i,I.
This question paper contains 4 printed pages]
Roll No,
S. No. ofQuestionpaper : 7461
Unique Paper Code : 2342203602
Name of the Paper : Deep Learning
Name of the Course : BA (P) DSC
Semester :VI
Duration : 3 Hours Maximum Marks : 90
(Write your Ron No. on the top inmediatel,y on recei,pt of th,is question pcLper.)
Question No. 1 (Section A) is compulsory.
Attempt any foz4r questions from Section 8.
Parts of a question should be attempted together.
Use of Scientific Calculator is allowed.
Section-A
1. (cB) Differentiate betwe'en Perceptron and a Multilayer Feedforward Neural
Network. 3
(a) Given a 5x5 input matrix and a 3x3 filter Ofemel) with a stride 6f 1 and
no padding. What is the size of the featur`e map, that will you get after
one convolution operation? Show computation. 3
(c) What are the challenges in neural networks optimization ? 4
(cZ) What is an autoencoder and wh`at are its applications ? 3
(e) Why is dropout. used during the training of neural networks ? 3
cO Discuss any fzuo applications of deep neural networks. 4
P.T.O.
•(2)
® Give o7®e word for the following :
(i) RNN variant designed to remember long-term dependencies.
(££) Normalization method used between layers to speed up training.
(j££) The operation that extracts features from images.
(ju) The technique that reduces the spatial size of feature maps.
(u) The problem that occurs when gradient becomes too small.
(I/i) A key strategy for reducing the cost of inference in a machine
learning model and to replace the original expensive model with
a smaller model that requires less memory and runtime to store
•and evaluate.
(UZZ) This parameter in a [Link] network determines the strength of the
connection between two neurons.
(uZZ£) An optimization algorithm which updates weights in the direction
that reduces the loss function.
(ir) Optimization algorithm that combines momentum and adaptive
learning rates.
a) The process that involves adjusting parameters like learning rate
and batch size to improve a model's performance.
sectioniB
2. (cb) Describe the structure of a feed forward neural network with appropriate
diagram. 7
`®
(I;) How does the backpropagation algorithm work ? Also, explain the chain
rule of calculus with respect to backpropogation. 8
(3)
3. (cb) Discuss Adam and RMSprop optimization techniques.
(a) What is the role of activation functions in neural netivorks.
Write the activation function for the following :
(j) Sigmoid
(Zj) ReLU
(Zzz) Linear
Also compute the output on the above activation functions when input
is x = 2.
4. (c» Perform a 2D convolution on the following input using the given kemel,
stride = 2 and no padding. 7
FliterAIemal : 1 0
-1 1
Input Matrix :
1 2 0 1 3 3
4 1 1 0 2 2
2 3 2 1 0 0
1 0 1 3 2 2
2 1 0 1 1 1
3 2 1 4 1 1
(a) On the above obtained matrix from part (cb), apply Rel.U activation
function and max pooling (2x2) with stride = 1. 5
(c) Which regularization method leads to weight sparsity ? Explain why ?
3
5. (cz) Consider a simple feedforward neural network with : 10
• One input layer with two inpdis : #i = 0.5, %2 = 0.4
[Link].
(4) 7461
• One hidden layer with two neurons (far and fa2), both using the
ReLU activation function.
• One output layer with one neuron using a linear activation.
The weights and biases are as follows :
• Weights from input t6 hidden layer :
• Neuron hi : ZUii = 0.2, ZUi2 = 0.4, bias = 0.1
• Neuron fa2 : zu2i = -0.3, zu22 = 0.1, bias = 0.2
• Weights from hidden to output layer :
• Zuoi = 0.6, ZU,02 = 0.5, bias = 0.3
Draw the neural network structure for the above and perform a .forward
pass through the network and .compute the final output.
Assuming the target value as 1.5, compute the Mean Squared Error
(MSE).
(a) Define convolution and pooling in the context of CNNs. 5
6. (c}) Explain architecture and working of Recurrent Neural Networks. 10
(a) Give fz#o benefits of using convolutional layers instead of fully connect6d
ones for visual tasks. 5
7. Write short notes any flue of the following : 15
(ci) Globaal Contrast Normalization
(a) Overfitting
(a) Stochastic Gradient Descent
(cZ) Bidirectional RNN
(e) LSTM
@ Graphics processing unit
® Mean Absolute Error.
7461