INDIAN INSTITUTE OF TECHNOLOGY 4 KKHARAGPUR ‘Stamp / Signature of the invigilator EXAMINATION ( End Semester) SEMESTER ( Spring) Section, | Namo s|s[o]o] 1/0] subjecttiame DEEP LEARNING aaiomeaoe ranean Violation of any of the above instructions may lead to severe punishment. » Do not leave the Examinaton Hall without submiting your answer script to the Inviglator. In any cago, you are not Important instructions and Guidelines for Students, ‘You must occupy your seat as per the Examination Schadula/Sitng Plan. Do not keep mobile phones or any similar electronic gadgets with you oven inthe snitched off mode. {Loose papers, ciass notes, books or any such materials must nat be in your possession, even If they ae relevant tothe subject you are taking examination, Data book, codes, graph papers, relevant standard tablestcharts or any othor matoriss are allowed only when intructed by the paper-setter Use of instrument box, pened box and non-progiammatia calculator is allowed during the examination. However, ‘xchange ofthese tems or any other papers (Including question papers) is nat permits. ‘Write on both sides ofthe ansier script and do oot tear off any page. Use last page(s) of the answer scrip for rough ‘work. Report tothe invglator ifthe answer scspt has tom er distorted page(s) Iki your responsi 1o ensure that you have signed the Atendance Sheet. Keep your Admit Cardildenty Card an the ‘sk for checking by helavigator. ‘You may leave the examination hall for wash room or for crnking water fora very shart period. Record your absence from the Examination Hall inthe register provided. Smoking and the consumption of any kind of beverages are srtly Pechsted inside the Examination Hal allowed to take away the answer seript with you. afer Ihe complation ofthe examination, do not leave the seat until the ingilators collec al the anewer scrips, During the examination, etter inside or outside the Examination Hall, gathering information from any kiad at coureas or exchanging information with others or any such attompt willbe treated as ‘unfair means’. Do not adopt unlair means and co not indulge in unseemly behav. ‘Signature of the Student To be tilled in by the examiner em Ae ace ce awe a li Marks Obtained | | | | | Marks obtained (in words) ‘Signature of the Examine ‘Signature of the ScrutineerINDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR End-Spring Semester Examination 2022-2023 Date of Examination: 20/4/23. Session:AN Duration: 8 hours Full Marks: 100 Subject No CS60010 Deep Learning Department: Computer Science & Engineering Instructions: Answer the questions in the baxes provided in this answer booklet only. You ‘may use supplementary sheets for rough work which must be tied to this booklet. 1. Consider the logistic regression model r-e(Ensrs) at (a) Write down an appropriate loss function including a L2 repularization term for m instances in the training set. (b) Determine the gradiont descent update rules for the regularized cost function above: Wy be (4)‘css0010 Endterm - Page 2 of 17 2023 cso Bint = Page 2 of 17] (4) (©) You have created a neural network and you find that your training error andl cost are high and your validation cost and error are sliost equal to it, Answer the following questions: i, What does this mean for your model? i, What actions would you take? (3) (@) You design a fully connected neural network architecture where all activations are sig You initialize the weights with large positive numbers. Ts this a good idea? Explain your answer. 2. Consider the following 1-dimensional ConvNet, where all variables are scalars: 2 = sla)? > hg Gon) B= war + w902+ 190 + a . ay =relu(21,0} 2 =relu( 2,0} a3 =relu(z,0} a1 shizy + kame + hyty +b ar 4 cow GN 2a ahisa + bys + kyra +b{2) List the parameters in the network, aL OL ob, ob (0) Determine the folowinss J Fos Fae O84 Fn as Guy aL 2 Se aL. Bins aL Buy (©) Given the gradients of the loss L with respect to the second layer activations g, derive the gradient of the loss with respect to the first layer activations 2. More precisely, given Oa; at aL, ab determine the following: 22, 2% ana 2, ermine te folving: 2, 9E ana BA a Ba | aL, | 2 Pe | ob oes a @ @)esoo010 Endterm - Page 4 of 17 2023 (4) (8) Givon the gradients of the loss I with respect to the fist Inyer activations 2, derive the gradient of the loss with respect to the convolution filter k- More precisely, given aL aL Gant Gaon % icseo010 Endterm - Page 8 of 17 2023 3. Consider the following RNN, which takes a sequence of numbers as input, makes a numerical prodiction at the last time step. ‘The hidden units (/(¢)) use the following activation function oe ‘The only trainable parameters are w; and wy and they are both initialized to 1. 2eo(2)~1 a(t) =wyh(t — 1) bwax(0) AG) = ED) (0) Wee te oma for de dave SH an fant of 2 fort P (b) Suppose initially ws Consider an input instance (1,1,1,1) of length TP =4,.0., (8) (1) = 2(2) = (8) = 2(4) = 1 and the corresponding target output is y = 0. Assume that the learning rate is = 0.5. Show how 1, is updated when you do stochastic gradient descent on this instance, ‘You may use the following approximate numerical values if required: 078, o(1.46) =081, 0(2.2) 0.46, 9{1.46) = 0.62, (1.62 0.885, (1.67) 067, $(1.67)eseooi0 Bator = Page Sof 27 ng 4. Consider a basi transformer architecture where both te encoder annd the decoder we stack of selfattention modules, Astame 2 attention heads, The transformer is applied to the task of English Hindi machine translation and daring traning, one instance of the input sentence and output translation areas follows. Inputs “A bind came to a forest stream as it was thirsty" Output: “pyass lngane par ek pakshee jangal kee dhaara mein aay” Treat each word as a separate token, 30 that the input soquence contains 11 items, and the tpn sequence contains 10 items. (6) (a) Consider the Sith ontput word “pakshee". Describe how the representation ofthis output word is computed. List the key vectors used, query vectors used, and value vectors use Horr is each of these key /query/ralue vectors used computed?6860010 Endterm - Page 9 of 17 2023 (b) To produce the input and output shown, how many key, query, and value vectors are computed in total by the encoder? How many by the decoder in total (not just for the 5th output word)? Fill in the following table and also explain your answer. Vector Type | Key Query Value Eneoder Decoder rr ©)cas0010. _Endterm-Page 1of 17 a (9) 5. (@) Brietly outline how selésupervised learning is used in BERT to learn the representatc of text, that i, how is the model trained? (9) (b) Briefly outline » method to set up selfsupervised learning to learn features for compute Vision task using a Ingge collection of raw images.‘c860010 Endterm - Page 11 of 17 2028 6. (a) Write down the loss function for the Variational Auto-Encoder and the loss function for a simple auto-encader. {(b) Let d be the latent embedding dimension, What is the output of the VAE encoder? basic autoencoder cannot be used for generating new objects? promi ices ma cee eet @) (3) (@)cs60010 Endterm - Page 12 of 17 2 (4) __(d) How docs the VAB architecture allow it to generate new data points?cssoa0 _Endterm - Page 18 of 17 on 7. Consider training a GAN with generator G(z) and discriminator D(G{z2)). The perfect dis. criminator outputs 1 on a real instance, 0 on a fake instance. ‘The figure above shows the training loses for two generator loss functions. Each point in tho plot shows the value of D(G(zi)), J1(G) and J2(G), For m generated points, # = 1,....72, JUG) and J2(G) are defined below. XY wet) XY ma ~ Dex) 20 : DG) ' (a) Barly in the training is the value of D(G(x)) closer to 0 or closer to 1? Explain why. ®) (b) Which of the two cost functions would you choose to train your GAN? Justify your answer. (3)(3) (¢) A GAN is successfully trained when D(G(z)) is close to 1. True or False? Explain y©360010, Endterm Page 16 of 17 2028 8. (a) Draw a block diagram of the reverse diffusion process clearly indicating T stops and spec (3) ‘fying the input and output of every step. | | (b) What is the input to the neural network used in the reverse process and what is the (3) output?2) 4) cseoo10 Bndterm - Page 16 of 17 x (6) What is the motivation bebi Intent diffusion model working ou the latent space? (@) How can the stable diffusion model be used for generating an image from a text?cs60010 Endterm- Page 17 of 17 2023 9. You have a large collection of grayscale images of frames extracted from old black and white (8) movies and also a large collection of frames from new movies, Suggest a method to convert a grayscale image to a color image utilizing this data.