Firstrelease:26October2017www.sciencemag.org(Pagenumbersnotfinalattimeoffirstrelease)1Theabilitytolearnandgeneralizefromafewexamplesisahallmarkofhumanintelligence(1).CAPTCHAs,imagesusedbywebsitestoblockautomatedinteractions,areexamplesofproblemsthatareeasyforhumansbutdifficultforcomput-ers.CAPTCHAsarehardforalgorithmsbecausetheyaddclutterandcrowdletterstogethertocreateachicken-and-eggproblemforcharacterclassifiers—theclassifiersworkwellforcharactersthathavebeensegmentedout,butsegmentingtheindividualcharactersrequiresanunderstandingofthecharacters,eachofwhichmightberenderedinacombinato-rialnumberofways(2–5).Arecentdeep-learningapproachforparsingonespecificCAPTCHAstylerequiredmillionsoflabeledexamplesfromit(6),andearlierapproachesmostlyreliedonhand-craftedstyle-specificheuristicstosegmentoutthecharacter(3,7);whereashumanscansolvenewstyleswithoutexplicittraining(Fig.1A).Thewidevarietyofwaysinwhichletterformscouldberenderedandstillbeunder-stoodbypeopleisillustratedinFig.1.Buildingmodelsthatgeneralizewellbeyondtheirtrain-ingdistributionisanimportantsteptowardtheflexibilityDouglasHofstadterenvisionedwhenhesaidthat“foranyprogramtohandleletterformswiththeflexibilitythathumanbeingsdo,itwouldhavetopossessfull-scaleartificialintelli-gence”(8).Manyresearchershaveconjecturedthatthiscouldbeachievedbyincorporatingtheinductivebiasesofthevis-ualcortex(9–12),utilizingthewealthofdatageneratedbyneuroscienceandcognitivescienceresearch.Inthema妹妹a-lianbrain,feedbackconnectionsinthevisualcortexplayrolesinfigure-ground-segmentation,andinobject-basedtop-downattentionthatisolatesthecontoursofanobjectevenwhenpartiallytransparentobjectsoccupythesamespatiallocations(13–16).Lateralconnectionsinthevisualco
2023/2/15 22:41:07
14.88MB
FCN网络
1