Reg.

No:

CHENNAI INSTITUTE OF TECHNOLOGY

Sarathy Nagar, Pudupedu, Chennai– 600 069.
Model Exam I - JUN 2023
Date Max. Marks 100 Marks
Subject IF4071 Deep Learning Time 3 hrs
Branch Computer Science and Engineering – Set A Year/Semester II/III
Course Objectives
The student should be made to
1. Develop and Train Deep Neural Networks.
2. Develop a CNN, R-CNN, Fast R-CNN, Faster-R-CNN, Mask-RCNN for detection and
3. recognition
4. Build and train RNNs, work with NLP and Word Embeddings
5. The internal structure of LSTM and GRU and the differences between them
6. The Auto Encoders for Image Processing

Course Outcomes
CO1: Features extraction from image and video data
CO2: Implement image segmentation and instance segmentation in image
CO3: Implement image recognition and image classification using pretrained network
CO4: Traffice information analysis using twitter data
CO5:Autoencoder for classification and feature extraction
Blooms Taxonomy (BT) Level:
K1-Remembering; K2-Understanding; K3-Applying; K4-Analyzing; K5-Evaluating; K6-Creating

PART-A 10 X 2 = 20 CO BT Year of
(Answer all the questions) Mapping level Univ. Exam
1. Define a vector and give an example CO410.1 K1 Nov/Dec 2016

A vector is a quantity that has both magnitude and direction.

It can be represented geometrically as a directed line
segment. Vectors can be used to represent a wide variety of
physical quantities, such as velocity, force, and acceleration.

Example:

A force of 100 Newtons acting in the north direction can be

represented by the vector (100, 0).

2. Define time series data CO410.1 K3 Apr/May 2017

Time series data, also referred to as time-stamped

data, is a sequence of data points indexed in time
order.

These data points typically consist of successive

measurements made from the same source over a
 fixed time interval and are used to track change
over time. Examples of time series data include stock
prices, weather data, and sensor data.

Example:

The daily closing prices of a stock over the past year can be
represented as a time series.

3. Name a loss function used for binary classification problems. CO410.2 K2 Nov/Dec 2018

A popular loss function used for binary classification

problems is the binary cross-entropy loss function. It is
defined as follows:

loss = -y * log(p) - (1 - y) * log(1 - p)

where y is the true label of the data point and p is the

predicted probability of the data point belonging to the
positive class.

Example:

If we are trying to classify emails as spam or not spam, we

could use the binary cross-entropy loss function to train our
model.

4. Define overfitting in the context of neural networks CO410.2 K1 Apr/May 2017

Overfitting is a phenomenon that occurs when a neural

network learns the training data too well and is unable to
generalize to new data. This can happen if the neural
network is too large or if the training data is too small.

Example:

If we train a neural network to classify images of cats and

dogs using a very small dataset, the network may overfit the
training data and be unable to correctly classify new images
of cats and dogs.

5. Explain the concept of Linear Time Invariant (LTI) systems CO410.3 K2 Apr/May 2019

Linear time invariant (LTI) systems are systems that satisfy

the following two properties:

● Linearity: The output of the system is a linear function

of the input.
 ● Time invariance: The output of the system does not
depend on the time at which the input is applied.

Many physical systems, such as electrical circuits and

mechanical systems, are LTI systems.

Example:

An electrical circuit that consists of a resistor and a capacitor

is an LTI system. The output voltage of the circuit is a linear
function of the input voltage, and the output voltage does not
depend on the time at which the input voltage is applied.

6. Describe the role of dense layers in a CNN. CO410.3 K4 Apr/May 2019

Dense layers, also known as fully connected layers, are

used in CNNs to combine the features extracted by the
convolutional layers. Dense layers are typically used at the
end of a CNN to perform a classification or regression task.

For example, a CNN for image classification might have the

following architecture:

● A convolutional layer that extracts features such as

edges and corners from the input image.
● A pooling layer that downsamples the feature maps
from the convolutional layer.
● A second convolutional layer that extracts more
complex features from the downsampled feature
maps.
● A second pooling layer that downsamples the feature
maps from the second convolutional layer.
● A dense layer that combines the features from the
second pooling layer and makes a prediction about
the class of the object in the image.

Dense layers are able to combine the features extracted by

the convolutional layers in a nonlinear way. This allows the
CNN to learn complex relationships between the features
and to make accurate predictions.

7. What is NLP, and what are its key applications? CO410.4 K2 Apr/May 2017

NLP stands for natural language processing. It is a field of

computer science that deals with the interaction between
computers and human language. NLP has a wide range of
applications, including machine translation, text
summarization, sentiment analysis, and question answering.
 Here are some examples of NLP applications:

● Machine translation: Translating text from one

language to another.
● Text summarization: Generating a shorter version of
a text while preserving the main points.
● Sentiment analysis: Identifying the sentiment of a
text, such as whether it is positive, negative, or
neutral.
● Question answering: Answering questions posed in
natural language.
● Speech recognition: Converting speech to text.
● Text generation: Generating text, such as news
articles, poems, or code.

NLP is a rapidly growing field, and new applications for NLP

are being developed all the time.

8. What is Part of Speech (PoS) tagging in NLP, and why is it CO410.4 K2 Nov/Dec 2016
important?

Part of speech (PoS) tagging is the task of assigning a part

of speech tag to each word in a sentence. PoS tags are
used to indicate the grammatical role of a word in a
sentence. For example, the word "run" can be a verb or a
noun, depending on the context.

PoS tagging is an important task in NLP, as it is used for

many other NLP tasks, such as parsing and dependency
parsing. Parsing is the task of constructing a syntactic tree
for a sentence, and dependency parsing is the task of
identifying the syntactic dependencies between words in a
sentence.

PoS tagging is also used in many other NLP applications,

such as machine translation, text summarization, and
sentiment analysis.

9. How do Generative Adversarial Networks (GANs) consist of CO410.5 K1 Apr/May 2018

a generator and a discriminator?

GANs consist of two neural networks: a generator and a

discriminator. The generator is trained to generate fake
data, while the discriminator is trained to distinguish
between fake data and real data. The generator and the
discriminator are trained simultaneously, in a game-theoretic
setting.

The generator takes a random noise vector as input and

produces a fake data sample. The discriminator takes a data
 sample as input and predicts whether it is real or fake.

The generator is trying to fool the discriminator into thinking

that its fake data samples are real. The discriminator is
trying to correctly identify all real data samples and all fake
data samples.

As the generator and the discriminator are trained, they both

improve. The generator becomes better at generating fake
data samples that are indistinguishable from real data
samples, and the discriminator becomes better at
distinguishing between real data samples and fake data
samples.

GANs can be used to generate a wide variety of data, such

as images, text, and audio. GANs are also used in many
other applications, such as image super-resolution, image
inpainting, and style transfer.

10. What are the two main components of the Actor-Critic CO410.5 K1 Nov/Dec 2016
algorithm?

The two main components of the Actor-Critic algorithm are

the actor and the critic. The actor is responsible for selecting
actions, while the critic is responsible for evaluating the
actions selected by the actor. The actor and the critic are
trained simultaneously, using reinforcement learning.

The actor takes the state of the environment as input and

produces an action. The critic takes the state of the
environment and the action selected by the actor as input
and produces a value. The value is a measure of how good
the action selected by the actor is.

The actor is trained to maximize the expected return, which

is the sum of the future rewards that the actor expects to
receive. The critic is trained to minimize the error between
the value it predicts and the actual return.

As the actor and the critic are trained, they both improve.
The actor learns to select actions that lead to high rewards,
and the critic learns to accurately evaluate the actions

Part- B 5x13=65 marks CO BT Year of Marks

(Answer all the questions) Mapping level Univ. Exam Allotted
11.a(i) Discuss the challenges and techniques associated CO410.1 K1 Nov/Dec2016 13
with handling time series data in deep learning
applications.

(or)
What are the fundamental principles of deep CO410.1 K1 Nov/Dec2016 13
 b learning, and how do they differ from traditional
machine learning approaches?

12.a(i) Can you explain the fundamental building blocks CO410.2 K2 Nov/Dec2017 13
of a neural network, including the input layer,
hidden layers, and output layer? How do these
components work together

(or)
b(i) Delve into the role of hyperparameters in neural CO410.2 K4 Apr/May2019 13
network training. Discuss commonly tuned
hyperparameters like learning rate, batch size, and
the number of hidden layers

13.a(i) Compare and contrast popular optimizers like SGD, CO410.3 K2 Nov/Dec2018 13
Adam, and RMSprop with respect to their
suitability for training CNNs.

(or)
b(i) Define feature maps and filters in the context of CO410.3 K2 Nov/Dec2017 13
CNNs, emphasizing their significance in feature
extraction.

14.a(i) Compare and contrast the Continuous Bag of CO410.4 K3 Apr/May2019 13

Words (CBOW) and Skip-Gram models for word
embedding, including their use cases and training
processes

(or)
14b(i) Explore the concept of global cooccurrence CO410.4 K2 Nov/Dec2018 13
statistics–based word vectors and how they are
generated in NLP

15.a(i) Describe the Deep Q-Network (DQN) and its CO410.5 K4 Apr/May2019 13
significance in reinforcement learning, addressing
the challenges that deep learning techniques
address in DRL.
(or)
15b(i) Describe the key components and working CO410.5 K1 Apr/May2019 13
principles of Generative Adversarial Networks
(GANs), highlighting their use in image generation
and manipulation.

PART- C (1*15 =15 marks) CO BT Year of Marks

Mapping level Univ. Exam Allotted
16.a(i) Describe the concept of Sparse Autoencoders, CO410.1 K4 Apr/May2019 15
highlighting how they encourage sparsity in
learned representations and their potential
 applications

(or)
16b(i) Discuss the principles and key features of YOLO CO410.2 K4 Nov/Dec2018 15
(You Only Look Once) as a real-time object
detection algorithm

Prepared by Verified by Approved by

Reg. No:

CHENNAI INSTITUTE OF TECHNOLOGY

Sarathy Nagar, Pudupedu, Chennai– 600 069.
Model Exam I - JUN 2023
Date Max. Marks 100 Marks
Subject IF4071 Deep Learning Time 3 hrs
Branch Computer Science and Engineering – Set B Year/Semester II/III
Course Objectives
The student should be made to
1. Develop and Train Deep Neural Networks.
2. Develop a CNN, R-CNN, Fast R-CNN, Faster-R-CNN, Mask-RCNN for detection and
3. recognition
4. Build and train RNNs, work with NLP and Word Embeddings
5. The internal structure of LSTM and GRU and the differences between them
6. The Auto Encoders for Image Processing
Course Outcomes
CO1: Features extraction from image and video data
CO2: Implement image segmentation and instance segmentation in image
CO3: Implement image recognition and image classification using pretrained network
CO4: Traffice information analysis using twitter data
CO5:Autoencoder for classification and feature extraction
Blooms Taxonomy (BT) Level:
K1-Remembering; K2-Understanding; K3-Applying; K4-Analyzing; K5-Evaluating; K6-Creating

PART-A 10 X 2 = 20 CO BT Year of
Mapping level Univ. Exam
(Answer all the questions)

1. How do you manipulate tensors in deep learning? CO410.1 K1 Nov/Dec 2016

2. Define perception learning algorithms. CO410.1 K3 Apr/May 2017

3. Name a loss function used for binary classification problems. CO410.2 K2 Nov/Dec 2018

4. Give an example of a commonly used activation function CO410.2 K1 Apr/May 2017

5. Explain the concept of Linear Time Invariant (LTI) systems CO410.3 K2 Apr/May 2019

6. Describe the role of dense layers in a CNN. CO410.3 K4 Apr/May 2019

7. What is NLP, and what are its key applications? CO410.4 K2 Apr/May 2017

8. What is Part of Speech (PoS) tagging in NLP, and why is it CO410.4 K2 Nov/Dec 2016
important?

9. Explain the Q-Learning algorithm and its use in CO410.5 K1 Apr/May 2018
reinforcement learning

10. Describe the denoising autoencoder and its applications. CO410.5 K1 Nov/Dec 2016

Part- B 5x13=65 marks CO BT Year of Marks

Mapping level Univ. Exam
(Answer all the questions) Allotted

11.a(i) What are the unique considerations and challenges CO410.1 K1 Nov/Dec2016 13
when working with video data in deep learning,
and how are they addressed

(or)

What are the fundamental principles of deep CO410.1 K1 Nov/Dec2016 13

learning, and how do they differ from traditional
b
machine learning approaches?

12.a(i) Can you explain the fundamental building blocks CO410.2 K2 Nov/Dec2017 13
of a neural network, including the input layer,
 hidden layers, and output layer? How do these
components work together

(or)

b(i) What is regularization in the context of neural CO410.2 K4 Apr/May2019 13

networks? Compare L1 and L2 regularization
techniques and their effects on preventing
overfitting

13.a(i) Compare and contrast popular optimizers like SGD, CO410.3 K2 Nov/Dec2018 13
Adam, and RMSprop with respect to their
suitability for training CNNs.

(or)

b(i) Define feature maps and filters in the context of CO410.3 K2 Nov/Dec2017 13
CNNs, emphasizing their significance in feature
extraction.

14.a(i) Compare and contrast the Continuous Bag of CO410.4 K3 Apr/May2019 13

Words (CBOW) and Skip-Gram models for word
embedding, including their use cases and training
processes

(or)

14b(i) Explore the concept of bi-directional LSTMs in CO410.4 K2 Nov/Dec2018 13

NLP, focusing on their dual processing of
sequences and applications in tasks like sentiment
analysis

15.a(i) Explain Q-Learning and how it is used to estimate CO410.5 K4 Apr/May2019 13

the quality of actions in reinforcement learning,
including the Q-table and Q-values.

(or)

15b(i) Describe the key components and working CO410.5 K1 Apr/May2019 13

principles of Generative Adversarial Networks
(GANs), highlighting their use in image generation
and manipulation.

PART- C (1*15 =15 marks) CO BT Year of Marks

Mapping level Univ. Exam
Allotted

16.a(i) Discuss the fundamental differences between the CO410.1 K4 Apr/May2019 15

Actor-Critic algorithm and other reinforcement
learning methods, such as value-based and policy-
based approaches.
 (or)

16b(i) Discuss the principles and key features of YOLO CO410.2 K4 Nov/Dec2018 15
(You Only Look Once) as a real-time object
detection algorithm

Prepared by Verified by Approved by