pce DL Cunit-3) = “| =Convolulion __ Neual,_nelwork CONN — _ a Inlroduclion to cnn: : is [A convolulional —neuxnl nelmonn on cual is 0 Nelo int ee del inant Cano — mode_up ‘of an ——liseneform an inpu a =| lycun a cw 4o do image fla da gge.aaalysia —— loss including a Scene classificalion abjech idelec lon _¢ seqmeniation ¢ image proces 1ag-_$§ L$n avder 1a ‘undersland how “Cie work Itneve cre three Kew cancenls - local seceptive itieida, shared wig hls € biases , & achivalion SS output anpuls Sear oulp i | Bayo seul ty. —tanluinal { “NelwoyK err NeboPogo No. NS Bate 1s */ Aachifecture of cM = L fonvs — conva __ronv3 con on — er (A+ | ae! S Booting | Peating Pooling Ssnoul Vidden layer ee clogsificatiog = Bp cnn architecture consisis oF 4wo fundamental _companen}s — fenlurve extaattion is a _paacedure tha} uses a ronvalution tool _{o Seperate & identify the dis) acl characteristics of la piclawre for study : i = there ore several pains of © convolutional oct! 6 in the feature extaachian nelwatk —!A funy connected lauer is a makes use of [eanvalubional process's oulpul ¢ delermines l4he class of the piclure. using the — | features thal weve _prreviausly extracted =) enn has ther e layers - 1) convolutional layer |» ¢ aroy_P layer 2 ta_tonnected layer ————_——_#i the basic _stauchuae of “a eanvelulianal | nelwork - X| Padding :- =\ i er the dimensions _| r — by inesc ening or decreas -lat ciloms ua. use a convo Mtienal 1 -___ without necessavi L widlh of the vol = Padding € stride influence operalidn 16 pevfarmed = Due to padding , information an the hander} ho of images are also preserved Simeloaty rot the “enfer df image =| Padding Terminologies - 0 vatid padding ouput = inpul - kernel + £ Sine sine Sine | |0) gs " f Bubp Inpul ' ! gine sine _ 13) fu padding Z e it (oulpus = “inpul_+ kernel = 4 Hine 512e SineSteride *- 1 ; how convolution | d_slride influence operation is performed $$$ — over. Sleide conjaals how filler _convoive: inpu ~ 46 Glvide js sel to 4 filler moves across. 4 ipixel of ao time ¢& if mide ia eel to 2 | Filer moves 9 pixels of a dime ; = imone the value of sleride, smaller witt_he the isesuiting oulpul <¢ vice versa = T 1 | 4) 3 THz! S| n | elle = [alls|el| > fires! [ES {2/9 Input Joride =4 lglmide io tep far suding the convelutt= Na) twin Sieride Sse es equal fo 4 like pa used jo atler +he dimensi adie input | oulput yeclora ei lhe by inc easing ov decteasing | RelLU Layer : : _ "ely siond for eclified linear unit RelV 16 computed _aflea convolution —___ \_\5 mos! _cornmanly deployed achivation funclias Maal allows the neurral network 40 account for __non-tineay velationshi ps ee -||this function has _}wo_ tenjart—ashiant ges | — over _4}9moida.!— Such gs 6¢%) ox tanh Ct iPage No, Dato! =| Relu mi be —coleuleled —exleemeliy- a because Yui thal is evequins the inpul tothe Value _o = Addiliy depending on__whether ow not — its i na Sh cnegolive thas a _derivadive — of eithey _o Relu (x) = f 0, fay uo bt | | C4 for x0 logisdic ££ Rely + détw/da => | vd Re : L oo. 2 4% 6 8 10 { —hg ReLU_Vs__814moaio/ Se { *) Poona 3 r= = Pootiny layer operates on | ndependen oo i a ee eee fe ee =| thie reduces resolution of the feature map — hay seducing, —heighl @ width of fealures maps bol redoing Fealtires of the map requived fax clos Wieadion. This _is_ca\ted dawn -samp Vigg | i T = each fealure map| it = aura i di ease i Tpiclure thal the Kernel has covesred, _weduened by ma i - iq} is the mos! common better crear tS ying: 2 average of att the values faom the oren of the picluare couered by the Kernel “is what is celucradd by avetage pooling approach os it — yo} | IF oo loo] 19] a ral ig) [rs fia Ye} t 3 [oof a3] ufie{ 24 7 ie wf sii Max pooling Average | t r PONG, pany =f Poa ing can—conlarol the over OO ING, Ee tata ae same__tresolution 05 input i connected__Layeas— Sub connected layer loons ike neural nedwork connecling—all_neurans’ & forms the last few layers in _+he network 4 the oulpul ferom —fiablen nye Fed _jo__this- fully connecled layeroverFitting 3 more time for Y_Joro ing Dropoul 3 Reduces the eH CitLingS. Le. improve éning—specd. 1 =/Aan the inputs — feram _+his loyev acre connected to every activation “unit of the Nex} lover -| Since. ur the ponane lees are occupied tala. fully connected lay l ar kithing -| Derdpoul is one of +the nigues thal r - L seduces oven titling This _daapout aale is ——— (Usuaiy oS I aut | 7 =a oes = OL Oni OO Onmis {| Tp One, convolution : : One oye =U oO ©. i anpul Max pool __Flathen ia conneded en : Volume : layer See fig. funy connected layer =| fuuly connected {ayer genevole final _cludsificalion these different layers in Cry cum _be mixed. On crepeated ony. number of times this is -l qh interleaving of layers_in_crval += Halewleavi is “an imporlan} feabure parovi= -ded in which makes reasearch & ;____| expevimenialian in can thal much more leesling i~|there ore several cam designe thal m 5 = Below 15 |) Lonel the ibe used ¢ these _accchilecl lute have heen lis) of a few wrchilectures | 2) plexnel 3) vac we} | _5) Res fel I 6) 2¢Nef 4) Google tved | 4 aa ] PAS Biegers [8 toyers Wwanaiiew Tal== fe} = fat pa Alexwe} Fig LFF” own archilechure —cuith Leen layers mT =n to mf Ste 71S, h SI a AISTA_ lalate 12 F|felo|slal | -l4 Probab mage} lei 1 2 ISIS NoLL at Li /l2 lat 22 Sslelellels\slsllsie|q_ wees a6 eS | = | = fig. em VGGNe} ancchileclure. n5_have been explcuned tn | the above MFEN crn _desiq £ 9 Wess =[Pee Wo.: S Date Li 7 I i lI | Loca Respanse__Normaliaadion = _______— : local _Respans ——+mos] important normatization techniques —__— Beli wad einplnged. “aa the adlivabian an? “instead of the comman tanh ¢ 41 Luohich ed 4o the initial jnleoductian of local Response ormatiration clan) in the Blexiel | —___Larchifectune Il =| is the tolal number of channels =|€K.diB.n are hyper. parameter Hat cre 9) Inlew = channel LAN ~|the pixe) values a} Ca.¥) location hefaae Lotter normatization are ata.u) € bCx,4) gesp,| _constant -|K ig the nowmanisa si } rol ciny _2ero division (singularities )_ 9) dnioa-channel_ LRN peared allel ain tok — n Nee ore eo nid & ee Solve image a Neighborhood is defined around the ipixel under considevation in intra _-chonnelPages No. =a 4 T aconvolutional wvedwark Generally inorder }o build & foan oa cn following Sheps acre followed —_____ { SLnpal Layer - dhs slop ~rresels the dala -| d-Sine- is equal bo the square of the aol it =| fo Example - if picluve has is6 pixels, the fig ucre 15 96 X96 contains color on na} 2 convolution layer - -\ Here, we need: to create cansislen} layers -i We a various Fillers ta learn imp Ifeatunes Of the nefwanrk —\ We define the sire of keyne! & Volume of | | the Fitbeer ! | | Pooling layer = j_ dn the Haier Slep. we adda pooling { layer - o - chia layer reduces the sine of Hye __ i input = ; 14—does__by ta The maximum value of the Sub-malrix Additional Convolulion_& Pooling ayers - an this 4tep._we can add as mony conval Jion_¢_.poaiing layers a5 we dant.[Pepe Wo, Pete ti S| dense _tayer | fur ly—connected layer —________ §n his lop. we “ean use a differen! acdiva-, ——+Henfunclion_@ add the de dropout effec! Splat layere = : -| 4} final _teye Slop ia the prediction 4) Train the model - = once the archilecture has heen finai2od th CNN can he dacined 8) Tes} 2 evaluate the madel- }_=lonce the model has been denined step is jo jes} the mode & evaluate iss | performein cecurrent ¢ Recursive mets aaa unil- Recurven| € Recursive els eed fee a Graph - the strruclure of a collection of computalior fuch_os_‘he-mappingaf~inpula €earomelens. [io tunable 4 loss.—may_he _formauised diag -a_compulal tonal _gaaph wWe can Unfold a recursive _on_crecurren}| Network Example of unfoiding a cecurren} eq? — classical foam of a dynamical Sysiem 16 72 £0675 9) : _where, 6{) is rated the stale of the syle. for £=% time steps we gef | 6. FE 5 @) | = CF (5s @); 0) » 7 y Fare <= Ft FAI? aA? = Ter fig Unfoiding = Dynan ical_ysiem Deriven by « external al goal : As another example , consider a-diynamhical_— Het glem deriven_byexdernal_ signal ab f Os Fg yO 5 x 30) slale now fanlains informalion ahoul Whe whole pas! _inpul sequence ._Paga Ho. S Data: 7 unfording aecunrent Neural Ned woork To indicate thal the alate ia hidden jeweile using variable bh for state 2! ¢ =f Chcb-D wt» we cy SOSH ewO PP! KF 1 ae {| ¥ untotd - ya ie) S@_O ¢ Adv ee 5 Py | imadel hos same inpul sive : | yl eossible to use same function £ sity sam | porameters of every sleps 3) learning 4 single shared Coode| | Mou *!) Recurven} Neural elu - ~TRee yp ureren}| Neuval Nelwork (RNA) 16 q eof neural nefwork In wh which the __ | iepesutle of one slep ave fed inta the next | step's compulalians =| traditional neural nelworks have inputs— é_oulpuls that ave independen} of ane another. ___ | fidden\ : tnpul Po Neyer {oul pul fig. General RV 2eco No: Date: iS i Run have a ‘memory’ thot cred cin, atl data elated +o talculation .a{ executes the same acfion an al) of | a the oulpuds Mons Ran ons - consider a deeper network thal has 2 hidden layey one input ¢ # biases far hidden layer 4 are ‘Cw bt) (we be) for the second hidden layer € (Wa, bs) faa the ied hidden la Tapul Sho, by) +©2b2) G1.) outpul] output] ineul fig. Recurren| tote in Rn vridden—toger snpul layer XN owout layer ‘ | Recurrent Net | ae fig. Recurrent sabe In RNAto use AAA There were a few paoblems for wand neural network which developmen} of Rv Niean't deal with cangecufive data Merely bakes ina account cutrent input BS 4) unable to cemember _eamiiow inputs i The RWN offers a mremedy far these problems pS RNN can handle sequential | | data accepting both the inpul data being Teed —al ¢ 2 Advantages of Rn) 7 | | a), ¢ L | previous inputs through their hidden s4ates a) Ans hove the obinity to relain info Rul can perform onmtine ea Raw trained on a ange _dalase: + | pisadavantages i ; )| computationany —imbensive Dt have a tine apacily Anns 15 highly sensitive 4, the gigs of initial eweighls . pifficurty, Ta_capluring Global__infarmatian | [ark of ialenpetobigigh ioe ie i =| Recurrent neural pee thal ave biclimechi ~ onal ave basicatiy 1 twa sepevale ens — | combined. — sos =| for one nelwork it ig fed in +Hhe~ inpul i sequence 19 fed in orequla time. onder, | foro dite” nelwomn it’ is fed in | ImeverSe lime onder -lpt lime Step. nelworks care typi aliy concatenated + however there arre athaa choice. uch 96 summation. -| th word info" about the sequence af each time 1of honduseriting recognition =the concep] seems easy enough bul-wiher tt omes to actually inoplementin ga neura\ Net — which ulirniees _bidixectianal structure, confusion _ arises(ae ‘Tas Sequence concat concat | ‘[eoncol | a RR / {s71 sir (is1\e}— J > {= ] Se, Est) tals} (sr x Se = hy He Xa 1 | i IL Fig. 0 confusing formulation. la i |_# lEncoder- necader Sequence fo Sequence architecture 1 eB shen_iele ile._ate_of same leagl =| Quins a very useful & can be used under iYpree cages - |) 10 i ixed sae veclay |) ¢o map a fixed sire veclor to a sequence. 4) To mop an inpul sequence ta an output sequence of the Samelength RNN when ‘Jo are ef hol same length = | Encoder- decodey imace] t =| Me-re_use—considehow_an_enn_can be Jonni | | | Input Sequence to en oul : = | Sequence —arbls_is_not necessarily the etl “ a 4J _ =| the encodea- decoder model _1s composed ii on three pa mover building blacks + J 1) Encoder | 2) Hidden vector / Encoder. oar | ay Decoder fu ola imensior a AMeclar Ceram the inpu} sequence Chidden ‘ : Nechorr) | . =| the hidden Neclow wit he 1 | tne audpul sequc . uy Yas { | 7 7 Ei 4 Encoder & Lendl = ave 5 } & 7“ | ann alan Fen) % | ( i \ \ &\ necader EF | \ % qT 1] ka as] ane 4—___|__ __fig_Enendeer- pecoder mode} ___ x ft I En =the encoder may be caeated by 4 Jelacking many Ani cells RNIN Guccessively _ ES rN 0 mA eee noe a j_9) Encoder veclar- This is model's final hidden stale created by the encoder 2) Decoder - The final hidden Veclow ablained af |the conclusion of the encoder model serve sa xhe decoders oupulQuestion nawe Speech erg ynitlon Time Series apph Most recurrent neu 20 val compulatia nelworks ‘parameters £ rela 2) Moving an to the d Foran the prreviou. imay he broken down inta three 1) ferom the inpul to the concealed alate 12) from the oufput othe concealed state blocks 4 ni cd trans Cowmations : | ed following concealed state one canbe Depth inlroduced in In RANA S ways i Vinecurren} state broken down into agqupa Tak is possible ‘le thal lower. levels oF the hierarcht ve a por} fo IN convertin, iE the wow “input into a crepreseniation thal ig more B6ulable for £ chi shate y GZPepeNo, Date: Deeper computation in hidden_}o hidden Adding the extra depth makes oh Jime “af oa voriahle from time Step - fo LO Variable in time Step t+1 become longer Oo 7 Te) y Ome o fig. Deep RNIN 2 antcaducing Skip connecfions - 4 Adding kia connections in the hiclden - 40 - hidden path o | O @-DrE Fig. neep Ran 9see SS, ; | oe ful alouchucred prediction = | Recursive refers 4a the neural _nefwark's { Lapptieations Jo ils output = Recursive neural nelworks are capable of handling hierarchical “data ——* | computa i Ne}work =|} generanres a vecurren} nelwark from la choin to a trree. ES la Jariable Sequence ot), x¢2), ---- LCL) fan he “mapped Jo a fixed sire cepresenlalion with 1a Fixed Gel of poaamelors Recursive Neural Nel. “TreeTF Need foe. mecursive neural eknork in vile | Jexi generation & machine taanslation andin