Steps twelve oy Pon Tovial 1 —. | Stp i Genter the datasd Step 2» Calealale the woomriance mabiia af the cenbued dala Skp 3 Conyrote the eigenvectors ain ergerwalns Ste he Sort the eigenvaturs in descending order Ard chee bp k — tigenwectors wvusponrding to the highust ciquwalucs . Step 5 Transform the original eet multiply sikoilty the Adectad etgenvesborsl Pinapal comapronents) Fm oltain a Lowe-dimunsional re, tabiont ‘ ae | Danes with Pep | Eigen face example jou Face revognation Of employers , tn 4, Company les ashume | 32000 ermploryce ee fortyvany - oe mio * De we need tw thre alli S2e00 employee’ Jace 9 Oy tore top fue projections /PCs { eigen dtections) Dam) conunns with £CA Time wompleity how muh re docs it fake te oct & re PCA? firme intenare FARE) yy gay ay - Lgfincling erqervaluis and “eigenvectors. Normal muthod of finding eigen valurs fake i [eca) | typically ep] Ps dota / er Ti vn de loge, a, Tak do Pen dinday [URtar eorrtinabon of Plo for « delved 4 i rans folk dake points having hon inca velabonahip' bub eh cant find it Fg: Data rraynot teessarily dee iy bos —Acmennional Linear Sporcereads Same colr to both issu | Saprinng honclinear dabusd and plane) (anal Fare _Omyplorely ivcue with Pea) * Le blame ad (Lu dean) “he of Tosuc large C {© datepor nts feabures xx: Tj 2 eet x) xy . } } JL Say >Ee+ xx] uk Let %, be te eagenvector wrvieaqeonding ee Ham langue sigurualue 0b. Cs) Wesginwilne / Cts _ pet bi ingtion z CWye Aiebie Chg deere 2) G5 Wes ArteHow get Ae? We Xd ek | Cue Ak Ok ¥7) [xe4e) = Ay XE) bert) xp) es Ae (Xt) Prone x dy | xT Ga xd) 2 nde xTOeae) PERE (xTx). OT) Xe = Und OT) 4 xt on™ | bets calf | XUXERR, ani +! Rras n Ale Te we won find oy thak sabisies | - Rake (ARAL oe tie an spe (As we knot thot, El ape Mp MF od tr ay We “sy wot wae s OXLEY TORREY eel Ey EOD, 4, we kro \|log)) =H ee oa oly should ale cpbishy tHe “47 cE] x | | ah ice =i ah uh 8 i of xx rexT and xx | ine ne ard Wn | ore eapetly’ tel-sdene})) whure | ce xt Elgrwedtors = (erp sh i Bh 5, Eiquurdus= L@A> Ares J i @ xxTe ne > same ergerwectors i} > ergenvale =f md 2 rhe Y- yo ox eigenvectors 2 (Bi, 0° PY Mey R eigenvedanes, =f ndppnda nc? PMY Tees rAd pl b 2 e227" ) ge, we ten set reek = ne Deeg 2o\A dude Oo " WR = =i ; avn ot ering \ergen eset valu given x ER ey Dedmy-: Raog 2te en [A soi) wsstep SS Ganapait “pe xtx, ER [ » nen ime eStep 23 comput, ‘eigndeconegitn, of mR LS nc asa Py corres cere 1 “Any! fat) ) | MW fp + Hepsi Peo ke Le. 145) fe sang edu MAA yt - Kok, wk sajna Ny wSsp9. casedevin il hte ge neod nicl the aches beac ted. ees Spang tbe this ides can help tS 2.) “hon Ucar aig Oe ee de Lat nitiY - amdasy ach (eretlebs sedothFe ato vrans}rmalasin Tosue 2. Won hear relabionsbuj What would Pth gue? _ (a: E D Gntre te dakarck 5 wf]: 2) Then Cortina “the [Link] —y 1 poudure | Ci heb us avnune it gars £14, 424 | ard both axe imputank. Relation biw feakures (fi-a) & (Faby = SO fi Raha Pap h enece] J hon JL of? Ph HE) go t : bs” bt) igen | Let anel e n Vee We Bory any o waza ~25] fe S i esr 193. Clokack datenent sabi fyi (gds CB)” op ns TIT i of R bo A PAS the datapornta ie tn a [bes ee 2h R11) Tako; Transfyim fractures ‘from | Uns ditnsriion po %& high dmenren > Prot rh sy att piheanadh RP be) Ya phan) an 34) Be yy stay dB alt know how to handle che We already need & find GT trrlzad when Daan [how ot +e] 1 ar fog un awe four eabertss Teatre rlahons] Ch fe fy fol ———” | Ty fb Cth, a _ aly ifs 2 Jfealuver Va labed ut thre be 2 pl? power L Face 2 otal) tee Total A feahures, | Wa geen | | Kernel dunclions. whol Fue cat find $i? yteted R) x (inner k ‘ote A717 — ' pins to comapate Mb) PU) wundal firding #4) from 0G, 9G 227 5 en this a | Lets perder VED ote Ch FJ x= [9 ad tenrider re fundion nie txt ats (Lt “fs woe 2 UF 91. + fage es)” - tanbes 4 = (fio) Jilfega e+ ated. “AES ae us + 2194. e[r a& ! er eileen . ' = gly) Tp to) Sis92 Sify where WwW eC): | ae ' i W alpen ; aN i 4 angtal the, “dated a : tn “tranapruned apace” fou ee "Th Was, Ca geod pGWe aA ce > FRI Vig, ng] gren! To9 a [ar ae| ur | expbatly — compittng btw) d- dimensional ae me | for some pay lobe point ) Can we da it be “More examples)! | k fae) = boitatany? Ahowh to be a " valid” purchon — tan be foc ws valid . ready Jerernise : (eg p: nf — pe Auch thal ) compute ¢ ¢$ — \ k loa’) = bul gta) =e Vt wt ae } One more cxanyte , Find Aeon ss]. op) « exp | ~ ln) ") for @A0m ea ‘ ( ser J TP tgif map | : ' Radial basis furclion Alswn to be a valid _feaction map. tan be 1 , 7 | Ls tatirtestingly, gh in Chis tare maps “ot tle an | " infinate’ damenaionat ane an | [technicalities aride,“*can Unink of Uris | port to a ‘funchis’ and dat roduc, lays \ mapping? aka < beliveen pordion become entegrals ] Som this valid maps /tunctions have (tee eee a mao? Ya, that io ; 4f keRvet_cowet i dal [7 “1 ERNEL GonetTon, KERNEL FunieTton, Ang furction kk: pix pi— BR which is a Meld map A) called a, Kernel funcion: akbvas Eotxtar)? — polynomial! kernet =n 20" ) — radial bars) 2 gaussian Keel Gra Let kl ae) =, capMethod 2: Mercen’s thestem (in Given! a function Ke RX IRE ih 4 valid ar kernel > af reer Sis ay method a Palvbit « map > eoyircitly Umeght be hard aontetinis) k: Pawn —R a valid kernel iy rand only ifs | © K is “aymmetric te, ky!) = Kha) (Since dot producct f \a symnutric operation ) ® For ang datusek L2,,-5- =n}, tre madrix Ke where kij = las, oy) i9 positive define aul ie ae are) non-negative al ec enceurtanedba posliee} kernd PCA 1 Faulk La, Y WE Ae, kernet kg Rixe dor + Step ie Compute ERO where ky 2K low, %§) 45 4 + Slip 2s Compute Bice Be eigerwectors and nhyb nay se 20g) eigenvalues ok ip xl cod mre to ge oF] st Sepee A= Agint SORE” hy, = blk. _ becouse Le i it needy dh) Swe cadt “reconstruct” eiguweckors “ Ue covariance pnabrix- 1 hw Buk wef lll compute he “ems reprarbabion (ww?) ee b awreatrean Gat HD BBs) 7 et sn F fi gta $ bu) gos Ke| 1 | we tan reconstruck edgowetors eapleatly ote a | compute the dot products or He projections tes —cigenveclow oH highe dimensional Apacs. | , | The reasen why” we do Ph is, we get Ute | Projechions alang each of “these data, paints Lets say for eg, the original datn point! twas | teo dimchaien but We [only icare of bp B59 we rly map) ‘the ° te0-D cata, ola @ S-dimensional projection. .° ae ay ki,.. 3 Hi BAO Eas Creda inareniy > Wmenion sh darbrstn Ne y® taplire non. Petarity, | | trivin Ke RU y= lai oy) Hes \° Creake a. new. kernel Io oy Getic 1 Po | | Ky = ky - Or 4; A ot PPA; at ; . where, mahix ¢f with . ‘ye n Fo coy ny AN a Gp BE = ey, Ke 42, - — keys pom es | Lay, Mayes Any , La pvaye Ply LAS A dl) may ton pphd-ad MeL 3g la) || | | Fir long 2 \ , = ~ A ae , Dis) ear olag | ree (pla 0 i plas doy) “WT gly) ea glud) fat 8 we i {Week-3 eek ara lysans { Tntwwcluckion to ustoring Another paradigm of Unsupurised Learning jp Clusboring + 37% datapoind are | _ 2 clustered together ; Ve sats A voy RM fig 3 a] MER Lo . Ren [with proxy) Parktion the data ino k different Ge lm XH Xs ty Ay tr : k=3 a | [me X25] f - (ce a4) me) sy . the ‘paren? xwhot wb the si (eco datos © es) attr gi, netlare 4,2 partiHionny ' 2 = date pov axcodats By Ba, 2 Rn ee Cluster ted calor a cluxter indicator fae ar a Gucstion Oven * Gecaraidemant, hour good 7 Fle;, vs ba), 2 EMMA Maell bo aquarred (Performance measure) = ome wasj>. wee want parbtins to be hemogencovs » They show | toe abe. How apd. teers Gre) he differed bye there hares from, ream trbac of tee dale 9 fee L celedates the disteunce ‘bho bare. Sie by iaonge the um of aqnared vector vals + jon 2 bayer te Ab Javviage “Phyo esduborq did. . . a T- tadicabor jie Zoe T2124) iF ag belong Hales) Ben T= 2 T(e=) °r Teoty boty ey yl eds M, 2 1 oi | 3 \ / | oak i . | mn else ne fay my My pot bso ; . | “Too many poviiities | (K” ways) AVP- Hard _problein [[¢8: engineer ‘lera} Usgpper Weis) | Yun tine Ys polynomiat tn’ Paramulirs of Intowute > | \ syay | iyyans. Bee CS be means algorithm Sip 1 ae 2 | 2 Compute many 7 SAG Wtbn=tn, why yaa « oy 2 5 2FLk)' visthat wigs sy 5! Ss | A oh t undd asbant 9 5 tnt (oe Reawsignenent ep M3 20 Seer mn |} a “abit "Gif the ign 8h sprmeat avignnwenk Seats ES pa badge) we ne ten awsign ts fay cluster ushase mean io Closer to ox; ) | Ip thre ove a eee aren pevibes conbinged rate ot) datists soadny ast | Fad tyods algerie ‘hve es) ethos “ «at a tb imal” | ns me \o “ a |e Buk Prodi) Nreaenalt ‘Ww ts 8S" pin ebiees Vs “ (Az ys} de 3 ie | ext Geis) ¢