MIT Art Design and Technology University

MIT School of Computing, Pune

Department of Computer Science and
Engineering M

Theory Assignments
Subject - Deep Learning & Neural
Networks (21BTCS031)
Class - L.Y. Core (SEM-I), <1,2,3>

Name of the Course Coordinator

Prof. Dr. Anant Kaulage
Team Members
1. Dr. Sunita Parinam
2. Dr. Mayura Shelke
3. Dr. Aditya Pai

A.Y. 2024 - 2025

 MIT

Theory Assignm
(Unit-I & Unit
 MIT Art Design and Technology University
MIT School of Computing, Pune
Department of Computer Science and Engineering
Last Year B. Tech
Academic Year 2024-2025. (SEM-I)
Subject: Deep Learning & Neural Networks (21BTCS031)

Theory Assignment-1
MIT
Date: 06/08 / 2024 Marks – 10

Q. 1 a) The linear perceptron can not solve XOR problem..

<L1/CO1>
Justify your answer with a suitable example.
b) How does Relu help in dealing with the vanishing <L1/CO1>
gradient problem? Demonstrate the significance of
Sigmoid and Relu functions with suitable applications.
c) Perceptron learning is a special case of SGD. True or <L1/CO1>
False. Explain with reasons
Q. 2 a) We simplify our Mnist computation by assuming our <L1/CO1>
“image" has two binary-valued pixels, 0 and 1, there
are no bias parameters, and we are performing a binary
classification problem. (a) Compute the forward-pass
logits and probabilities when the pixel values are [0,1],
and the weights are:
.2 -.3
-.1 .4
Here w[i; j] is the weight on the connection between
the ith pixel and the jth unit. E.g., w[0, 1] here is -.3
(b) Assume the correct answer is 1 (not 0) and use a
learning rate of 0.1. What is the loss? Also, compute
Δw0,0 on the backward pass.
Q. 3 a) Discuss the following challenges motivation deep <L1/CO2>
learning in detail:
● The curse of dimensionality
● Local constancy and smoothness
regularization
● Manifold learning

b) Differentiate between the weight update rule for linear <L1/CO2>

neurons and sigmoidal neurons. Derive the expression
for a sigmoidal neuron for squared error.
Q. 4 a) Discuss and derive a backpropagation algorithm for <L1/CO2>
the feed-forward neural network.
b) Discuss the following terms in detail: <L1/CO2>
a) Capacity. b) Underfitting. c) Overfitting.

MIT

Theory Assignm
(Unit-III & Uni
 MIT Art Design and Technology University
MIT School of Computing, Pune
Department of Computer Science and Engineering
Last Year B. Tech
Academic Year 2024-2025. (SEM-I)
Subject: Deep Learning & Neural Networks (21BTCS031)

Theory Assignment-2
MIT

Date: / / 2024 Marks – 10

Q. 1 a) State the number of neurons required in the input layer of a

<L1/CO3>
multilayered neural network for an image of size 124x124x3.
b) What is the difference between a filter and a kernel? <L1/CO3>
c) a. If the size of the input image is 24x24x1, kernel size is 3x3, <L1/CO3>
stride =1, and padding = valid. What is the size of the output
feature map?

b. If the size of the input image is 24x24x1 and kernel size is

3x3, stride =1, padding = same.
What is the size of the output feature map in this case?
d) Given Sequential model for classification of the input image <L1/CO3>
32 by 32 by 3:
model = Sequential ()
model.add (Conv2D(32, kernel_size=(3,3), activation='relu',
input_shape=input_shape))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(64,kernel_size=(3,3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(128,kernel_size=(3,3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
 model.add(Dense(256, activation='relu'))
model.add(Dense(128, activation='relu'))
model.add(Dense(no_classes, activation='softmax'))
Compute the number of trainable parameters of each layer in
the given model.
e) The input size is 24x24x3, kernel size is 3x3, stride =1, <L1/CO3>
padding = valid, number of filters = 5. What is the size of the
output feature map along with its depth?
Q. 2 a) What do you understand by sparsity of connections? With an <L1/CO3>
example explain how it is different from dense layers used in
multilayer network.
b) What is translation invariance? Explain the concept of <L1/CO3>
translation invariance using MNIST-10 dataset.
c) Briefly summarize the difference between VGG16, <L1/CO3>
RESNET50 and Inception networks.
Q. 3 a) What are Sequence Learning Problems? How are they <L1/CO4>
different from that of Feed Forward Neural Network and
CNNs? MIT
b) Give the architecture of a Bidirectional Neural Network and <L1/CO4>
state why a hidden layer is added in the backward direction.
d) What is Deep RNN? State a few applications of it. How is the <L1/CO4>
architecture of Deep RNN different from that of a
Bidirectional Neural network?
e) Explain the concept of the Encoder-Decoder model with the <L1/CO4>
help of a speech translation application.
Q. 4 a) How is the training phase of the Decoder different from that of <L1/CO4>
the testing phase of the Decoder in the Encoder-Decoder
model?
b) Explain the parts of an LSTM cell. State how each state is <L1/CO4>
different from each other?
Q. 5 a) Name the activation functions used in an LSTM cell. How are <L1/CO4>
they different from each other?
b) How are GRU different from LSTM? <L1/CO4>
c) What do you understand by context vector and explain its role <L1/CO4>
in an Encoder-Decoder model?
 MIT

Theory Assign
Unit-V
 MIT
MIT Art Design and Technology University
MIT School of Computing, Pune
Department of Computer Science and Engineering
Last Year B. Tech
Academic Year 2024-2025. (SEM-I)
Subject: Deep Learning & Neural Networks (21BTCS031)

Theory Assignment-3

Date: 14/11 / 2024 Marks – 5

Q. 1 Explain the concept of autoencoders and their

<L1/CO5>
applications in deep learning.
Q. 2 Explore different regularization techniques (e.g., L1 or <L1/CO5>
L2 regularization, dropout) for autoencoders and their
impact on model performance.
Q. 3 Investigate the influence of layer size and depth on the <L1/CO5>
representational power of autoencoders.
Q. 4 Explore denoising autoencoders and their ability to <L1/CO5>
reconstruct clean data from noisy inputs.
Q. 5 Explore deep generative models such as generative <L1/CO5>
adversarial networks (GANs) or variational
autoencoders (VAEs).