1. Deep Learning Ia gi? [A Mot mé inh todn hoc da tran logic mer 8. Mot nhénh cla Machine Learning sit dung mang no-ron s8u dé hoe dc tamg, ‘C. Mét phwzong phép phd tich df liu dua trén cy quyét dinh Da Mot Ky thudt ma ha dr lieu trong co sé dieu 2. Tai s20 Deep Learning phé bién trong thei gian gin €8y? ‘ADU igu lon enn ph bién 8, Phin cing tinh ton manh hon . Ce thuSt togn hoe su da duge ci tién . Deep Learning la phuomg php duy nhét dé gi quyét bai todn Al 3, Mang no-ron nhan tg0 (Artificial Neural Network - ANN) bat churde hogt d@ng cia! |A.H@ théing thn kinh sinh hoe B. HE théng logic team hoc c. May Turing. Diy thuyét xéesust 4. Trong mt mang ng-ron co’ ban, fang du vlo (Input layer) c6 nhigmm vu gi? [Link] ton trong s6 céa mang B. Nn di ligu du v2o va truyn dn céc tng tgp theo [Link] chinh trong s6 68 etm 18 D. Bus radu dodn cust cing 5. Him kich hogt nto thurimg durge sit dyng trong cde tang fn ca mang no-ron? AReL 8. Sigmoid . Softmax. [Link] 6. Tl 0 ReLL thudng due wu tién hom Sigmoid trong ede tng dn?[A Giipgidm vn dB vanishing gradient B. Tinh ton nhanh hon . LuOn dm bio gid tri ura nd trong khodng (0,1) . Cung ed sv phi tuyén tinh cin thit cho mang 17. Tong he su, "overiting" xy ra khi: [A.M hinh hoc que t6t rn tap hun huyén nhumg khong téng quét ha téttrén a igu moi 8B. Mé hinh ed s6lugng tham s8 qua it C. Mé hinh khéng hoc due tir dU igu .M@ hin dat dO chinh xc 100% tron tap kiém tra £8, Kj thudt ndo gidp gidm overfitting? ‘A. Thém niu I6p vao mang no-ron 8. Sir dung Regularization nh L1, 12 .Thém Dropout layers . Thu nhé kich thu tap dir Bgu hus luyén ‘9. Qué tinh cp nhgt trong s6 trong mangna-ron duge thyc hign bing thudt todn ndo? A Gradient Descent B. K-Nearest Neighbors , Backpropagation D. Apriort 10, Ham mat mt (Loss function) ¢6 vai tré gl trong hoe s8u? [[Link] gid si hdc bit ita dau ra dy don va gist thy t6 8. Gidp ti wu héa trong s6 cia mé inh C.Xée dinh efu true cia mang no-ron . Gidm 46 phir tap cia mé hin 11, Adam optimizer la gi?|A. Mét thu ton t6i uu hoa dua trén Adaptive Learning Rate 8B, Mot bién thé cia Gradient Descent . MOt thuat tod foc khOng gism sit D. Mét kién trie mang no-ron 12. Gia siy ban c6 mot ma hinh CNN v6 du vBo hich tue 226%228%3. Sau eéc lp ‘convolution va pooling, ban nhan due feature map kich third 7x7%512. Layer tiép theo la ‘mét Fully Connected layer v6i 4096 neurons. hi chuyn FC layer nay thanh m6t Convolutional layer, s6 long filters eta Conv layer nay ld bao nhi@u? asiz 8.1024 4096 2048 13, Mang no-ron hdl tiép (Recurrent Neural Network - BNN) thurdng duige sir dung trong bai ton no? A Phan oal doh 8. xirly chub thai glan C. Nhdn dig cho’ vide tay . Dy dosn chuai vin bin 14 Tal sao LSTIM thuréing durge sir dung thay vi RNN truyén théng? ‘A. Gia quyét vin d@ vanishing gradient B. C6 thé ghi nbé théng tn dai han Tinh ton nhanh hom CRN . kh6ng cn training dt ligu 15, Convolutional Neural Network (CNN) éuge sir dung chi yu trong inh vie nao? Airy ngon ngtr ty nbién B. Phan tich anfDy bio ta chin D. Choi cd vua 46, Thanh phidn chinh cba CNN bao gm: A. Convolutional Layer 8. Pooling Layer . Fully Connected Layer D. Decision Tee 117, Mot mang CNN c6 3 convolution layers, mai layer e6 kernel 3x3, stride 2, padding 1. Néu {nh du vio €6 kich thu 128x428, hich thué dura sau 3 layers la bao nhieu? Absa 8.32082 16x15 Des 16, Dropout trong mang ng-ron git: A. Glim overiting 8. Tang Be €8 hoc . Logi bé trong s6 khéng quan trong, . Ca thign kha ming téng quét hoa 19, Phép tosin Max-Pooling trong CNN cé mye dich gi? A Glam chit dir igus 8. Gif la dc tang quan trong . Thay thé hoan toan Convolutional Layer . Tang 35 rgng tham s5 cia mang 20. Néu mot mang CNN cé 3 convolution layers, mal layer cé stride 2, nh fu vdo c6 ich ‘thurée 256256, kich thuréc Anh Bu ra sau 3 lop nay la bao nhieu?A 12BK128 B. 64x68 6.3032 D.16:16 21. Th gléc may tinh (Computer Vision) lagi? |A. Motfah vige eG tri tu@ nha tgo lip may toh higu va xij hinh dnh/video '. Mat phureng php ma ha di liu hinh aah Mot KY thugt d® hoa dé xr Wy nh 30 D. Mot cong ng xij van ban ty nin 22, Nhding tie vu phé bién trong thi gi méy tinh bao gdm: 'A. Phan loai anh (Image Classification) 8. Nhan dign d6i tueng (Object Detection) . Phin doan an (Image Segmentation) . Dich ngén ngtty nhign 23, Phép bién di ndo KHONG dug sir dung trong Data Augmentation? [A Xoay Bn (Rotation) 8B. Thay 461 66 sing (Brightness Adjustment) . Thay di s6 rong pixel trong anh D. Lat nh ngang (Horizontal Flip) 2A. Mot inh du vio 6 kich thu6e 6&x643. Néu sir dung mét convolution v6i kernel 55, stride =2, padding = 1, kich thuréc du ras 8 bao nhieu? AaLaDa 8.304303 ©. 3323 2.16163 25, Transfer Learning gp ich gi trong th gide my tinh?[A Giip hun luygn m6 hinh v6t it dr igu hom 8. Tan dung dc trang 43 hoc tir mé hinh pretrsined . Gidp md hinh té:nén nh hon . Loa bé nhu cBu sir yng GPU 26. Mot s6 mé hinh pretrained phé bién trong Transfer Learning la: A ResNet 8. ves Vision Transformer D. Inception 27, Trong Object Detection, mé hinh nde host déng theo ca ché one-stage? ‘A Faster R-CNN [Link] css D. nas R-CNN 28, Trong YOLO, "You Only Look Once" cé nghia lagi? |A.Mé hinh chi cBn quet anh m6t lin dé nhan dign 66) tueng '. VOLO chi chay mot lin trén CPU. C. Mé hinh sir dung duy nht m6t convolution layer . YOLO khong cin hud luyén nu fan 29. Ban dang husn luyén mét mang CNN, sau mgt s6 lép convolution v3 pooling, thu éurge he rung eb kich tude (H.W. Ce(8,4,512). Sau d8, de trung ndy éurge Flatten va di qua mbt F layer 1024 neurons, tiép theo la mt FC layer 10 neurons dé phan loal. ‘Ban muén chuyén hai FClayers nay thanh cée convolutional layers c6 filter FxF. Hi git F cia hai lap convolution twong ing la hao nhiéu? Aaa 3.16 cat Daa130. Mang nio dug sir dung phé bign trong Semantic Segmentation? AUNet B. SegNet [Link] [Link] 31. Kj thudt no €6 thé gidp gdm 56 lrgng tham s6 eta m6 hinh CNN? ‘A Dropout B. Batch Normalization . Depthusise Separable Convolution . Fully Connected Layer £32. Trong mét CNN, téc dung cia Pooling Layer lagi? [A Glim kich thude cla 8c trung 8. Giam s6 lugng tham so ([Link] aut ec trang quan trong . Logi b6 nihiBu trong dnh 33, Mang GAN (Generative Adversarial Networks) éurge si dung &@ lam gi? [A Sinh nh mei tr dr ligu c6 sin 8. Nhan align di wong trong anh . Dich vin bin . Giém nhiBu trong anh 34. Thanh phn chinh chia GAN gbm: A. Generator 8. Discriminator . Fully Connected Layer D. Pooling Layer35. Mot Max Pooling ayer 2x2, stride 2 duroe dp dung lén mbt anh 6 kich thuée 128412868. Hol kich thutée ura? A besa B. 329264 127612764 D.2pea2E<64 36. Kj thugt nao gidp ay anh €6 du kign dn sing khéc nhau? ‘A. Adaptive Histogram Equalization 8. Gaussian Blurring Batch Normalization D. Affine Transformation 37, Optical Flow éuge sir dung trong bal toin ndo? A. Theo dot chuyén done (Motion Tracking) 8. Nhgn dién d6t trong PhS ogi nh . ng t8e 4 xi y Anh 38. Trong que trinh backpropagation cia CNN, thanh phn no sé durge cip nhgt? ‘A Trong s6 (weights) cia eéc kernel trong convolution layers 8. Bias cla efc convolution layers . Trong s6 cia Fully Connected (F¢} layers D. Tit cede phuong dn trén 39, Kign trde Transformer ndo dug ding dung trong Computer Vision? ‘AVM (Vision Transformer) B. Swin Transformer BERT [Link]440. VIT 6 gl kde bit so wei CNN? ‘A Ding Self-Attention thay vi Convolution B. COthé hoc méi quan hé toan eye trong anh Cb ft tham s6 hon CNN. . Chi hoat d6ng t6ttrén de lgu nnd 421, Trong md hinh CNN, tal sao khi chuyén mge Fully Connected layer thanh mgt ‘Convolutional layer, ta thurdng chon kich thuréc kernel F bing spatial sie cta Bu va0? ‘A. B6 gl nguyén s6 lugng tham s6 cla mang. 8, 86 tinh mat thong tn spatial . Dé chuyén toan b@ feature map thanh mét single output vector (ging Flatten). D. Tét cb cde 6p dn tren. 42, Opencv lagi? ‘A. Mot thu vi M8 tr x nh va video 8, M&t leat mang no-ron nhén tao C Mgt cng ey lp rnh din cing cho NLP D. Mét framework ing hus lnyén m6 inh 43. Khi no nén sir dng Data Augmentation? A. Kh tap do ligu husin luyén qué nh 8. Ki m6 hinh bj overfitting Khim hinh 6 qué niu tham 58 . Khi mudn tng khd ning ting quit héa cia ma inh 44, Néu dCrligu hud luygn c6 sur Khde biét lon vé mau sSe so vét dt liu thye té, phurong. hip Augmentation nao phu hyp nhit? A. Gaussian Noise B. Color JitterHorizontal Flip . Random Crop 435. Khi sir dung Transfer Learning, yéu t6 nao quyét inh c6 nn Fine-Tuning toan bg mé hinh hay chi hun luyn phn Fully Connected Layer? ‘A. BO twong dng gita tap de liu hud luyén va de ligu pretrained B. kich thu tgp di igu hudn yen 56 lvong epochs cia mé hinh pretrained .Kign tric ia m6 hinh pretrained 46. Trong trurong hyp no chi nn husinluyén phn Fully Connected Layer (Training Head) thay vi Fine-Tuning toan b6 ms hinh? ‘A. Kh tap dt ligu downstream 6 kich thu6c nhé 8. Khi dt igu downstream cé 49 twong GOng cao vei dir iu pretrained . Kh c6 t8/ nguyen tinh togn han ch . Khi muén mé hinh thich nghi vot dir ligu mot hoan toan 447, Néu dirligu cia bal todn downstream cé kich thc lém nhumg rt khéc bigt so v6 dt liu lia m@ hinh pretrained, chign lege nao phi hep? A. Fine-Tuning tea bé m8 hinh 8. Chi hudin lyén phn Head cia mé hinh . Git nguyen trong s6 pretrained va sr dung Feature Extraction . Thém Batch Normalization 6 tranh overiting 448, Néu ban c6 tp di ligu nhd vi mun 4p dung Transfer Learning trén mgt mé hinh lon nhur LIT hoje ResNet,chién lunge nio glip gidm overfitting? [A Git nguyen phn Feature Extractor ila mé hin va chi huda luyn phn Classification Head 8. Sirdung Data Augmentation . Them Regularization nh Dropout hodc Weight Decay . Huai luydn lai toan bg m6 hinh tr Bu 449, Trong bil toén OCR, kj thut Augmentation no khong nén éurge ép dung?A Horizontal Flip 8. Colonitter . Rotation D. Translation 50. Trong bal todn nhgn din bin béo giao théng, kj thudt Augmentation no nén durge wu Hen? |A. Random Rotation 8. Horizontal Flip, ©. Color Jitter . Random Erasing 51. Khi hudin luygn m@ hinh phn foal anh y khoa ti anh Mi lending nén dp dung? key thugt Augmentation nb, [A Gaussian Noise 8. Random Rotation Horizontal Flip . Color ster 52. Trong bai toin nhgn dign khudn mit, thudt nao nén duge han ché sir dung? A Horizontal Flip 8. Random Crop ©. Color iter . Random Rotation 53, Trong bai toin phat hign 461 turgng (Object Detection), tal sa0 kf thut Random Erasing <6 thé itu ich? [[Link] m6 hinh hoc ich nhan dién vat thé ngay cd ki mot phn bj che Khust 8, Lam tng te €9 hun yen mo hin CC. Gip mé hinh b6 qua cé chi tét Khong quan trong . Tyo ra dt igu da dang hon54. Khi ndo nén dp dung ky thudt MixUp trong hudin luyén mé hinh thi ge may tinh? [A Khi mush ting kh nng t6ng quat héa ea mé hinh B. Ki di ibu e6 qua it mu hud luyen . Khi bi todn yu cBu elt nguyén b6 eye cba d6i twang trong dah D. Kh dt figu chira niu ohiGu va cin lam 18 d3c trang 55. Mot op Dense (Fully Connected) c6 128 du vio va 64 du ra. HOI s6 lucmg tham sé {parameters) cia lp ndy [a bao niu? Asia 8.8256 8320 6400 56. Mot bloc (kernel) 343 4p dung trén dnh €6 3 kénh mau (RGB). Hi sé lvong tham sé ‘trong b6 Ige ny La bao nhibu? ag a7 ©.30 p32 57. Khi Sp dung mgt convolution véi kernel 5x5 va stride = 1 lén anh du vdo o6 kich thud 3242253, kich thurde du ras [8 hao nigu néu khdng ding padding? A2B283 8, 285281 ©. 2612603 D. 26126x1 58. Khi str dung padding “same” trong convolution, kich thuéte au ra sé thé ndo so v6i du wae? ALNhé hom 8. len hon . Bing kich thude Bu vdo D. khong xéc éinh durge58. Mot lop Max Pooling 2x2 v6i stride 2 4p dung lén nh 32%32%3 s& tao ra dau ra c6 kich ‘thuréee bao nhiéu? A166 8.320323 6.300303 D. 15x153 60, iu gl xa ra nd learning rate qui cao? A.M hinh khéng hoi op 8B, MO hinh hgc nhanh ahumng khong 6n dinh , Mé hin hg tu dé nghiém t6i vu nhanh chéng . khéng anh huéng dé kt qué hoc 61, Ham kich hogt nao phé bién nhét trong CNN hién dai? A Sigmoid 8. Tanh Rel .Softmax (62. Tai sao Batch Normalization gidp tang tc di hun Iuyén mang CNN? [A Glém sy phy thude v8o kh t90 trong 58 8. Gir au ra cia mi layer trong mot phan phéi én dinh . Gidm overfitting [Link], 8, CdBu ding 163. Khi sir dung Dropout vr ty 160.5, diéu gl xdy ra trong qua trinh hudin luyén? [A M@tnira s6 no-ron trong layer sé bi vo higu héa ng nhién M8 hinh bj gidm kich thuse [Link] 6 tham s6 eta mé hinh . Ting 66 lveng no-ron due kich host164, Néu mot mang cb kn tric CNN - Flatten -> Dense (512) - Dense (10), $8 hrgng tham $6 trong layer Dense cubi cing i bao nhiéu? Asi20 8.5130 51200 D.51210 65. Trong backpropagation ela CNN, gradient durgc tinh todn nhu thé nio? [A Sirdung dao ham chia him mét mat wl @Bu ro 8, Sir dyng dao ham cia ham Kich hoat . Sir dung chubi dao ham thdng qua cic layers rude ds . Tét o ede phuong dn trén (66, Khi sir dung kf thugt Xavier Initialization, trong s6 han du durge chon da trén el? ‘A. Phan phi ngdu nhien trong khodng (0,1) 8. Phan phdi Gaussian vér trung binh 0, phuong sai phy thuge vo sS ugng no-ron ©. Gi ti e6 inh [Link] s6 bling 0 667. Néu mt lép Convolution sir dung kernel 3:3, stride 1 va padding 1, kich thurde dau ra ‘thay déi thé ndo so v6% dau vio? A. Kring d6i 8. Gidm 2-den vi [Link] met nt D. Ting wip dot {68. Tal sao str dung nhiéu convolution layer thay vi chi mt layer c6 kernel Ion? [A Glim s6 lugng tham s5 B. Tang khd nding hoc dc trerng phir tapDB ity hom kh hus ay [Link], B,C abu ding 169, Phan dogn dn (Image Segmentation) Ia gl? [A Chia nh thanh cde ving dya trén mau sée 8. Gin nhan time phe! trong nh wer mot lap cu thé C. Nhan dign ede d6i tugng riéng bit trong anh . Thue hién phan tich biéu a pixel 70. Gid sir mde dc trung CNN cé kich thu de (H.W,C}=(7,7,2088). Dc trung nty éuxge Flatten ‘va di qua mt FC layer c6 4096 neurons, tiép theo lé mt FC layer c6 1000 neurons @& phn loai. Kkhi chuyén ee FC layers thanh Conv layers, b6 Ige F cia hal lp Conv twang ding la bao nhieu? Aza B14 C77 baa 71. Tai sa0 CNN hoat dong t6t trong Computer Vision? [[Link] dng tinh eye b9 cba nh 8. Gidm 6 lueng tham s6 so véi mang Fully Connected CHoe duge dc trung quan trong tir anh . khong bi anh hizing bai 46 nhibu cis ot ib 72. Tap hun luyén (training set) vit tp kiém tra (test set) c6 val trd gi A Training set 68 husin luyén m6 hinh, test set €é én gié m6 hinh 8. CS hat du chi dng 48 husin lyn . Test set glip m6 hinh ty didu chinh trong s6 . Khang o6 sy Khde bigt gta chang73. Mgt convolutional ayer c6 128 fiers, mdi ite hich thutée 33, 6p dung ren du vo ¢6 64 channels. Hoi 6 lrqng tham s6 trong convolution layer nay (cura tinh bias) la boo itu? ans2 8.73728 18432 . 36864 174, Tai sa0 Data Augmentation quan trong trong Deep Learning? A. Gidp mérrOng tap d0 fou hudn uyén B. Gidm overftting CTaing t8¢ 66 hun luygn md hinh 1. Ca hig €6 tng quat cia mé hinh 75. Mang nao éugc sir dung phé bién trong Face Recognition? A FaceNet B. OpenPose ©. DeepFace [Link] 76, Mt mang CNN €6 5 convolution layers lidn iép, mai layer sir dung kere! 343, stride = 2, Padding = 1. Néu nh du vio cé kich thu 2562256, kich thie du ra sau 5 layers la bao nhieu? 2883288, 8.128128 ©, 2560256 0.1616 ‘77. Batch Normalization 66 téc dung gl? [A Tang t6c qué tinh hun luyén 8. Gidm internal covariate shift . Loaibé noise trong dt eu . Ting 49 phic tap cla md hinh‘78, M&t mang CNN cé du vio kich thé 323203, sir dung mét Convolution layer véi 64 ‘ters (kernel 5x5, stride 1, padding 2). H&i kich thube dau ra? 203064 B. 3203264 202864 D. 16x16x64 79. Trong qua trinh backpropagation cia mgt mang neural, néu learning rate qué lon, éitu gi 6 thE nay ra? A. Mé hinh hoi tu nhanh hen ma khong gap vin 2 gi . Mé hinh ed thé dao dng manh hodc khéng hoity C, M6 hinh fuén hi ty én gi tr 161 uu tod cue .Mé hinh 58 laén mc IS underiting 180. Gia sir mgt mang no-ron ¢6 3 lép: W1, W2, W3 .Trong qus trinh backpropagation, néu ‘gradient cia W3 rt nhd, nhurng gradient cua Wi gin nhu bing 0, ma hinh e6 thé gp vin d@ ar A. Vanishing Gradient 8. Exploding Gradient . Overfitting . Dropout qué mite