Ma­chine learn­ing is much more than a set of clever al­go­rithms; it’s chang­ing ev­ery­thing


Ma­chine learn­ing is ab­so­lutely ev­ery­where now; it’s the com­put­ing zeit­geist. Dig­i­tal as­sis­tants use it to trans­late speech. Face­book uses it to au­to­mat­i­cally tag im­ages (Deep­Face has a claimed ac­cu­racy of 97.25 per­cent). Uber uses it to cal­cu­late ar­rive times and pickup lo­ca­tions. Ama­zon, Spo­tify, and Net­flix use it to se­lect the rec­om­men­da­tions they serve to you. Google in­tro­duced RankBrain to its search en­gine in 2015, to un­tan­gle query se­man­tics, Gmail has its Smart Re­ply ser­vice, and last year, Google Maps started us­ing ma­chine learn­ing to ex­tract street names and house num­bers from im­ages.

These are just the tip of the ice­berg. Be­hind the scenes, big busi­ness has taken ma­chine learn­ing to heart. Fi­nan­cial in­sti­tu­tions use it to track mar­ket trends, cal­cu­late risks, and delve into the streams of fi­nan­cial data. The race to the self-driv­ing car is be­ing pow­ered by ma­chine learn­ing. Any busi­ness that is try­ing to see pat­terns in the un­tidy mass of data hu­mans cre­ate can make some use of ma­chine learn­ing, from the FBI, med­i­cal re­search, and in­sur­ance down. Cur­rently, the five big­gest com­pa­nies by mar­ket value in the world are Ap­ple, Ama­zon, Al­pha­bet (Google), Mi­crosoft, and Face­book. Every one of these has made mas­sive in­vest­ments in ma­chine learn­ing, and has ac­cess to the kind of big data its ef­fec­tive use re­quires.

One thing to re­mem­ber is that ma­chine learn­ing and ar­ti­fi­cial in­tel­li­gence are two dif­fer­ent things. Ar­ti­fi­cial in­tel­li­gence is a de­vice, pro­gram, or sys­tem that ap­pears “smart,” one that can act and re­act in a quasi-hu­man way. How this is achieved is a sep­a­rate thing. Ma­chine learn­ing is a method that AI can em­ploy to ap­pear smart but, more im­por­tantly, to learn as it op­er­ates. It is this abil­ity to adapt that is caus­ing the revo­lu­tion. PROGRAMING VER­SUS LEARN­ING A com­puter pro­gram is a set of in­struc­tions, all neat and log­i­cal. Given an in­put, you can trace the data through the pro­gram’s al­go­rithms, and ac­cu­rately pre­dict the out­put. This is the tra­di­tional way that com­put­ers work—in­flex­i­ble, and quite un­like the real world. Each task re­quires

the pro­gram to be writ­ten to deal with that spe­cific task.

Ma­chine learn­ing is a way to cre­ate a pro­gram that apes the hu­man brain, rather than a cal­cu­lat­ing en­gine. In­stead of try­ing to pro­gram for every even­tu­al­ity, every task, every type of data, you cre­ate a ma­chine that thinks more like a hu­man. En­ter the con­cept of the neu­ral net­work: a sys­tem de­signed to mimic the fluid way the hu­man brain cre­ates and changes its in­ter­nal con­nec­tions as it learns.

It all sounds won­der­fully mod­ern and fresh, only it isn’t. Like many programing con­cepts, it’s ac­tu­ally not that new. The prac­ti­cal ca­pa­bil­i­ties of ma­chines have long lagged be­hind ideas. The first rec­og­niz­able com­puter al­go­rithms were writ­ten around 1837, at a time when hard­ware was an un­fin­ished mass of cogs and gears.

Ma­chine learn­ing isn’t quite that old; the idea is gen­er­ally cred­ited to Arthur Sa­muel, who coined the term “ma­chine learn­ing” in 1959, and helped write a check­ers-play­ing pro­gram that learned from its mis­takes. Its roots are deeper still, though. Back in 1943, neu­ro­phys­i­ol­o­gist War­ren Mc­Cul­loch and math­e­ma­ti­cian Wal­ter Pitts de­scribed how a neu­ral net­work op­er­ates, and mod­eled one us­ing elec­tri­cal cir­cuits. In 1949, the psy­chol­o­gist Don­ald Hebb gave us Hebb’s Law: “Neu­rons that fire to­gether, wire to­gether,” fun­da­men­tal to the way in­ter­nal con­nec­tions are re­in­forced by learn­ing.

Progress was slow, mainly due to the tech­ni­cal lim­i­ta­tions of avail­able hard­ware, and the lack of key al­go­rithms, such as back prop­a­ga­tion (used to cal­cu­late weights, more later). De­spite decades of re­search and the­ory, it took un­til the 1990s be­fore we had wide­spread use­able ma­chine learn­ing pro­grams, and it took un­til the 2010s be­fore large neu­ral net­works were fea­si­ble.

A neu­ral net­work con­sists of a set of in­put and out­put units. Be­tween these is a grid of ar­ti­fi­cial neu­rons, called nodes. These are ar­ranged into lay­ers, the out­put from one layer feed­ing into the next. Con­nec­tions are as­signed a weight, a level of in­flu­ence, which changes as the net­work learns. Each node is fed data, per­forms an op­er­a­tion, and sends the re­sult on, ac­cord­ing to the weight.

A neu­ral net­work needs teach­ing (OK, not all of them, more later). A com­mon

use is pat­tern recog­ni­tion in im­ages. So, for ex­am­ple, if you feed your net­work an im­age, la­beled as “con­tains cat,” the net­work knows the de­sired out­put, so the con­fig­u­ra­tion of the node weight­ing is sim­ple. Then you feed in more im­ages: black cats, ginger cats, run­ning cats, sleep­ing cats, par­tial im­ages, and im­ages without cats. Each time, the con­nec­tion weight­ing needs to be ad­justed, so that the cor­rect out­put is main­tained, not only for the cur­rent im­age, but for all pre­vi­ous ones, too. As the net­work learns, the weight of each con­nec­tion is es­tab­lished be­tween the nodes.

A sys­tem of pat­tern recog­ni­tion now emerges within the net­work—it will learn the out­line of a cat, the po­si­tion of its eyes and ears, and so forth. Even­tu­ally, you will be able to feed in an im­age without re­veal­ing the cor­rect out­put, and the sys­tem will cor­rectly see, or not see, a cat. This is the clas­si­fi­ca­tion model; it at­tempts to pre­dict one or more fixed out­comes. The re­gres­sion model is sim­i­lar, but the out­put is a con­tin­u­ous vari­able, say a dol­lar value, or float­ing point num­ber.

The magic of a neu­ral net­work is that the in­ter­nal process is hid­den. You may have cre­ated it, but once trained to a task, you don’t ac­tu­ally know what the weights are at each node. It may be us­ing the shape of the cat’s eye to iden­tify them, or it may not, you don’t re­ally know ex­actly how it achieves the re­sults it does, which is a long way from the fixed logic of a tra­di­tional pro­gram.

There is an ob­vi­ous draw­back: It is fal­li­ble. There is a level of ac­cu­racy, and there is never the guar­an­tee of 100 per­cent; an un­usual con­cept for a com­puter pro­gram, where out­put er­rors are seen as bugs, er­rors in the code logic that can be cor­rected. A neu­ral net­work, in be­com­ing more hu­man, has gained our abil­ity to err, too. This isn’t too much of a is­sue when we are tag­ging pic­tures of cats for Face­book, but is some­thing to bear in mind when de­sign­ing a sys­tem for a self-driv­ing car.

There are other is­sues, too. It’s some­times dif­fi­cult to work out why one re­sult out­ranks another, as so much of it is hid­den in the in­ter­nal weight­ing. This can also make it dif­fi­cult to fine-tune. There has re­port­edly been much in­ter­nal dis­cus­sion at Google over the rel­a­tive mer­its of ma­chine learn­ing over its ri­vals for rank­ing search re­sults and ad­vert tar­get­ing—it doesn’t lend it­self to a quick tweak when you want to push one re­sult over another.

A sim­ple neu­ral net­work can run on a hand­ful of nodes. De­tailed work re­quires some­thing a lit­tle larger. Face­book’s Deep­Face runs on nine lay­ers and has 120 mil­lion con­nec­tion weights. For con­trast, the hu­man brain is widely quoted as hav­ing 100 bil­lion neu­rons, al­though a re­cent, and rather grisly, ex­per­i­ment in­volv­ing sam­pling a hu­man brain soup re­vealed

a fig­ure of 86 bil­lion. Each is con­nected to some­where be­tween 1,000 and 10,000 other neu­rons (no­body is re­ally sure; 7,000 is of­ten cited as a fair guess). This puts the num­ber of con­nec­tions into the tril­lions, more than there are stars in the Milky Way. Wow. The largest ar­ti­fi­cial sys­tems built so far touch one bil­lion con­nec­tions, and these have been short-lived re­search projects. We still have a long way to go.

As you might have guessed, we’ve stuck to the ba­sics here. Ma­chine learn­ing is not all neu­ral net­works, for starters; sup­port vec­tor ma­chines are another pop­u­lar method, which are trained in a sim­i­lar way, but use a dif­fer­ent math­e­mat­i­cal model in­ter­nally. These are sim­pler, don’t re­quire huge amounts of com­pu­ta­tional power or big data sets, and the in­ter­nal work­ings are more open to ex­am­i­na­tion. But they don’t have the power or scale of a neu­ral net­work.

Ma­chine learn­ing is a sub­ject that be­comes com­pli­cated very quickly as you delve deeper. We can’t even be­gin to list the ba­sic ma­chine learn­ing method­olo­gies; there are over 50. These use a whole slew of sta­tis­ti­cal anal­y­sis tools, de­ci­sion tree al­go­rithms, di­men­sion re­duc­tion, re­gres­sion anal­y­sis, and heaps more. This is high-grade math.

As well as the su­per­vised learn­ing out­lined, there are semi-su­per­vised sys­tems, which use a min­i­mum of la­beled data, and fully un­su­per­vised sys­tems. These work with no la­bels at all; you just feed in raw data, and let the al­go­rithms go to work. From these emerge pat­terns of clus­ters and as­so­ci­a­tions that might not be ob­vi­ous any other way.

You can only train a sys­tem if you know the out­put cri­te­ria you’re look­ing for. If you didn’t know you were look­ing for cats, you can’t train a sys­tem to find them. Un­su­per­vised sys­tems are use­ful for cre­at­ing data la­bels, which can then be fed back into su­per­vised sys­tems; for ex­am­ple, find­ing a clus­ter of im­ages that ap­pear to con­tain the same ob­ject. They are also good at find­ing anom­alies in data, ideal for se­cu­rity sys­tems look­ing for signs of fraud or hack­ing, but when you have no idea where or how these are go­ing to be made.

Another much-used ma­chine learn­ing buzz­word is deep learn­ing, es­sen­tially just used to de­scribe large and multi-lay­ered neu­ral net­works. In im­age recog­ni­tion sys­tems, for ex­am­ple, lay­ers may be used to di­vide im­ages into ar­eas or blocks that may be ob­jects; the next layer may try to de­fine edges; and fur­ther lay­ers iden­tify spe­cific shapes, end­ing with a train­able out­put. The more lay­ers, the greater the so­phis­ti­ca­tion, as the in­put is bro­ken down into an in­creas­ingly ab­stract rep­re­sen­ta­tion of the data. Sim­ple neu­ral net­works may only have a few lay­ers; a deep learn­ing

sys­tem can run to three fig­ures. They scale well, but do re­quire sig­nif­i­cant re­sources.

What ma­chine learn­ing needs to thrive is ac­cess to a lot of data. This has now been provided, thanks to the In­ter­net, by us. We’ve typed in un­told search re­quests, emails, and blogs, up­loaded mil­lions upon mil­lions of im­ages and videos, cre­ated pur­chase his­to­ries, travel his­to­ries, we’ve shared things we like, what we’ve seen, heard, and read, and more. This is big data, a set large enough to re­veal un­der­ly­ing pat­terns, as­so­ci­a­tions, and be­hav­iors. We’ve been feed­ing data into the In­ter­net for years now, and an aw­ful lot of that is sit­ting in data farms ripe for pro­cess­ing.

The other thing it re­quires is pro­cess­ing power. The GPU proved to be just what the sys­tems needed for the sim­ple but repet­i­tive op­er­a­tions, and now we have ded­i­cated hard­ware from Google and IBM, with In­tel and oth­ers to fol­low (see box be­low). Wire­less In­ter­net is in our homes, and the hard­ware re­quired to con­nect de­vices is cheap and plen­ti­ful. Add these fac­tors to­gether, and we have the per­fect storm for an ex­plo­sion in ma­chine learn­ing.

We have reached the point where it has be­come easy and rel­a­tively cheap to add ma­chine learn­ing voice con­trol or ges­ture

recog­ni­tion to some­thing as mun­dane as your tele­vi­sion. Mod­ern sets are al­ready con­nected to your wire­less hub; it’s a sim­ple mat­ter to route your voice com­mand or im­ages to a server run­ning one of the pop­u­lar ma­chine learn­ing frame­works— there are dozens, in­clud­ing Google’s TensorFlow and Ama­zon’s Ma­chine Learn­ing. Here, a neu­ral net­work quickly trans­lates your com­mand and routes it back to your tele­vi­sion. And as if by magic, you can in­struct your tele­vi­sion to turn down the vol­ume by voice or move­ment, rather than be both­ered to press a but­ton. Ten years ago, this would have been dif­fi­cult and im­pres­sive; now, it’s noth­ing un­usual.

We’ve be­come ac­cli­ma­tized to vir­tual or dig­i­tal as­sis­tants quickly, too. These have only been on the gen­eral mar­ket since 2011, the first be­ing Ap­ple’s Siri on the iPhone 4S. Last year, the mar­ket for smart speak­ers grew by over 100 per­cent, and they were Ama­zon’s best-sell­ing item. By the end of this year, it’s es­ti­mated that there will be over 100 mil­lion in peo­ple’s homes, and by the end of 2020, there will be over 225 mil­lion, all lis­ten­ing for the magic words.

These dig­i­tal as­sis­tants won’t stay in our homes; there are plans to bring them into your car, and your of­fice, as well as build them into other de­vices—a re­frig­er­a­tor you can talk to, any­one? These as­sis­tants won’t just talk to you, ei­ther. This year, Google demon­strated Google Du­plex, which will make phone calls for you. Cur­rently, it can only cope with mun­dane tasks, such as book­ing a reser­va­tion at a restau­rant. The sys­tem has a nat­u­ral-sound­ing voice, and even adds “ums” and “ers.” Add a neu­ral net­work sys­tem to an­swer calls, and it won’t be long be­fore we can have our dig­i­tal as­sis­tant call their dig­i­tal as­sis­tant.

One of the big­gest, and most public, ap­pli­ca­tions of ma­chine learn­ing in the real world is the self-driv­ing car. This lit­tle project is cur­rently soak­ing up bil­lions of dol­lars of re­search fund­ing, and has a dozen or more of the world’s big­gest firms chas­ing it, from gi­ants such as Google and tra­di­tional car com­pa­nies such as Ford, to a new gen­er­a­tion of firms, in­clud­ing Tesla and Uber. It is a huge test of the tech­nol­ogy, and re­sults have been gen­er­ally pos­i­tive, al­though not without mishap.

The kind of things that can throw a sys­tem are myr­iad, and of­ten un­ex­pected. While test­ing its sys­tem in Aus­tralia, Volvo found that it was con­fused by kan­ga­roos, un­able to work out which way they were go­ing. Ap­par­ently, the hop­ping was reg­is­tered as both mov­ing to­ward and away from the ve­hi­cle. In In­dia, Ta­tra’s sys­tem is strug­gling to cope with the strange va­ri­ety of mixed traf­fic on the roads. Auto-rick­shaws have proved a spe­cial prob­lem, as they are

of­ten dec­o­rated and cus­tom­ized to such a level that they be­come uniden­ti­fi­able.

De­spite tech­ni­cal hur­dles, the race to au­ton­o­mous cars ap­pears un­stop­pable. There are dozens of test pro­grams on public roads, and Ford plans to launch a fully au­ton­o­mous car by 2021, “in scale.” Early cars will be ex­pen­sive, and prob­a­bly rented out for use as a taxi ser­vice, rather than per­sonal trans­port. They also look likely to be ini­tially limited to care­fully mapped lo­ca­tions. We can be sure of a fuss— crashes by self-driv­ing cars still make the news even now, and the le­gal im­pli­ca­tions are in­ter­est­ing—but they are com­ing to a road near you soon.

Ma­chine learn­ing will soon be en­demic, hum­ming away in the back­ground, help­ing to run vir­tu­ally every in­sti­tu­tion with a net­work con­nec­tion, banks and gov­ern­ments, down to bots in our games, and as­sis­tants that sug­gest we or­der a pizza be­cause an al­go­rithm has cal­cu­lated that our be­hav­ior model in­di­cates this as a likely re­sponse to the four-hour gam­ing ses­sion it de­tected. You never know, per­haps one day we will have a ver­sion of Win­dows that can de­tect and cor­rect bugs as it op­er­ates.

A brave new world? Not a flat­ter­ing anal­ogy per­haps, but there is a hint of dystopia here. Along­side pa­tently wor­thy ap­pli­ca­tions, such as med­i­cal re­search, and use­ful tools, there is go­ing to be a moun­tain of mar­ket­ing. Our dig­i­tal foot­print is go­ing to be sifted and sorted, as com­pa­nies look for pat­terns and clus­ters, try­ing to pre­dict what we want, where we want it, and when we want it. You are the cus­tomer, but you will also be­come the prod­uct.

Our rights to on­line pri­vacy are still be­ing ham­mered out, too. How much of your life do you want pro­cessed by deep learn­ing ma­chines? Be­cause there is a whole gen­er­a­tion grow­ing up that sees a smart­phone as an es­sen­tial item, and ap­pears happy to share a de­tailed view of their lives on­line. The ram­i­fi­ca­tions for so­cial me­dia are al­ready be­com­ing clear, as in­creas­ingly so­phis­ti­cated bots learn to mimic real peo­ple. Twit­ter and Face­book are al­ready delet­ing mil­lions of ac­counts a year. Gov­ern­ments and or­ga­ni­za­tions have al­ways prac­ticed so­cial ma­nip­u­la­tion in one form or another, from slo­gans and bill­boards up­ward. In­te­grated ma­chine learn­ing sys­tems bring a whole new ar­mory into play.

At some point, bound­aries will have to be drawn. Some of them rather im­por­tant. It looks as though we are go­ing to have to ac­cept that peo­ple will oc­ca­sion­ally be run down by au­ton­o­mous cars, but what about be­ing shot by an au­ton­o­mous mil­i­tary ma­chine? These are al­ready out there (see box­out be­low). Are we get­ting too dark now? Prob­a­bly. Like any tool we in­vent, it is up to us how it is used. How­ever “smart” ma­chines be­come, they are still kinda stupid, too. A three-year-old child

doesn’t mis­take a pic­ture of a cake for one of a dog, an er­ror that even the best im­age recog­ni­tion sys­tem can make.

Many cur­rent ma­chine learn­ing projects have a nar­row field. They are aimed at spe­cific prob­lems with limited goals: im­prov­ing search en­gines re­sults or guid­ing a car through traf­fic. The next step is a more gen­er­al­ized ap­proach, where com­mon sense and an un­der­stand­ing of hu­man be­hav­ior and in­ter­ac­tion are the tar­get. Ma­chines that don’t just talk to or­der food for us, but can carry out a con­ver­sa­tion that you can en­joy, too.

The world of hu­man cog­ni­tive de­vel­op­ment is pro­vid­ing clues. We learn not only through trial and er­ror, but guided by a set of in­ter­nal in­stincts, na­ture, and nur­ture. Among the re­searchers in this new field is In­tel­li­gence Quest, a group based at MIT, and the Allen In­sti­tute for Ar­ti­fi­cial In­tel­li­gence. The next gen­er­a­tion of ma­chine learn­ing sys­tems should be able to re­act to the chaotic and un­pre­dictable hu­man world with more of that most hu­man of qual­i­ties: com­mon sense. So when it sees a cake shaped like a dog, it won’t be fooled.

Soon, we may look back at the “dumb” com­put­ers of the re­cent past, with their in­flex­i­ble logic, with some amuse­ment. In the 1980s, John Gage of Sun Mi­crosys­tems coined the phrase, “The net­work is the com­puter.” He was right, but it took years to sink in. To­day, we view any un­con­nected ma­chine as crip­pled. Per­haps we could now add one word to that phrase: “The neu­ral net­work is the com­puter.”

Nvidia trained this deep learn­ing im­age noise re­duc­tion sys­tem with 50,000 im­ages. The most im­pres­sive part is that it was trained only us­ing cor­rupted im­ages.

This lit­tle bird was gen­er­ated by a Mi­crosoft re­search project called the Draw­ing Bot. It was cre­ated from a text cap­tion: “This bird is red with white and has a very short beak.”

This stop sign, cre­ated by re­searchers at the Univer­sity of Wash­ing­ton, fools self- driv­ing cars—the added stick­ers are de­signed to throw the im­age recog­ni­tion off bal­ance.

A team at MIT has cre­ated RF-Pose, a sys­tem that uses wire­less sig­nals to track peo­ple, even through walls. It’s trained us­ing a cam­era and ra­dio sig­nals.

If you train a neu­ral net­work to iden­tify dogs, then process the same im­age 50 times, this is the re­sult; here thanks to Google’s DeepDream.

Feed thou­sands of im­ages into a neu­ral net­work, and you can cre­ate life­like, but ar­ti­fi­cial, im­ages. These were gen­er­ated by Nvidia us­ing a GAN, Gen­er­a­tive Ad­ver­sar­ial Net­work.

Adobe’s up­com­ing Cloak fea­ture can re­move se­lected el­e­ments from videos. No more painful frame-by-frame edit­ing re­quired, and the re­sults are im­pres­sive.

The Su­per aEgis II Sen­try gun can op­er­ate au­tonomously, track­ing tar­gets up to 2km.

These com­i­cal cars have been au­tonomously trundling around the Google cam­pus for a while. De­spite the con­trolled con­di­tions and low speeds, it has not gone without in­ci­dent.

A self- driv­ing Jaguar iPace car from Waymo (part of Al­pha­bet); these are start­ing tests in San Fran­cisco.

Us­ing ma­chine learn­ing to study all of Rem­brandt’s paint­ings, a team in the Nether­lands gen­er­ated this: “The Next Rem­brandt.” It’s com­pletely new, but un­mis­tak­ably in his style.

Newspapers in English

Newspapers from Australia

© PressReader. All rights reserved.