Page 1 of 3
Ref. No.: Ex/UG/BE/ CSE/PE/B/T/413C/2021
B.C.S.E 4th YEAR 1st SEMESTER EXAM 2021
(non-contact online mode)
Subject-Deep Learning
Time: 3 hours Full Marks: 70
Read carefully the instructions written on the body of the email through which the question paper is
sent. Answer scripts as single file must be submitted via Email: jukamal2015@[Link]
Time: 3 hours Full Marks: 70
Answer Question No. 1 (Compulsory) and any FOUR questions from the rest
1. State whether the following statements are True or False. Give reasons in favour of your
answer .
a. Momentum can slow down the training convergence of an ANN model
b. Logistic Regression Model always gives the probability of being in the positive class
c. A perceptron is highly capable of classifying non-linear data
d. In the case of an autoencoder, the desired outputs are set to be equal to the inputs
e. In overcoming the gradient vanishing problem in deep neural networks, ReLu is
preferable to sigmoid activation function.
f. CNN is rotation invariant
g. Gating mechanism used in a LSTM network helps in learning long term dependencies
h. Dropout is used for handing the overfitting problem in a deep neural networks
i. In residual networks, skip connection is used for handling the underfitting problem
j. An ANN is powerful enough to approximate complex functions even if all nodes in it
are linear . 3 x 10 = 30 marks
2. a) If L2 regularization is used, show how the weight update rule is changed? What is the objective
of applying regularization?
𝛿𝛿𝛿𝛿
b) Consider the following computational graph and find using chain rule of calculus.
𝛿𝛿𝛿𝛿
g(y) p= g(y)
y= f(w)
o=K(p,q)
w K(p,q)
f(w)
𝑥𝑥
output
input weight
z= f(w) h(z) q= h(z)
Page 2 of 3
3. Consider a dataset in which two data points {(-1, -1), (1, 1)} belongs to one class , and the other
two points {(1,-1), (-1, 1)} belong to the other class . Start with parameter values of a perceptron
model at (0,0), and work out stochastic updates (incremental mode) with 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 ∶ 𝛼𝛼 = 1.
Show calculations upto 3 epochs and answer the following questions.
(a) Does the algorithm converge in the sense that the change in objective function becomes
extremely small over time ?
(b) Explain why the situation in (a) occurs. 10 marks
4. a) The each gate in a LSTM unit contains a sigmoid function-Why? Explain.
Mention two different learning tasks where LSTM can be useful.
b) Discuss briefly gating mechanism used in the LSTM network. What is its purpose?
5+5=10 marks
5. Consider that a Convolutonal Neural Network which have 3 convolutional layers with a
combination of average pooling and two fully connected layers is decided to be used for an image
classification task. Can you mention the name of a well known state-of-the art deep neural network
which has the similar architecture as mentioned above? This CNN model is reconfigured with the
parameters or hyperparameter values as shown in the following table, but some cells of the table
are left blank. Now fill in the blank cells of the following table with the appropriate values(show
necessary computation). You may leave any cell blank if you feel that the corresponding parameter
or hyperparameter is not relevant to the architecture of the proposed model.
Layer No of filters Filter size Stride Size of Activation
feature map Function
input 32 x 32 x 1
Conv1 6 6x 6 1 ReLu
Avg pooling1 3x3 13 x 13 x 6
Conv2 6 x6 1 8 x 8 x 16
Avg pooling 2 2 4 x 4 x 16
Conv3 120 4 x4 1 ReLu
Fully 84
connected 1
Fully 10 softmax
connected 2
10 marks
Page 3 of 3
6. a) Mention the different types of machine learning with an example for each type
b) Draw an architecture of a deep auto encoder. Mention the steps of layer-wise pre-training
algorithm. 5 + 5=10 marks
------------------------------
