Page 79 - The Economist
P. 79

The Economist December 9th 2017                                                  Science and technology 79
       2 shows and finding ways to make that ma-  Environmental engineering   waters is unlikely to tempt anyone into
        terial more engaging.                                                harvesting them. Previous studies have
           Besidesfundamental algorithmswhich Clean-up mussel                shown that the ribbed mussel is both har-
        firms hope to apply to their own opera-                               dy and adept at collecting a range of trou-
        tions, NIPS is also home to applied re-                              blesome materials from its environment.
        search, particularly in health care and biol-                        DrGalimanyand DrRose thoughtitwould
        ogy. Becks Simpson from Maxwell MRI,a                                be ideal to help clean up the Bronx River
        startupfromBrisbaneinAustralia,showed  Anasty-tasting mussel could be justthe  Estuary in New York. With an industrial
        a way to combine magnetic resonance im-  job forcleaning rivers      waterfront and wastewaterrun-offs from a
        agingwithdeeplearningtoimprovethedi-                                 dense urban environment, the estuary has
        agnosis of prostate cancer. Elisabeth Ru-  HELLFISH thrive in waters rich in nutri-  a long history of suffering from harmful
        metshofer  from  Johannes  Kepler Sents. These include the nitrogen used in  bacteria and high levels ofnitrogen.
        University Linz presented a system that  fertiliser, which passes from the land into  With a group of colleagues they
        could automatically recognise and track  rivers and then into the sea. The shellfish  moored a six-square-metre commercial
        proteins in cells, helping to illuminate the  grow, as do the profits of those who har-  mussel-growing raft in the estuary and
        underlying biology. A team from Duke  vest them. The problem comes when dis-  populated it with ribbed mussels. They
        University in North Carolina had used  charges into the sea are tainted with more  closely monitored the health of the mus-
        machine learning to detect cervical cancer  noxious material, such as bacteria that  sels over six months and, using a flow-
        automatically using a pocket colposcope,  pose a threat to human health. Once the  through device, also analysed the chemis-
        to the same level of accuracy as a human  bugsare in the shellfish, theycan be passed  try of the water both before and after the
        expert. Some used AI to mine doctors’  on to anyone who eats them.   mussels had done their filtering. The re-
        notesto estimatethe chancesthata patient  This problem—and another, of excess  sults were impressive.
        willbereadmittedtohospital,tocategorise  nitrogen that can cause poisonous algal  The researchers found that not only did
        and understand the allergic reactions of  blooms—might be mitigated by shellfish  the mussels thrive in the polluted waters
        children and to model the geographic dis-  thatpeople don’teat, reckon Eve Galimany  of the Bronx River Estuary, but they also
        tribution of naloxone, which can help  and Julie Rose at the National Oceanic and  collected a lot of pollutants. More specifi-
        blockthe effects ofopioids, in order to get a  Atmospheric Administration at Milford  cally, a fully stocked raft ofmussels cleared
        bettergrip on the use ofsuch drugs.  Laboratory in Connecticut. As they report  an average of 12m litres of water daily, re-
           Other applications ranged from re-  in  Environmental Science & Technology,  moving 160 kilograms of particulate mat-
        searchers at the Federal University Lokoja  their chosen candidate for the job is the  ter, ofwhich 12 kilograms was absorbed by
        in Nigeria trying to use machine learning  ribbed mussel, more formally known as  the mussels’ digestive systems and inte-
        to identify potential child suicide bombers  Geukensia demissa.      grated into their bodies. The remainder
        to the DondersInstitute in the Netherlands  The ribbed mussel is edible, but it tastes  was excreted as waste, which drops down
        presenting a system that can reconstruct  terrible and so has no commercial value.  and is ultimately buried in the river sedi-
        pictures of faces that a person sees simply  This means growing the mussels in tainted  ment. The material filtered out by the mus- 1
        by scanning their brains. Google research-
        ers used machine learning to hide a com-
        plete image inside another picture of the  Gender in academia        University seminars, relative* share of questions
        same size. Whattheymightdo with that re-                             asked by women…
        mains to be seen.                    ONE theory to explain the low share of  Ten countries, 2014-16
           New hardware for machine learning  women in senior academic jobs is that  …when the first question
        was on display, too. At its party Intel un-  they have less self-confidence than men.  was asked by a man  Number of seminars
        veiled itslatestchip dedicated to solving AI  This hypothesis is supported by data in a          20
        problems.  NVIDIA, a chipmaking rival  new working paper, by a team of research-
        whose share price has increased ninefold  ers from five universities in America and               15
        in the past three years thanks to sales of its  Europe. In this study, observers counted
        graphical-processing units for deep learn-  the attendees, and the questions they                10
        ing, displayed its latest wares. Graphcore, a  asked, at 247 departmental talks and
        British startup, caused particular waves. It  seminars in biology, psychology and                5
        presented benchmarksforitschip’sperfor-  philosophy that took place at 35 universi-
        mance on common machine-learning     ties in ten countries. On average, half of
        tasks that tripled speeds for image recogni-  each seminar’s audience was female.  Women asked  Women asked  0
        tion and delivered a claimed 200 times im-  Men, however, were over 2.5 times more  fewer questions*  more questions*
        provement over NVIDIA for the kinds of  likely to pose questions to the speak-  …when the first question
        machine learning required for speech-rec-  ers—an action that may be viewed (right-  was asked by a woman
        ognition and translation applications.   ly or wrongly) as a sign of greater                     15
           Among older hands at NIPS, especially  competence.
        those who can remember its origins, there  This male skew in question-asking was                 10
        is a sense that the corporate obsession  observable, however, only in those semi-
        with machine learning will not last. They  nars in which a man asked the first ques-              5
        should not be so sure. The systems being  tion. When a woman did so, the gender
        developed are justbeginningto be a broad-  split in question-asking was, on average,             0
        ly useful technology, and new algorithms  proportional to that of the audience.  80  60  40  20  – 0  + 20  40  60
        presented at the conference are likely to be  Simply handing the microphone to a  Percentage of questions from women minus
        adopted rapidly. Powerful computers and  woman rather than a man when the floor  percentage of attendees who are women, % points
        large volumesofdata lie waitingfor exploi-  is opened for questions may make a  Source: Women’s visibility in academic
        tation. The world’s most valuable compa-  difference, however small, to one of  seminars: women ask fewer questions  *Compared to their
                                                                             than men, A. Carter, A. Croft,
        nies have grasped the power of machine  academia’s most intractable problems.  D. Lukas, G. Sandstrom  share of attendance
        learning, and they are unlikely to let go. 7
   74   75   76   77   78   79   80   81   82   83   84