SYLLABUS [Academic Session: 2016-17] MACHINE LEARNING (TCS-402) Instructions to Peper Setters: 1 Quan Na Istould bcompton ner thentn plu. This qt should ate bet or shor ane ype quaton sould bef 5marke. 2 Apart fam Quin NI rote te paper shall oat orate orth ls [Bon an stould haved qustont Hower edn may eased engl] ueton om each unit Each quetin sa bea 2. Sore jective: To intros theses aout be koawledge f bs concep of machine Tearing etm, peseflerang Introduction: Basic concepts Definikon oflering tens Goals and appiatonsoachine orig. ‘tf rps rig ote, pean tos ‘Types of Learning: Supervised learsiog and unupervized (Steieatio: nt, aii tes taldaton dtc, one tng probate cdma end ener) geaet peighber [Fy T2}No, of He 13} ‘UNFFAE Logistic rgession, Prepon, Expnenil family, Canaratvelaring alert, (Gnusian csrnnant aaa Nave Baye, Suppers vector machines: Opal Me? ening, Overview of (ha TaNe. of ese) Uneupersed learning: Clustering Krmean. EM Alp, Mistr of Gavan Factor tips PCA Principal components analy CA Unepeent components anal), Intent semanti inning Spee dareiag, Marker models Hen Marky dsl Gnas. (74, No of Mee 1 Reinforcement Learning and Control MPs elinan equations, Yate tration and poly feraton, Lier qendatirpslaon LQG earings fon ruin eysnchRentrce POMP GP, THM Hs 11 MODEL PAPER-I END TERM EXAMINATION [MAY-2016] EIGHTH SEMESTER (B.TECH) MACHINE LEARNING [ETCS-402] Rae @.NO 1 isconpuen. Atonpanyfurfrometes @.L() Erlanusing anexamplethe conceptof learning sytem? tern eara om egrience wi rps st eas aaa Es praromnee stain eased 7h ior ipeece “uxt png crs 1 etrane entre pester won ais oops [Ahundoritingecopston learning preblent ‘uo gsning und daring handritn words within imaet 1 peturmans menor spect of werd coecty canied “ipatiperpenc «dtc fhandritan wed with velit [Arebt diving learning problam: ‘Tne rvngon uc erlaneihway sing ison seers 1 Retrmanormennre average tes eeeedbfre anergy a7 o we renga deve, ‘Q.6) What arethegoalsof machine learning? o ‘Ans The pinay golf mane ering ito eel geeral purpose algerie otgrata vee Suc elertan shld be efiest. earing rita shel lo thera! purpne as pune We re oking or alontae that eam Be aly Shlled ia bend hu of arin eS ‘we wat the result oflearingto bea prediction rule that he pedctions hatter Ocsioally, we may aso be Te tide by imps impressions or perhaps an examination a nly a flatly Shall mamber example.eae ight Semester, Machine Learning Q1.(6) Whatare Generativelearsing algorithms? o ‘Ans Consier acasifatin problem inwhich we want learn to tngish Sewean elephants =D and dogs 0), basen tne fete tan ana ‘Givens training vet an sigorita ike lngtie regret othe poet sigrth bse tes tofnd estate line-in einer that separates the phased dag Then, edn new animal sie an ‘lpn rg on hh to a ein dy fal a sp ent rat te ea SUSIE aeuerioeten meneame err tere a eer ro ‘Algorithms that tolear piety (uch a lepton alga hatte to learn mappings directly fm the pace fnpateX tothe abel eh [asthe porepton algrith) are called ernie easing algerithma ‘The algoriths that instead try to mode lx) and pare elle generative learning algorthne. Fer inatane y odcateswbeher an eampes = gern ‘lophane (hy then ply =O mode thedstrbaton af age fare, abd platy =D ‘oleate dtnbuton of epbants features ‘The diferent geeratie leering algorithms are + Gaussian dscriminant anaes Nave Bayes 1. (d) What dosouunderstand by Reinforcement Learning?) ‘Ans, Reinforcement Learning refers tothe problem ofa goal-directed agent Interacting with an uncertain environment The goa efan RL agent it maximizes Itnpstn tala rewerd by saung the state ofthe oavironment and taking stone inblh ale he sate. etch tay an RLeytem get evaluative wdc sbout te [Perbrnanee ote ation allowing to improve perfrmanco ef susaqaet actions. ‘Bojer sethods hav ees developed and suceeflly applied in machine faring to loam opal polis for fitestateinto-eton greene Markov Decision Process MDP! LP Unlverny-B.Teeh-AB Publisher wera ‘An analogous cota ytem shown Figure where the cone, awed co fifteen eback sbttprevow ection calulin te aoa peciead to mnimprv prorance The renfrcement ial it eset eh mans evaluator on, whch iia auntie sae ae enya ha sinulecjestive to on optimal conto ahh ra ie te ter performance entaron while maintaining atablty a cia the vlue fn, which ea eeasure of goodness of = ‘Brenachon fr apfven ste 'Qi.(c) What novefitting? ‘how, Oveciting refers to 2 mods that madels the traning d ovectitop happens when a model ears the detail and nee i the a eA t negatvey impacts the performance on them FaeSE ent tne nie or pando Mactntions i he training datas picked op saree sncpts bythe mel he problem i that these concepts do nat aang creatively topos! the models bit to generalize, Overiting SPREE th nuparawetis aod onlnentmedels that have more Desbity Fae nese fonction. Ay tuch, many nonparametse rachioe learning esa parametar or teehsiqes to lit and constrain bow mech geri ~ ‘etal tbe mo "mle, econ tees ae nonparametric machine Tearing ryote and in subject to overiting training data ‘Thin pret orto Tore some she dtl as ped oP “Q2 Explain modes flearingin deta ap ‘hee Leasingutrtine arena vgn rounds partir paroien’ eee ie reall aprac leanne Acomplia earsng eta soubor sou te towing apes ene: rat = igh ing on lent ater propre entire eaeded more gene torre eo Teolttgtotemntarentsor maybe noid in pyc obese oot anda Tcucworeef cern naligont enon * pomain: What eng learned ea anton, oa eancet. Armen the many anor pois re operon ta Jv fe ogee lanpnge a peer, smytbe tte cane eferosepn ta ofcncep hat er consierd far tearing we Sistema cae. ~ Gon iy thelersings dna, Thelerngsanbe done ta eee setof aes rom pets dta obewnepodeimalter or tmephyeal phenomenon. 0 tae font rer ten, and ‘Tepresntation The my the objets tbe eared are represented g the way heparin bererseted ye mpats ogra Ton ypabers which he poem arp whtelenming maybe eprennted ite sameway ein broader or mare ‘sited ora Algorithme technology The sgn emawo tobe ued. Amore the acai areal psa rtwork, cane tase [enwaing, dec tee, pamiary, Ins state machines, probate network, ‘elearangspport ector machines, and these faci One may ls spect)em Bight Semester, Machine Lersing eanigpetenetey eter nh tn ‘oats rg ee al Sacto! cman et ei it a rpm sari cre Sot cere inna meson en Szenlee Ses aerate Sek oocanamwemereoe fag omni nrpon s ‘Bet ome lean dss Ina nine aig nar he rogram Sa en ereacams pecan Eee eet deteerermee eet itd Zancai Se See Sipanatenaee eter crea ome ichanattginetesintsaeomemarnanrates inuitraeninenanetone cepeenoncs reggae ethane ee “herctatas rnet epee im nie pg i een rent se mreltis Trina nO esti eens eat pee Se ea ania eet Sonics Shortridge sets ent toca es iiinaaipane ieee remcirseete ie dins emey notioeeengenco noes i tpaemamnetherintr dt “rams etre fin an gn i eee sisi saben veer’ npg orate emia rate ecinaoe teas isneterted repeat ier ae Stet iymawittpse oe “eons Tio eaten pt ee anata Sos ae ace sieberiana Seopa meen ateadopoeeloon ‘Conpatationa ering models muy pend on many moreereria and specie tain ofthe earring rece. wees 1 Univer. TechAB Poor cco algorithm In deta ‘han The tnt sgt fo sive re lenring, rvesznds approximate eat ln tier dco es omc ies td. eee eso ahaa abo etnad atthe ot ofthe re? The atin atthe rote tthe ten Ae 3. Eel rade ot te eae tae srt, ad the eng easy ered tothe appoprste devcendant node (oe, down the bran a es vein thle) Th nti procera then een a eater tncined with ech decondant tds selec the tsk cae a tat pn in toe ee Ti rms weedy search ran ecrtable RERENECIESREN Wt tect never tka to resmnier ener ces v potspy, that ebaracterzes the pany fanaa colecton of xaos. nnn nang isn ag rata re a Entropy S) * ~peleesP9-Pees Fe etre the propaton of tv examples in ani she proportion ugetve xpi in ‘Taformation gin eras howl agivn atribte separates te tining exact acnedng tote target asieatan IDS woes ths infant pin meas CETTE Stag ie eodidateatrtotn each step wile growing the tree Bae Some in stnpy te enpested reduction in entopy ese by parttining thecasplr coring ote tbe, Mere peel he infra xn Gan A) lam aurbte relative to colcion of examples Ss eine a2 there rede then rested rec aint. = Eato)~ here Vet A) ett flyout vals fer tebe A, and Sy the nharat' far whieh wre Abt aleve, Sy= le SACS) wD Te vale ‘Gait ite mumer of it eaved when ebcfing te target vale ofan arbitrary ‘nenter by knowngtbe val atribte A. no Yeromen ighth Semester, Machine Leasing ‘To ilatat, repose Si elation of M examplon a some Relea concept, 1 University CBTADAB PB ‘eladng9pstveand negative examples we ade th tation 9,5] ha eamplecf data). Then the slepy orl o ha bcleacaeention Batropy (0 +,5-) = (10 eg sOA) deg) = 090 Sera Sacco angered winning | ‘ind, whieh an have the values Weak or Strong. Ofthere 4 eamples,vppon ot {he patve and 7 oftheneetve examples have Wind» Weak, andthe remainder ave ‘Wind » Strong. Th nraaton gain dc to sorting the ergnal 24 eazpln by he stile Wind may hen be cleat a ‘Vaves(Wind)= Weak tong 8 = (04.59 Baa 166.24 Same © 84.341 Stearns) sin, Wind) = strony) — iat + Eneey)-(SLOEDYSyuy) (OLENA ae) = 0900-t@nans11-en0100 Information gin ie precizl th marred by Do ale the bet atrbat chatapin roving tee Th we enoraation gino valuta he rlevancet Phles In tien gurethaformatonunel ws iret tris, Humily td Wind incmpied in ererto tarsi wish tetera for lasing the taining eames ‘Which tribute the Bet clair? Eos ee ay \ pom vans \oors Both Bat BB ENS (an. iy) ans in Seiogmeesiesse Tegra aeenens 4 (0) Consider the following set raining examples: es Testance [ Ghifcaton | Ani [ames [Ames] 1 et * 7 2 a . t 3 a » Fr “ a . t 5 3 . F ‘ e > t ® 1 e . F © ® a > T « (o) What is he entropy ofthis olectionoftraining examples with respect tothe target fenction casifiation? o (What the information gun ofAUHD! elatve to these training examples? (c) Crete the decison tree for thee traning examples using TDS. (Convert the decision tee into the rules ‘Ana (a) Entropy» -81016510)- 2100p 2/10)- 010g 110)» 148 (8) EAU wa) = si4ogi34)— VAlogtV4h= #08 BiAWwht =)=-Dlog24)- LAlogt)~ VHlgt4) = 18 EAA se) -12ag{.2)- log) «1 (e)GainctAsibl «E~ UIOEAurbt=4)- 410 Bebb) -210ELAtrb 0) 036 bd (tase =) 4 (Ath? Then Clason vet dusts wa (rbd =F) Then Cassifeation «ch Meats «by Atebt =) The Cassieton et caer =) Ate ob) (AtEA? «7 Then Casifetion «edeur ight Sometr, Machin Larne Hate «Ata = 9 Ate = F)TRenClsieton wet ted +o Then Chasiestion heed =A) Then Casifeation xe (QA. (Give decision trees to represen the folowing boolean fenetlons (@AviBACL (iaaBvicaD ‘Ane (@)AVIB AC] oy tds BiviewD 5. What the dtference between supervised and unsupervised learning? 028) ‘Ans, Sopervised Machine Learning: ‘The majority f practi machine lnring wie sopervae earning Supervised teacrng ofl conenan inlet problene because he goa is Ren ge tho amputer to learn m clasalfiaton aystm that we hav created. More generals ehncaicn eanioginepropite fran prclem where daducnga cassifestin 1 Unverng-B.Toeh-AB Publber MP9 etl andthe claniiatin is cay to dare 1 eelled upervined learning th prem oa aot onrlng rm the taining datant ean be oven ASP EE2o'Sipervining toe Tnening procne Tho algorithm eratively makes aon ee eining ata and io correstad ky the eather. Learaing ops when ‘Esalptn achiever an ceptable evel of performance ‘Soperin earnings where you hve apo arable) and an output arable ‘op anda an gor oar the mapping fenton fom te apt toot v= 60) ‘Theat istoappoxinate the maringurtion rowel that when youhave new sngat ta ha can predit the otpatvarales fr that ata. Supervised leering probleme can be further grouped into regression and ouiteton problems = Clasfctin: Aclasifton problem when he output vaable sacar, such and Ml oie” ad "no das’ ‘*egrssion:A regression problem ia when the outpt variable rel val, sucha lar” eight” Soe poplar examples of supervised machine leraing ‘Suet ar * Leite Regression + Support vector mache (SVBD ‘se Nearest Neighbours Naive Bayes + Svat for rgrston ‘Unsupervised learning? ‘The elit han te computer arabe todsomthing tat we dan telithow tou! Taare scaly tow approach toupensne leaning Th ie aprons {Stench the agent not by eng ean eterna, bot by using sone sot of ‘arden to indie acne Assn Ope a wspervaed ening is ealad ‘Sontering Int pe oleering he goat oto mane way fn bat Simpy tnd sian inthe aig data Toe sumption fe tha too ‘Sater sovered wil match reaznay wll wih untae dsiaton, asupried arigis whee ouonlhaelnpu ita and no een utp variate, Th gal er unspent sled undrying aucun ristribution nthe data inorder learn more bout the dat, ‘These called uupevined leringecue nt supervised — shove there eect ante and ther fencer: Ago ele to eros Abvise to sce and prevent th lteretng strate athe fa nspervisn leaning proloms ean be farther grouped int cst sociation problems, baie aesighth Semone Macioe Learning Acustering problem where you want dicover the inherent ose an grouping costomers by purchasing behav. “An stcation reaming probes ie whee you want to diner ‘les that dese large pros of oar data, rh an popl hat by Kalo edt tart ‘Aillutering bgt come under uneopersee earning sgpritms + K-meane clustering + Hirai cartering + Hidden Markov models 46. Whatisthediference between Model selection and Feature section? 023) -Ans(a) Feature selecion:Ia machine leasing fstre seta, le known sriabe election» attbute selection or variable subset selector, the proven of Sclcting a subset of relevant featarer (arable, predictors) for ute in model Cnctnocton. Feature elation achniges ere used for thre ree = Sinplieaton of mls to make the ease to interpret researcheainers, + Shorting tine, + EBahanced gnerlztcn by redoing overttng ormally. reduction ef aries) Feature alec Lechniqos shuld be distin rm ature extraction, Fetare ‘acion creates we fetes fro fini ofthe oil eatures, whereas fate Btoctan return eubea th features, Feature selection terriques ae on used {domaine where there are many fencres and comparatively few sample (or data ‘Bint Archetypa cows forthe alin feature selection include the analy of ‘riten cna and DNA msrarry dat, where hee are many thowsand of festares, (nda few ton to hundred of amples. “Afeatareeclecton lg canbe soen at the combination ofa earch techie tec popcng nem feature subsets, long nth an evaluation meatre which sears the ‘Aerene feature subeta. The sspletaleitim Io to tet each pousible sub of {eatures Gad the oe which minimise the ere rts. Thisis an exhaustivorerchof {he apace, and computationally intractable oral but the alls of fostare st Feature Selection Algorithns "Thee ae three general clases of feature selection alorthme: filter mods, wrapper methods and embedded ethods 1. Filter Methods ‘er fsture election methods apply a statisti! mesure to aesign searing ech entre: The feature are ranked bythe sar and ether eslored to be her se riom the dataset. The method are often univariate and conser ce feature with regard tothe dopendent variable. Some examples foe ter independently, aperasain the Ch oqeared tort iformation gun ad creation eoelclent 2 VeapperMuthods arch pete, lection of at of fatres a Wrapper methods conser the ‘prepared, evaluated end compared to other where different combinations a 1, Univerty 48 Teeh)-AB Publisher mesa td to evaluate a combination of features and tng scr lted on model accuracy The ware process maybe methodical nich ag sn pe Tends any mocha noch ata raptor hll-imbing algorithm, it NER CERSSIie Ite toeerd and bacoaed pases to add and rove featares. A LIN ef a wreper method the rsurive fate elimination algorithm 2. Embedded Methods ‘bedded method learn whic fstares best contribute tthe acarac ofthe sndl wile the mol i beng eased. The mat comsnen type of embadéed feature ‘Riedidsmathctsarererulanation methods, Regaarzationmethodn are also called {enaliantlon mothe tht itrodceadiionalenstzants nt the optimization of « Frodaivecigorithm Gch an eosin algorithn) Ua bia the model toward lower ‘ompleny donee ences). Examples ofregslavzation lgrithms are the LASSO, ‘tane Net and Ridge Rereen. embintlons. A predictive model aah tt tt at echhirtice aes ners sprig matic gacieara nme Soest Somes See ae each a psec agate oentitr ‘often considered. * er eee eee ates eae sna cetaceans fanction that generated the points, waa = ppemee ste Salita pease eer tat atin cpt eee nemesea ttarmfegyareiageaea te aay Sta aeine wa tcemmam ete 25)weet ighth Somovtey, Machine Leasing + Aisa setofocton.(Porexampe the et ofa prsbe directions in which yoy canpushthe elioper entre tks) +P, roth tat transonpeabte, For eachetats eS and action 8, Py _adstnbasen oe the stat spre, Well sty more abut hs ltr, bat biel, ses ‘he dtbtion over what states we wil analon ef wo take seon nin sas’ +16 10,2 aed the dicount tn + R:SxA-+ Ris the reward farton Revans are sometins ala writen fancionots tate ony in wbichese we wouldhave RS») ‘The oamies ofan MDP proces as flrs: Wester insomedtate ny and getto one sme ston ye Ats tak inthe MDP. Aa esl for chet sat the [MDP randomly trasiton to ome cua sate dan ard te - Pa. ‘Then, et te ick nater ation y Anarene thi ato, on ote tons again now tosome Pa, We npc, andeoon. Patil moc prt py Rey Up visting thenogoensaf ater gy, mihasine a a. oer aa Pe sapivendy Baga) seg) Rg) + ‘On when wear wing ewards aba funcinafthe stately thie becomes +e) PR ‘Ans (8) Bellnans Equation ‘ABelimun equaion{s also known asa dynamic programming equation Iie necessary covditonoptiality associated with the mathematical optimisation ‘method known st dynamic programming wie the vale devon peter at ‘certain putin ine in tars of the pao rm ome ial hac andthe ae ‘tthe emining deen problem that rel fom then alee, Thit beaks 1 dynam optimization problem ina staple uiproblems, ar Hallman’ "Principle sr Optialig” prescribes ‘Alsat any problem which cen beled using ptnal contre tniry can alee snlvod by analyzing the appropriate Blinaneqaton. However, the ean ‘elina intr aul refers tobe dynamic programing equation avnated with dirt {ine optimization problems ‘hi the Bellman equation, Ian bo spied even Srter i we op ine sabecrips and plugin th alee the next tate vio = pay (Peadepvereed) ‘he Blinn equation iclasied a factional equation, because by selving ltmean nding th unkown fnton V which the value neti Recall hat the ‘ale function deertos the best posable value ofthe ebjectiv, asa funeton ofthe latex By calculating the valve fnction, we will lo fad the fnction a) that 1 University orin vector formas vos> 0 Forbes 9 ill sometimes write the pete Ln very 87 AB Pah vests a) = note ve 1990 vote 8 cnet pp psingspernedcion rfa the "oints).The perceptroneutpatsa 1 fr instances sa ia ae ‘png sada! the ypeplane anton din machine learning? ® tow Dudhovrting se ndertingean lad poor mel perforance. oS by turthe tent amen ion napled atin earengvorriting Overiting {chao enn vation ache ering eget taining aa {Stren fm ese we aay creo ot aaa aly how ele Sign pera onsen athe we tapartan techies tat ocak ‘Sten ete mache eaming signs oes vein 1 Uses ramping tehige a etinate mel asuracy 2 ld back alain cate ‘oem poplar resampling eget aidan, tallow 3009 train andes mel inet tent ars afro sta sed ed oe ‘tinateofte perarmans aca ning ne oe ean da ‘Avalain date any a eto ou ning dts tat you ol back tyr aching sll ry endeear ret A hace sted entree ache rig lbs on or niga oe tli earned mols he vlan dataset os ak thee Saat othe ee ight room morn ata Using ces ants ee na lntupled mation rang fer estinntng eal ataregyonancee eee ‘Meds, naga validaton xtaetis aan ecoiot ee (0) What doyou understandhy POMP? Aas. A Partally Observable Markov Decisi ‘sweralztion of Marko deanseat Eighth Samet, Machine Learning ‘alatenance and planning under uncertainty in general. An exact slton toa FOND? {elds the epimal action foreach poste belt oer the world states, The epi ‘eign mantles (or minimise) the expected renard (or cost) ofthe agent over sil infinite hoon. The sequence of epinel actions i known eth epinal oly ofthe agent for interacting With seinen ‘A disrete-time POMDP macels the relationship between an agent and its ‘ovroament Formally, POMDP is TuplS, AR, 03) whore +8 inns tater, + Aisasetfacions + Tis set of enitonal transition probabilities Been tates 1+ RSKA-+Ristheronard fenton + ies et observations +0 ie et conditional abereton probable, and 1 re (041 the discount far 2 Give thediferent appectsofdesigning alearningsystem. 126) ‘Ans, Ta order to strate some ofthe baie design itues and approaches to ‘machine learning, le ut consider designing program to leva ply checker, ith {heal festering inthe word cskeretoormaent 1. Chooring the Training Experience "The fie desen ese we fetches the type of tralaing experienc from vrhich our system wil era The typeof tuning experience avaabe ean have @ ‘Sonica impact on sucess or ult ofthe learner One key attebtels whether te {Gaining exyenanc provides rect or indrec feedback rearing the cies ede by the performance rstem For example, neering play checers, teeter might learn fom tes trsning samples eonsitngot edu checker brad ate and the erect over each ‘Asecond important ttt ofthe training experience the greta which the tearm contr the sequeonfrinng example For example, the learner might roy the ncn tees nora ar ates rie th cect me ich ternative learoer might tc propos beard ete that fd pa Seeking and wk the teacher forte coret move Or the rarer may have complete ‘Solel ver bath the boned state and nrc) training eawletons, a does ‘Shen t Tears by playing against itself with no teacher present "Athind inportantattribte of th taining experience is how well t epresets the dioafoation of exaples over which the nal ystom performance P must be measured. In general, learning in most reliable when the ang examples flow 8 (Giurbutin ear to that offutre test examples ‘checkers learning problem: ‘Task T playing ebechere + Perfonwance measure P: percent ofgumes wen aginst pponents «Training experience E: playing practic game gaint tell trite AB Pale ons easter ac aoe scene aly to ea wi ll a io iin nb pone ra at ahh Sige ra amen cme soe Foren re ane rch ee moe ere rein mates oe ee iron ret anon gis nt nt ee a i hon stars eer ah hotie aerate aaa ee ty ioat tenet HEN aa eI toes ae dnd nt steep rp tin Veg a hat iterate scaeerae te erates dad ent at en = alpaca eae OI— Feats dinncate teaming peor tt fal eee atten cee ohn cogil om nnreenase rae er eee tn ee ere rey Sones ences eee eee ala paar anes Renae tcltyestetnneewteecsicceet ee ee Nae an aoe ieetaes cecealcenep ec eerareee nese tot ar ha ped nl ut ae We mos chow Personas ta ecasgpeessou te ements tenance Sine ineeracausecetecana meetin aati Mtwrvintonaamnemmntoennane terse ae once “retary niente eer ce oer apranin areca en eccrine ga pena nn roreuanchrapueetaliinas any eee Sees = tp terete yon and GalniS, 0)saat ght Semester, Mucine Leasing 4.0 Give decision resto represent the following boclen func wane (AXORD An ARB ; ‘Al metas A, and A? meane 4. Explain in detall Support vector Machines? How Io the optima hnypetnlanecompated? 038) "Ans Sopport Velo Machines In china ering. uppert vector machines (svitjaresuperied earning tsi ected aringsgoitims tat arly (ie ned avseation ad roprennn sly Civen fining expen, cried ectoging ton eriinaterefeocugotas an SY sleet ‘ita madd tat asp new example tone cater thn tie inary near laser An SVM model isa representation ‘rpintsinepce tapped tht the examples ofthe eparaecatagre ae vided {ear guy hat ao wide at poste New examples ne ten mapped ins that ame ‘hme ond reid telong oa ctepry based on ch sie of he gap he al Talons perorming linear classieation, SVM can ffeny perform anon User ciation tong what called th Lee ek, inlet mapping their {spats ine high-imensacal eaare space 1D Uaivernity-(Bec-AB Publier esto _Asopout Vea Machine SVIO ea diniintive lanier fray defined by aerating iyperpane Ine words, gen bled ein data peri orig) ‘hetlpyrn atpts an opin pepe which entegrze em examples. a ole the fling single problem: For a linearly parable st of 2 es finda nepaatng aight ie y wee as 7 ° “kh \oe oo a tine bore pstace you cm ee tht there ests multiple ines that offer 2 setsnn oth poem any of hen erent thers Ween tle die ‘Tonto aetna tho woth ofthe Hoes:Alie sb it pases to case to te {ins bcos wile moe soni and wil at generalise corrects Therefore, rel sh be ond the ine pasing a fora posile rom al pols. ‘Ton the operon of be SYM alrite asedon Sng the hyperplane hat si ett mim dita the rig enlace How isthe optimal hyperplane computed? Atrio ots yas tof typenannins igh cite incscnalogny sidan wed sana ecro-mpat ighh Semater, Machine earning {ash Itstvly god separation le nehevedby the hyperplane tha has ha aga,