Say LS % Convolution Ne ehuovk (ent 9 vral ay ve ( —= JOE ) ® Archietecture of CNN 5 Feature erbactton Li Class Resiree ss ir fe VA Peela Hidden ovlput ; ly conneched CNN ; Be 88 ep oF thre diffoent pineal A} convolutior aye 8} Posting tayey S Sere connected (ayer (ecrons innct , tare] or Cael a n fonction ave eee a A] convolutional Ja » There arc several fF onal gf pooling Jayes in Feature ext to entract di ffeent %) convolutional photos character} s tlc wy) Tk applic! kemels or filttk to inpy® images the kKemels or Filters are smaller matrices vsvally bye exe, sxe share input \ \ bet? the Filter and inpy is obtained be see orca Image: nts by sliding Fil wh as Featvre Map fas been applied to inpyt convolvtton Jape Eeand Fe 1) DoT prodvct pictures compone ») The result ts kno w) once convolutlon operationoutput to nent tayer 8] Pooling Layer : 9 Uke convolutional Layer, pooling Aayer ts inchary« of shrinking the tonvelved Feature spatial stv 1) Through dimensionality seduction , amount of compol ing power nected to process Ale data will be reduced foams of pooling axe? 9) Max pooling §) Average pooling a) Two 4) Max-posling ‘ en max pooling we we that the Kernel areq of pithy as to downscole an input by take the davgest valve form the as covered: > TH primary objecltve redveing 4s dimen sonality. D average gooling’ overage of all the valves From the — Here we take Che that the Remel has coversd: ava of the Pi uly, connected Layers (60) GER TH MATHS classification oF regression Ect computes the Final task: we mrethed of Jearning non: tinear wy) Te is an anexpensi tombinaltons of high tevel charactristes represented by osttet of convolutional Sayer. i) tp FEL, nevren applies a Hincar hea na fortnatiann aie input vector thravgh « wecgh matin: fy A non-tinear bansformalion is then applied to the predusle ‘though a non-tinear activalion funciton £wraps the dol prs YD The activation functton bet™ the input of faye and the weights matin of thet tayer Hy myped sates deyer Q Rely Layer + (tom tines aateation font) Ce Ans D ReLU stands for Rectified Linear Unit - i) Range of RelY fon? > [co tea) Graph + # ) EOD ce releo * i.e ‘ aw the 4) Bs shown in graph, any negative input given te The Rel for? durns the Value into zero- immediately in the graph © The Purpose oF ReLY Sayer ts to amprove_nenlinesity of the tmage pixels dota. ® In kelY fayer we remove every negative valve form the Filtered image and yeplace tf with 20+ © so when input is below zero , the output is 2e¥S @ However when inpyt rises above a certain threshold | Ht has dinear velatonship with the dependent varietk ® Rev layer works beller than other activation fon dike Tanh, sigieid HM ete. because ERAS faster wither more -accurocy- making significant difference fo accureg © 1t is also immune to vanishing gredient problem« ReLv operation on an image * convolution operation of an image, oa feature map, which might be ReLv Jnye When we apply then generates Drewes Fo week het nani ot SE nip we don't infreduce none Linens Uy, than out Deets network modal will become teadiNonal Jinear classifier only: = When apply cenvelviion, original image gets converted te feahne map whelt, black portion» are negatye and white portion a positl® valves all negative valves 5 so when we apply RELY fon™, she gamle toy 209 soy ON ee Black povtons ay \emoved when we apply RelY fon’, 50, Feature Map new only has non-negative v9 after applying ReL¥ dayty Q _propout layer: satis Le od Rely artivntton, mah ling functor soFiman Ja For eleesi fenton ves end © propout Jayer ads as o mask, that sliminates tHe contibsiion of some neurons towards nak Any er while maintaining fonction ality of all other neurons © ie we apply Avopput Jayer to inpul_vecto, some of ils chavadaitts a eliminated, byt tf we apply fea hidden deyer, some hidden_nevynns 9¥* etimninaled B. provp Dropout layers avid overfitling on baining data, go they aye ess important +MoM) Klemm UI subject dm Or call here 9130509737 Price is extremely nominal. You just have to gpay/phonepe the amount Tam M Rl O hse Pi linkY or pdfs will be shared to you for this subject. You can easily clear your semester. These notes are in highly simplified form and from exam POV. These are premium notes. Very easy to study.| feel everyone should pass easily. It took lot of hardwork for preparing these notes.| have worked hard for this.So please don't ask for free or bargain. | will be happy if these notes prove helpful to any1 of you. Thankyou frnds N All the bests. BE computer all subject notes are available. Any doubt feel free to ask here sé omkarpawar_0707 Edited 6:29 pm -7Se Srey ae” Page 3 © there are different Jaye in cnn. © They can be mined or repeated any number of times This if inketeaving of Jaye in cunts D unterdeaving bet fayers i imp. featuye iN CNN os it makes vesearch ond experimentation in Cun intersting B The comclotion, pooling and ReLU layers ave typically interdeaved iM a neural nehwork morder to increase the expressive power of the nehwork co D Raw dayelPoFten follows the convolutional Jasper D afler two or the set8 of convolutional -REWY |, combinations ,one might have a mex pooling Lays BD For example: CRERP, CRERCRP repeated Fw mes B this entive pattyn might be nevral network - qnorder to create a deep wy Dike can vepeat the fink patrn fue Hmes followed by fully connected Jager (F) « CRERP ACRE FE vanolinkar Pawar's Easy Notes “Te anvdes J Aayans, encluding input tere? Is is the process of incitasing the step size by which you slide a filly over an input img = Th shows how Far Fille moves in every step along oh aiectton - = poy eg when aide is a, we move the Filey fe rine at ottins land chen shide ts 2 then we move the Ties te 2 pinds af a dime 4 So 0n-@ Local Response Normalization” ['F") fr @ LRN ts a non: bainable Jaycr - D It square- normalizes the pinel values in a Feat mop within a focal neighborhond D pn is always vsed immediately offey ReLY Jayer tia wee of Hie Dick aide generalization Dd Because RELY Nevrons have unbounded activations and we need CRM to normolite We DIF we nermelize around Jocal neighborhood of excited neuron, 14 becomes ever mor sensitive as compared to HS neighbors © thee are 2 Apes of URN, based 0m neighborhed A) tnter-channel LRN 4) Tnta- channe) LRN A) Hee the neighborhood ts defined across the Chane! B) Hee the neighbor hood wame channel + 9. faaarg] = rh involves adding he input featur roa adde eyivel 1032 ding valve 4 borders of the he pao of image shee the Let only valid part sf imye ts extended within the an semalines Fler dees not At perfectly te dhe input images so we SSE padding techn exta pixels ayoond the b p before convelulion J for padding cam be calculated Wer is zero padding wh inpot Rebor order of = nor of pinels te be based on sve of the most common fad involves adding erates inaps fo) ei walld) gaadings aes pve MEd GTS ay ekea Training a Convelu onal Netyuark SS Following are the steps inveled in Waning oq CNN H Yplond Dalasck ! ere we upload datasel. fe ef tage een renee Inpyt Layer This skp rods the data. Hoe we need fo specify whether image contains color or ot € Hee we apply Filder to exhact important Fraluien. we alee define site ot tell ASYEL' This fayer ts veponsible for reducing oF shrinking Ahe size of Input, & Additional conyolylienad : Here we can add as many and pooling Resjers ! convaluton and pooling buyers as. we wank’ © Bully connected Lenpey ;othis step Flatitns the previous te SF Dene Layer form Fully joined Layers = Tt computes Final cassificatlon ov vegyeuttory H Logit Layer Finally we define Sask Sayer with predictton of model + &. Train mode: once archtetecture [5 finalised enn can be tained. For practical implementation packager Fike Hnsorflow, Keres, pyterch can be used: Test and evaluate , ginal step ts to fest & evaluate the SM oaensinedel performance of _medel