WR AssIGNMENT 4 Q1 Dsus 2, a deta abaut ‘ Autbable oli grams. ee ee Conveluttonal eural Netwank Comm) Mo wnvolusonal newrot Retwank 4 eee Reheark that tg Jernatly ere ee, Visual a Pena ees Ge Progen by Procem ng Wh Gurls Ltke i oe 34's eho known an Conv Net. A convelutonal : Whe eK ES oned to detect and clang reeks in on Crue hie i CNN hos uttple Wololen lagers that netp ty Pethoching taphrmotien from on ch The a | Pe eet fi a ea a ch Conveturion depen th Rev bayer CN) Poot CO Fatty Connected upon 1+ Convolirhton Loe, TiS Hs the first step A tHe procems 4 eetaceh ete veluoble joatunes from on fonnges A conve tuxion fayen has serenet filters that pirfonm te Convelichion opehation- Eveny Tmage $ corrs'clerot et a meth Pixel valuer. 2 | Uba ho lo pret. Rte, Os lectin ae hl ela. ofelali fe © | onal Te ofits [ofo] 2+ ReLV ReLU attends fon he necnyed Bear uit: Once thy coment Mops one enthacted, the next step *s to move them fo a RELU Layen- ReLV performs an Uemant mst operation and aus oll He negate pixels to 0+RCr)= max Co, 2) 3. Footing Layon Poo Un. ¢: - W es g as Le "same ing oprrok Fi ol ns ton, | a roy 9 me fo, a Te rs mop ng SW goes trough a pool’ - a ° “a hoger, te Pn ten eee ae HUow ond Ste 2 mot (3, 4,b,2924. The poo Co meek Vertes filteng te telenhifay a e 4 the - Amoges Uke eolges , coh, Mya featrou, eyes, and beak a 6\8 e _tettenny og [TO] ‘The ee math 98 fal ob cpub te tne g ennectesl boapen to Hayy the Boge<7 7 Fully connected laren here ds how emackly CNN pecoprzs a bral 1> The pixels from He Imoge ohe fed fe He wnvelubonn( Aauyer. 27 St nesulls mm a convoluted mu - D> Te convo dred nap v's oppied to a ReLU GrrcKion, AZ Te fmoge precessel with murlhpte Lonvelittons and ReLU Myers por fovah ng the fatre,. S> differen poolng Layers with varrous EPS one used fo selentdYy speefter specipe parts op the wage. E> Te poolesl featur: wap %S Lottened arnol Row filly wonnected Layer 40 get the pol eutputyYon + RELV + Mar eee * Featoe exthachon & mnuthple Hdden Layers Ussi'fieotion im the owtpuhQa. expla the following : CH Case based reason ng. Case- Baycol Reasoning tr mache Learning i's an AT fechnljue Bhat & usel to Ame problems based on past ewpetencess The techarque 5 oltu'vel prom huumet Preble solug apProacher, where People offen rely om hei pout Cwpertences to make decisions ta new attrepnfons. CBR Rw Type OY machdne Hearning thot uses a clatat ase of previously Selweol problems or coser fo sowe ew problems. CBR cs based on te Clea that sintlar Problems cam bewt sdwvilay soliton, and te uses t's Atertlosty ko pinch holut'ons fo new problems. Sn Cose- Based Reatoning th machine tewrumny » a protlem soled by metnteving — stanf Lan Past cases and adaphy Hum the uke Aateah'ons The Key termes fer prosent by CBR Coral soso: A case SO problem Hrat has been previously Salwec| ond stoned an the clotubase . * Similarity : Me Atenlortty, measure vs useol to clefernuhe the degree of resemblance betucen pest Cases anol Hye puhent Stuahen , * Adaptation, Adaptation § the process 4 wrveclifyny a retrieved post cose pik He ciumrent BELO Procem in Case Baseal Remsoniing The CBR proce typically Invalues fou mary steps: retrieve reuse, nevise ond neboum,eB CorRetntever The prmst step aw die COR prourm os fH Rebnleve wekevowt cases Jrom a case Library, M's uveles semrete! throug the Uenery ta pine Caaee thee are Avian to the carrer — problem, The goats fo Felemnify comer Mot ore O% clese He ste cmrnent problum as pom ble , as these and the mest Wkely fo provilelt usepul Caformat'on CH) Reuse: Once relevant cases have been retrtemed y He next BRP oy wR MELEE them te see the cent problem Tis frvetees aelapdry te Koluhions used, An post Canes wo pit Phe cUbhent pho ble, CO> Revises Apen a Selubton hoy been elLevelopedt using past Cases, ee ment step cs te Meuse ot # beblen 44 dhe Curent problem: Ts may favelue modifpng the gelutien based on feedback from the usen om Mew Mfohwokun that has become ayaostlatle-The goal re vo rape rte Boluben yo malts ah eYfechve ag por'ble Jor the cunt prs en, CA) Refatay The feral otep en the CAR precem ¢S fo Reber He neu fy developed soluben Jor fbrre Lde- Tus mvelves aolelirs the new cose fo the core Uony So Hat cb Gu be usel om the netrteval ater for putre pro blemss The york & to conbwwually Suprove He Yulty of He tie ditrory and te efechvenem gf te CAR procers oven AYne, Case - Boseol Reasoning COBRY cycle. Gu. Retain On Soe noesSS Th C > Locatly, wuyrted Regrariion qaly's Win the pled of maclune Learning and repens notable fealty weynted Lovear Rapromven CLWLR) emerges of © MPA Sch dinate bolsters prectichve eccuneey Hemouge OO MMEYAEA'n OF Loca aolaptanons dn entrant fo convent tens Rephess em midels, usw Prune & wnwou of CoN, gruppo * Cekmn Emeng vertaber , LWLE Gchnowledger He Sy! PE tocatered puttinns and ReWwhtolys pre ow te batt: Sr the subseyut UiScounse » We embark on on anple ranen ¥ - eat el nherert te fundamental prurecyler , oliverte oppltcanrons > aewantages sfferech by docadly weryitedd Linear Regrerse'er + Onn aim es Yo shed Mw on fs ewcephional eupacity © ply predichve howe anal PAnish ttricar, unclerstantt Cf mtvtcate datasets Femdamentalty , Lie maamiP ests as a non parometue naprewicon Atgontithm that ciscernk Hy tonnechin bebeeen. we clepenclent vortakle end sevthat while pendint vanialles. Metabey , LWLeR’s oistimcnvenes em anates Prom obs dynam clap trary , wore empowers ct fo berthew Lsteret ways pen onclfusla date — pomts umbygent on Aree frowimity to the tangebt Pome uncler — prebichion- In emendes ttuls eGontthe aconas Peater mynstpecnse 2 propiimare uta parry , Feemy Prem as more ntluennial ont JEWS He pred Non prseusQa Wher & Aaadierh descent, oleltr sue & nutitrtoyer perceptron” Gaeptent DESCENT: Graktent peice (3 Known ab ene of He most Comme ny wed op iimfrat on alportthens vo Maw mace learn wodels by eant a otal errors behuoven saga and enpected remls. Furthen, gradient decom % alo used yo trou Werral Metworts + dn mathe mation! ternvbro boyy, Opiimization algonither regent to the tuk gf mutating /monimiring an ebechve yunctten fOr poremetentried by a+ Almtlanty , ay machine dearsuing , op bieteation ss the HSK menietng HE cost function Poratrtrel ty the meodel'S parcmetens- The man ety 'echve YH yroctent clesrent 2s te miniatie ta wnver funchten wang Henabion g parameter updates. Onee these mackent Cennnuty moll are optimtredt, these medels can be Used as powerpr/ role for Andypeiet Inielltyence amd vawlous Comparten Suence applicohions- Sn thig Jer ttewal on Gradients Desert or Mactume Lesanag, we wlll earn tm oletatl about Pradtent descomts che male Of cost yonct'ons Ape et fealty we Oo banometen with Maclin Leaning 1 types gy Gadi clescents , Bunning ote, ete, Gradient Descent 6 cleypmed as ONC the most Corral sed Aenabve oph'mcchon algorithms of mnclume Eeheers fo tra ah machine learning and deeg Cooney models - St hetps de pork ta local mulaivnnim 40 forchon, The ber yong a olepre He loot nentmim or Local msn of & fonction ubdy gradient descent AS a fo Uns?BE we rave fowonds a negative gaulienr er aay bem : te ulvet of the punctven wth ye te 2 Whenever et the tUnhent pathy ie bout mintmnm gy hee prnution We move “dowands © peithve Gradient en towards THR pratent the function ot the euuenct Pathe, we wh get thy doent mammum op that panchtmy They amhdre procedure whe alte oun abe uve HS to m™inimde, YS Rac 08 Cracliont Ascent, lemon ok Stee Posh der conte The BONE © Pruttent descers atpaninn he cost funedion ay Heradton, DELTA RULE: Te dilim rule, Bron anneal Werol nemo O spterye find of backMepagation tieh ant oy peng te mnching Lemouny Jonntfiedel srtedipence AETWONK y mascteg Oso Pat ong One g Apt Awol eutpets with aiyferent Layers og ornyfeval reine, Te pel. me i obo Called Hes Delta Learmtg rule, Gemerobty ’ back propagation hos tbe de with recok oles eb teadng Pras RA neque — dean HS by Capek wets bn onkytal neurons Delta, Lemurs beAseen op He AY OneO forget achvenion anol an ehtoumed ectatan act as on. By mary a hear achvawen fumenion, nemuorit coune tions ore balanced. Another approach Yo ep low tte Delba pule PS dak Ce WEE an AOA punciuon We Rory om Paden Olesce re Reanruig, Dela rule Rifers f0 the Compeuson of actual outpt wily og Forget output the vechne topy. bey fo Uscoven, the match, and ete Crcfrom — makes ctongess Pre actual Srecution ef tte Deltr mule wilt decks, ag Pr the nekoerk — and Ms comporititons SEU, y pepe a Wnean achive tion, perenon, the clelte Cam be bat wh MPN, 8 few ser gg nena nehuortky wiley sree pe Kady 4 beck pop ganion, MULTILAIER PER CEP TROND: The prrceptrons whieh ane arnenged fy Layena are catlesl euath’ loyer perceptron. Mus wodel hog koree legend) cn input Royer, emtpuh Layer anc Lidelon ayes Br he perceptions mh thy bapek hoger, the Lenk Prontter funcbon used ard Joh tHe Perceptran Sa tty Wdden Lower ond eudput WY, the Aipmacd al on Bmashed S$ pumehion is atest, The apa Alpnot propa prverd Wreebvron, On aw bayer by Guyer bors, eo te Ll Pereaptnon bar BCn) & thentek as a ATrophe went PAA Tern te nemurk i acbuwven by gemek Sap equst wo 4L, Ss AEs Tah utn), te Cady oe tam] welare on dente, vie Nenation Hep on applying ate Masithn Cores poreltnaby , woe, Gefns Ma meine vector as; O)= Cbd, wy nr, MAOIs 9 wn Cn] econclingly , aha, ORR ombrhen Oxtput ds WA ey ba the compach form o - VO) s awe OD end = WoT On IK n(n)Oa. Eeplam about SOM alganithin with help 9 sultotle OLragrom. Sey - Orngan'uing Fechnique woluich helps olla by nediay phe dnteaons St fotos an wrvpearvired Leanning approach and troihert ATS nebeohh HAoUyh a Kemnay Compete Learning ayentn Som vy mech fon clubteriy and mopping Con dn ennasom a Sieg Decluchion ) teclutques te map multiclimensional olata onto Lower demensiont uslich allows people fo Reduce Complere froblems Pr sasy Alenpretantons Som pos too Luyeha, ane 15 wh Sayput foyer ond tre other one othe Cutpud ayes. Mop (sor) provides % elate. visualitahton Yo unclerttand hy climenfdenst of dasa fe map. ny, he Fs BW Ancbe cture 4 Sey Or gamely mop with two clusens nel n mput se features Ava tTHm: Sep ls Satnelire the wulyhts wy random value may be OMumed » © Insttalize tte learnvg rote &, Siepr: Colette squrshed Euctoclenn distance. DEP) = E Cuey - AeA ohrone won ant >L tm, ep Avol — srole~ Ty when DE) PS mentum that vill consiclened ag windy fdr:_ a Saep ur Ara each f wit a gpeape neiyrbonkoed gf and por all £ f, colette ee new wesght - PY Cowl = cory Coth) re Tout Coty Cota) Bepst yds ee Canning Mate by matng 1 ACtti)sose p. SHEE: Mak ee Shopping Conshition.Os Aiseus Mm deta wh suttable anet'te ere the datvahen of back propagahon atyar'th mn. wseedo we 04S wee eo So wee oss brso6o Tas 0-39 Now coture the values of Hi orl Ha byw fonwand pars. Forward Pars To fond tre value g HE me frst Wy are Brpd value from fe woryhts ay Are wa xwe + a Rwr tb} Hy 2 0.3945.ty colutode He Poel rete ef Hi, we perygoumed fre Btytond famehion os He pier = + Hi gwat Hagrnol = 0.593269992 We WALL Colertate He value of rin oh 2h le way os AL % ee Ha = 1 x3 be Yedy EL Ha = O-05X0-2540-10% 0-30 40-35 Ha = 0-3925. T Colerloe He penal nasult of Ha , we Peyohmed tHe Ay tel forchta as. a Ka gegat = — I+. ete He pets * @ oars: Ho bel = 0: 5969 G4 393 Nowy we caleulete tre values of yi ond Yo as we colartete the Hy To fee tn the some wi ond Ho “4 Hed te value af Ys, we drat multiply the put value 2 dhe Cutkcome of Hy ond Hi from the weight as Yrs ALRWS + Me KwWe +67 EL= 0-+593269992 KOKO + 0-5969R W378 XO-US HO-G0 Ys 1-L0S90S3.- B colette the frat reset of Yr WE perormel the Atyeustel qunchin as Seo So eee J! Soak ae en Mites © eee we presto 39nc Gi ginal = 0 FSL36S07 WE whl cafetesHe tre vale of yo on the ome Yo Yor 0-59 32698 weg ce YL Hytwat He ke eho 42 "0.50 FO*SIERBUITEK OSS FOG Yor 402 UQ2LH, Te cotewose tre Final MULE 4] Hi, we perprmred the sigredel foncdion. as °, = L Go Gat Tea et Yoginal = L wo ey Yoftncl = 0:-F72928465- Our Fanget values ane oor and 0-89-O4UN Yr and yo value 5 net motched with oun denget values Th and Ta New, we will fel dhe Hil enroh, wetch ws simply me difjerrnce bebween tre outputs from tHe target: otputy, The foetal whor fs calucloted ay Ertl = EC target - ovtpat)™ Bo, dhe total enroy cS TB CB dette) 2 OR Ys iat FL Coon -o- te ease 4 (0:93 - o- sats aees)® = 0-29 “811084 porobalecedagg ! Erotal= 0-293 3FLEL Backward Pays at tre output Coyen B update the waipht, we calustode the enon correnpencl fo each welyt with the help of a fotel OHEN- The enor on wert wf cokcutated by at erent ioting total exnor volt nespeer to We Enron, = IEtotat OE toto OwWe Perjonm backwort Proves gy PAS comadelen ee tags wort Ws oh Enron og 5 DEretel ta Bws Etotat = ATs "Yb grat)? + ACT, Hr Btant) Pe. 6 From egar ation deo, Sh Ts clenh wD Weak we Cannot park ifercnttae ak att pect to wos be cmuse theme fs we wer ONY Ws we aPlt equohion one tak PULA LE terms Ae that we com coat ly Lif fenrmbtede Mit, Prespe ct WK Ws ag ete DErtet ge Dy teal w 94. 660g) BY grout By ge Backword pas ot Hidden Layer, New, we wilt back propagate 4 oun Lidolem Layer o VMedott — te Wet Wi, 2,3, onl Wy oy We howe, Bre with Ws we, WR and We wotphss . we wih coburlote the ennor af WL as Gon 4, 2 DE totes Ow. E teint = SCT yy he 4 (T-Yagnat)? From arprabion(a) st te leuk vhat we Cannot poswatty Atferentrate fe ty respect €9 Wr lecause thine no Comp A we split equation (1) anso murthiple terms So Met we un Caily denen of wlth Reaper do Wi as DEvind = Zeit i HBr xe DHL G3) Ae ous OA gat on oeNow, we cefurtote each fenm by » one by one Yen enrte Etotel with nepeur te wi oF Beam = ACK Ce De FC ce) Bex —_ BMG Goal any we ago apd Atus because shen Sno Oy Hy boul gam ow 6 ea 2 Exral = 96: Ex hic leleed) OL Frrol OMG freak OM penal 8g DER eyo ple because 87 GL and Er BA pal OM gunad Khe Sno Hi dorm. Ap Uiteg v8 done od 06: _ dG x Oye te) 344 Grel dys OM gro Oe 2 Be x Oe ...c19) DHigoet Yn Mga wet re both SEL and AG because Here "> 79 ofan Aap ga yo Ong Yr and yo tenm so EL and E02. we aphit oR ab: gee > Re we OY ert... Ce) Oye Ss ot dy. b& 2. de 3 a) = oe * Yrdroat > “CS Oy Dn tinod Oye