Artificialintelligence,

MachineLearning

and

DeepLearning

aretermsyoumighthearoften,butcanyoureallytellthedifferencebetweenthethree?Let’sfindout.

ArtificialIntelligence

Abitofhistory

Theterm

ArtificialIntelligence

firstappearedin1956duringa

Dartmouthconference

tointroducecomputermethodsthatwouldbeabletodemonstratereasonandcreativityinsolvingtaskswithgreaterefficiencyandproductivitythanhumans.

Evolutionoftheterm

+442071835820

info@magora.co.uk

sales@magora.co.uk

Whenwe’retalkingabouttheAIoftoday,weshouldn’tinterpret“intelligence”inthesamewayas“intellect”.

Creatinghuman-likemachinesisafairlyinterestingconceptfromascientificpointofviewbutisn’twhatindustriesdemand.

Wedon’tneedemotionalrobotslikeinthefilm“BicentennialMan”.Whatwedoneedistoprovidelightning-fastcustomersupport,analysefinancialtrendswithadvancedaccuracyandincreasesafetybycheckinginvisitorsusingasystemthatcannotbefooledorbribed.Andthiscanbeachievedbyapplyingadvanced

mathematicalalgorithms

.

So,AIisascientificfieldthatistryingtomodelthemostsignificantintellectualfunctionsofthehumanbrain:

naturallanguageprocessing

,autonomouslearningandcreativity.

However,withinthescopeofthisterm,wecanalsoreferto

ITareaofexpertise.Thegoalistocreateintelligentsystemsthatcanmakereasonabledecisionsandtakeindependentactionsinordertosolvetasks,thusliberatingstafffromroutinejobs,optimisingbusinessprocessesandsoon;

itcanbealsounderstoodasthegeneralabilityofanartificiallymodifiedsystemtointerprettheenvironmentordatainput,learnfromitandusethisknowledgetoachievecertaingoals.

AIspecialistsaremainlygoingintwodirections:

solvingproblemsconnectedwiththedevelopmentandimplementationof

AIsystems

inordertobringthemfurtherinlinewithhumancapabilities;

creatingsoftwarethatconnectsallthelatestachievementsintoonesystemeffectiveatsatisfyingtheneedsofthemarket.

+442071835820

info@magora.co.uk

sales@magora.co.uk

InordertocreateanArtificialIntelligencesolution,weneedtoapplyoneorseveralofthefollowingmethods:

MachineReasoning–thisencompassestheprocessesofplanning,datarepresentation,searchingandoptimisationforAIsystems;

Robotics–thisisthefieldofsciencethatconcernsbuilding,developingandcontrollingrobots,includinghardwareissues(sensors,trackersanddrives)andintegrationofallthecomponentsintothecybersystems’architecture;

MachineLearningisthestudyofalgorithmsandcomputermodelsasusedbymachinesinordertoperformagiventask.SomeexamplesareClassicalLearning,NeuralnetworksandReinforcementLearning.

Allinall,artificialintelligenceincludesmachinelearningasoneofthemethodsofitspracticalimplementation.Withinmachinelearning,therearemanydifferentalgorithmssuchas

T-

distributedscholasticneighbour

embedding,

Leabra

and

Neuralnetworks(NN)

.Inturn,DeeplearningisjustoneoftheimplementationmethodsforNNalgorithms,alsoknownasdeepneurallearningordeepneuralnetwork.

AbitmoreaboutMachineLearningandDeepLearning

+442071835820

info@magora.co.uk

sales@magora.co.uk

YoucancallMachineLearningaclassoragroupofmethodsthathasthegoalofteachingacomputertosolveataskduringtheprocessofcrackingsimilartasksandfindingpatterns.Therearedifferentwaystoclassifythesemethods.

Thisisthesystemwehavechosen:

supervised,whereahumanguidesthecomputerandcorrectsitsmistakes;unsupervised,wherethemachinelearnstofindpatternsbyitself;

reinforcement-throughasystemoftreatsandpunishmentsthecomputerlearnstotaketheoptimumactionsinacertainenvironment.

Nowlet’shaveamoredetailedlookathowexactlytheprocessofMachinelearninghappens.

Howdoesthecomputerlearn?

DataScience

+442071835820

info@magora.co.uk

sales@magora.co.uk

DataScienceliesattheheartofAItechnology.WhatdodatascientistsdoandhowisitconnectedwithMachineLearning?

Forthecomputertolearnitisnecessarytohavethesethreecomponents:

Adataset–acollectionofvaluesthatrelatetoaparticulararea.Forinstance,aclassregisterisadatabaseofgradesofacertaingroupofstudentsinmanydifferentsubjects;

features–atraitthatrepresentsmeasurablepiecesofdatathatcanbeusedforanalysis.Followingourexample,itcantaketheformofcolumnssuchas“Name”,“Subject”or“Grade”;algorithm–computermethodsofsolvingacertaintask.Forexample,youcanwriteanalgorithmthatcalculatestheaveragescoreineachsubject.

Datascientists

arethepeoplewhocollect,filterandclassifydatainordertoprovidethecomputerwithclearmaterialbywhichtolearn.Errorsandlacunesindatabasesleadtoincorrectresults.So,withouttheworkofdatascientists,eventhemostsophisticatedAIalgorithmsareuseless.

Computerlearning

+442071835820

info@magora.co.uk

sales@magora.co.uk

TomakeMLworkyouneedahugecollectionofdata–thiscancompriseimages,videos,textorevensituations.Youwanttoteachthecomputertoperformacertainaction–forexample,findphotosthatcontainkitties–andputthemintoaspecialfolder.

Foreachimagethatyoushowthecomputerinthiscase,oneresponsewouldbegiven–it’seitherakittyornotakitty.Thisdependencybetweentheobject(theimage)andresponse(kittyornotkitty)iscalledatrainingset.

+442071835820

info@magora.co.uk

sales@magora.co.uk

IfyouchoosetoworkwithDeepLearning,yousimplydownload100thousandimagesofkittiestotheprocessorandwaituntilitfindsthepatterns–fourlegs,twoears,atailandsoon.Themachineneedstoretrievethehiddenpatternsinordertobuildanalgorithmthatisabletoprovideaclassificationpreciseenoughtoapplytoeverypossibleinputobject.

Aninductionmethodlike

ReinforcementLearning

impliesthatyouallowthecomputertolearnbyitselfthroughtrialanderror.Thecomputergetsarewardeverytimeitdoessomethingright.Forexample,inthecaseofadriverlesscar,nothittingthepassengerwillearnit+500points.Ifitmakesmistakesthehumanwilldeductthepoints–verysimilartothewayinwhichchildrenlearn.Inclassicalmachinelearning,youcaneithersitandhighlightthetraitstypicalforcatsyourself,oryoucanuseunsupervisedmethodslikeclassificationandclustering.Inordertoestimatetheprecisionoftheresponsesyouget,youneedtoinventfunctionalqualitycriteria.

Inreallife,thetaskscanbeverydifferent.Forexample,thedataconcerningtheobjectscanbeincomplete,imprecise,non-quantitativeandheterogeneous.Variousmethodscopewithcertaintasksbetterthanwithothers,whichiswhythereareso

manydifferentmethods

.

Asfortheresults,machinessometimesdoachieveimpressiveresultsin

diagnosisand

businessintelligence

,thoughthey’restillveryfarfrombeingabletolearnwithouthumanhelp.

Moredetailsaboutdeeplearningareavailableviathis

link.

Popularmachinelearningalgorithms

+442071835820

info@magora.co.uk

sales@magora.co.uk

WehavealreadytalkedaboutDeepLearningandReinforcementLearning,butthereareotherpopularalgorithmsthatweuseeveryday.Forexample:

NaiveBayesclassifier

–usedforspamfiltration,frauddetectionandsentimentanalysis.

Regression–oftenappliedtoforecaststockfluctuationsandmedicaldiagnosis.

Clustering–usedtoanalyseandlabeldataformarketsegmentationandconsumerbehaviour.

Generalisation–recommendationsystems,riskmanagement.

NeuralNetworks–betterthananyothersystemforfacerecognition,butcopeswellwithpracticallyanytask.

Todayit’sbelievedthattrainingcomputerstothinklikehumansismorelikelytobeachievedthroughtheuseofneuralnetworks.