches to early learn­ing skipped these steps. It sounded to me like a nice-to-have, not an es­sen­tial, but re­search showed oth­er­wise.

In the 1950s, Bri­tish psy­cho­an­a­lyst John Bowlby pro­posed a the­ory of “at­tach­ment”. He hy­poth­e­sised that in­fants, un­able to reg­u­late their own feel­ings, were prone to get up­set when they were hun­gry, sad or lonely. A carer was needed to help them “co-reg­u­late” their feel­ings, which over time would teach the child to self-reg­u­late, pro­vided their early ex­pe­ri­ences helped them do so. If neg­a­tive ex­pe­ri­ences weren’t al­le­vi­ated with love from a parental fig­ure, they could be­come es­tab­lished.

The im­pli­ca­tions for chil­dren grow­ing up in poverty-stricken or trau­matic en­vi­ron­ments were sig­nif­i­cant. This was why Pen Green took care to put the be­ing and be­long­ing of its chil­dren first. “Be­hav­iour is al­ways just a sign of chil­dren try­ing to tell you some­thing,” said Prodger.

As we toured the build­ing, Prodger told me that the skill of the prac­ti­tion­ers at Pen Green was in learn­ing to at­tend to what was go­ing on in the minds of the kids, and in­ter­pret­ing it as ev­i­dence of what the young­sters were sig­nalling, even be­fore they were able to ver­balise it them­selves. Chil­dren were con­stantly com­mu­ni­cat­ing with us, she told me. We just had to learn to un­der­stand.

“It’s about look­ing,” Prodger said. “What are the chil­dren try­ing to ex­plore? What are they try­ing to find out?”

Cre­ative play is the foun­da­tion on which cre­ativ­ity, lan­guage, maths and sci­ence are built. If you start too early with flash­cards, you lose this de­vel­op­men­tal stage. “It’s about be­ing free,” Prodger said. “It’s about risk-tak­ing.”

They take the kids out to the for­est a few days a week, light fires, let them ex­per­i­ment with scis­sors and ride BMX bikes. If they want to be out­side, they go out­side. If they fancy re­turn­ing to the snug, where the youngest in­fants roll around, that’s where they would go. The en­vi­ron­ment dic­tates the learn­ing. The adults aim only to con­nect and share at­ten­tion with the chil­dren. Read­ing and writ­ing could wait. Nurs­eries ought to be as so­cial as pos­si­ble, and fol­low kids’ lead in their play. Be­fore kids can get on with learn­ing, we have to en­sure they be­long.

The chil­dren seemed happy here, learn­ing to be­long and lay­ing down foun­da­tions for their fu­ture suc­cess through play. And yet I won­dered if we couldn’t do still more to ac­cel­er­ate early learn­ing. The im­pli­ca­tion of Deb Roy’s ro­bot ex­per­i­ment was that ev­ery mo­ment counted. Could we af­ford to leave so much to chance?

“The ac­ci­dent of birth is the great­est source of in­equal­ity in the US,” wrote econ­o­mist James Heck­man. It’s equally true in the UK to­day, where the strong­est pre­dic­tor of aca­demic achieve­ment is how much your par­ents earn. Though two-thirds of our kids at­tain a C or above in English and maths GCSEs each year, that num­ber falls to just over a third of kids on free school meals. Heck­man has also shown that the best way to tackle this in­equal­ity is to in­vest in chil­dren’s de­vel­op­ment as early as pos­si­ble in their lives. It isn’t enough to trans­form schools – we have to start much ear­lier than that.

At Tem­ple Univer­sity, Hirsh-Pasek told me that we can’t sim­ply drop kids in front of iPads and ex­pect them to catch up – but that doesn’t mean we should give up en­tirely on in­tel­li­gent ma­chines. Some of her lab’s ex­per­i­ments are aimed at clos­ing de­vel­op­men­tal gaps be­tween rich and poor kids. Oth­ers cover top­ics such as lan­guage de­vel­op­ment and spa­tial aware­ness, and all use tech­nol­ogy in dif­fer­ent ways. “What the ma­chine can’t do is be a part­ner,” Hirsh-Pasek told me. “It isn’t so­cial. It’s in­ter­ac­tive with­out be­ing adap­tive.”

Hirsh-Pasek’s mis­sion was to change the way we thought about learn­ing, es­pe­cially for the poor­est kids. “We had this vi­sion that it was so im­por­tant to get the ba­sics into poor kids,” she told me. “We thought we should drop re­cess – even though we know be­ing phys­i­cal helps kids learn, helps build bet­ter brains. And we thought we should just do read­ing and maths, and cut out the arts and all this su­per­flu­ous stuff like so­cial stud­ies.”

It weighed heav­ily on her. Pol­i­cy­mak­ers and lay­men had twisted the sci­ence to fit their own ends. No sci­en­tist thought flash­cards worked. No sci­en­tist be­lieved you should start learn­ing to read and write at an ever younger age. It was a fan­tasy of gov­ern­ments. More re­cent

re­search has added depth to the lan­guage lessons of Betty Hart and Todd Ris­ley’s Kansas Study. In 2003, the psy­chol­o­gist Pa­tri­cia Kuhl ex­per­i­mented with teach­ing Amer­i­can in­fants Man­darin. Split into three groups (video, au­dio and flesh-and-blood teacher) only those with a hu­man tu­tor learned any­thing at all.

Schools are still guilty of ig­nor­ing these in­sights into in­fant learn­ing. Erika Chris­takis, ear­ly­child­hood ex­pert and au­thor of The Im­por­tance of Be­ing Lit­tle, charts the slow de­scent in preschool learn­ing from a mul­tidi­men­sional, ideas-based ap­proach to a two-di­men­sional nam­ing-and-la­belling cur­ricu­lum. Daphna Bas­sok at the Univer­sity of Vir­ginia asks if kinder­garten is re­ally the new first grade. The ex­pec­ta­tion that kinder­garten­ers – aged five or six – can read is now com­mon­place. Yet this is counter to all the ev­i­dence. A Cam­bridge study com­par­ing groups of chil­dren who started for­mal lit­er­acy lessons at five and seven found that start­ing two years ear­lier made no dif­fer­ence at all to a child’s read­ing abil­ity aged 11, “but the chil­dren who started at five de­vel­oped less-pos­i­tive at­ti­tudes to read­ing, and showed poorer text com­pre­hen­sion than those who started later”.

These find­ings are clear: if you start on the de­cod­ing be­fore you have an un­der­ly­ing un­der­stand­ing of story, ex­pe­ri­ence, sen­sa­tion and emo­tion, then you be­come a worse reader. And you like it less. Treat kids like ro­bots dur­ing early learn­ing and you put them off for life.

In­stead, Hirsh-Pasek wanted kids to em­brace the joy in learn­ing and grow­ing up. Apart from kids, her other great love was mu­sic. She of­ten used to break into song, es­pe­cially on the phone to her grand­daugh­ter.

In her book, she sug­gested six Cs for mod­ern learn­ing: col­lab­o­ra­tion, com­mu­ni­ca­tion, con­tent, crit­i­cal think­ing, cre­ative in­no­va­tion and con­fi­dence. Truisms, I had thought, but un­like much ed­u­ca­tion pol­icy, drawn from sci­en­tific ev­i­dence. If I was to take away one thing, she said, it should be that “from the ear­li­est ages, we learn from peo­ple”.

It was the same in­sight that had prompted a pair of sub­ur­ban sci­en­tists to hit the “record” but­ton.

Deb Roy was dressed in black and still looked youth­ful when we met at MIT. A few flecks of grey in his hair were the only ev­i­dence of 11 years of par­ent­hood. Look­ing back, the Hu­man Spee­chome Project – as his and Pa­tel’s home-record­ing ex­per­i­ment had been named – seemed a quirk of turn-of-the-mil­len­nium en­thu­si­asm about ar­ti­fi­cial in­tel­li­gence. In all, they had cap­tured 90,000 hours of video and 140,000 hours of au­dio. The 200 ter­abytes of data cov­ered 85% of the first three years of their son’s life (and 18 months of his lit­tle sis­ter’s). But now the footage had been gath­er­ing dust. “I still have the whole col­lec­tion,” he said. “I’m wait­ing for his wed­ding day, just to bore the hell out of ev­ery­one.”

In a way, it was also a great lost home video. With his team at MIT, Roy had de­vel­oped new ap­proaches to visu­al­is­ing and study­ing the data they had cap­tured: “So­cial Hotspots” showed two tightly knot­ted lines, vis­ual traces of ten­der mo­ments in which par­ent and child came to­gether to chat, learn or ex­plore; “Word­scapes” were snow-capped moun­tains ranged through­out the liv­ing room and kitchen, the high­est peaks ris­ing where par­tic­u­lar words were most of­ten heard. The tools had turned out to be fan­tas­ti­cally lu­cra­tive as a means for analysing talk on Twit­ter. Roy and a grad­u­ate stu­dent had spent the decade build­ing a new me­dia com­pany.

Roy was now back at MIT. His new group was called the Lab­o­ra­tory for So­cial Ma­chines. He had given up build­ing ro­bots that would com­pete with hu­mans and in­stead turned his at­ten­tion to the aug­men­ta­tion of hu­man learn­ing. What had changed his mind was the process of ac­tu­ally rais­ing a child.

The first time his son ut­tered some­thing that wasn’t just bab­ble, Roy was sit­ting with him look­ing at pic­tures. “He said ‘fah’,” Roy ex­plained, “but he was ac­tu­ally clearly re­fer­ring to a fish on the wall that we were both look­ing at. The way I knew it was not just co­in­ci­dence was that right af­ter he looked at it and said it, he turned to me. And he had this kind of look, like a car­toon light­bulb go­ing off – an ‘Ah, now I get it’ kind of look. He’s not even a year old, but there’s a con­scious be­ing, in the sense of be­ing self-re­flec­tive.”

“I guess, putting on my AI hat, it was a hum­bling les­son,” he con­tin­ued. “A les­son of like, holy shit, there’s a lot more here.”

Roy was no longer sure you could bring a ro­bot up like a real hu­man – or that we should even try. It didn’t seem there was much to gain by de­vel­op­ing ro­bots that took ex­actly one hu­man child­hood to be­come ex­actly like one young adult hu­man. That’s what peo­ple did. And that was be­fore you got into imag­i­na­tion or emo­tions, iden­tity or love – things that were im­pos­si­ble for Toco. Watch­ing his son, Roy had been blown away by “the in­cred­i­ble so­phis­ti­ca­tion of what a lan­guage learner in the flesh ac­tu­ally looks like and does”. In­fant hu­mans didn’t only re­gur­gi­tate; they cre­ated, made new mean­ing, shared feel­ings.

The learn­ing process wasn’t de­cod­ing, as he had orig­i­nally thought, but some­thing in­fin­itely more con­tin­u­ous, com­plex and so­cial. He was read­ing He­len Keller’s au­to­bi­og­ra­phy to his kids, and had been struck by her epiphany at un­der­stand­ing lan­guage for the first time. Deaf and blind af­ter an ill­ness in in­fancy, Keller was seven years old when she got it. “Sud­denly I felt a misty con­scious­ness as of some­thing for­got­ten,” she wrote, “a thrill of re­turn­ing thought; and some­how the mys­tery of lan­guage was re­vealed to me. I knew then that ‘w-a-t-e-r’ meant the won­der­ful cool some­thing that was flow­ing over my hand. That liv­ing word awak­ened my soul, gave it light, hope, joy, set it free! Ev­ery­thing had a name, and each name gave birth to a new thought. As we re­turned to the house ev­ery ob­ject which I touched seemed to quiver with life.”

Roy had re­cently started work­ing with Hir­shPasek, fol­low­ing her in­sight that ma­chines might aug­ment learn­ing be­tween hu­mans, but would never re­place it.

He had dis­cov­ered that hu­man learn­ing was com­mu­nal and in­ter­ac­tive. For a ro­bot, the ac­qui­si­tion of lan­guage was ab­stract and for­mu­laic. For us, it was em­bod­ied, emo­tive, sub­jec­tive, quiv­er­ing with life. The fu­ture of in­tel­li­gence wouldn’t be found in our ma­chines, but in the de­vel­op­ment of our minds.

In­fant hu­mans didn’t only re­gur­gi­tate – they cre­ated, made new mean­ing and shared their feel­ings

MIT Me­dia Lab

Do­mes­tic ex­per­i­ment … Deb Roy and his in­fant son in their home, which was fit­ted with cam­eras to record the child’s de­vel­op­ment

Newspapers in English

Newspapers from UK

© PressReader. All rights reserved.