IT4653 – Deep Learning

Department of IT and ADS 2024-2025

IT4653 DEEP LEARNING
COURSE OUTCOMES
After the completion of this course, students will be able to:
C310.1 Understand the role of deep learning in machine learning applications
C310.2 Design and implement deep learning applications.
C310.3 Critically analyze different deep learning models in image related projects.
C310.4 Design and implement convolutional neural networks
C310.5 Know about applications of deep learning in NLP and image processing
Course

PSO

PSO
PSO
PO1

PO1

PSO
P01
PO

PO

PO

PO

PO

PO

PO

PO

PO
Outcome 1

9
0

2
3
4
C310.1 3 2 2 2 2 - - - - 2 3 3 3 1 2 1
C310.2 3 2 2 2 3 - - - - 2 3 3 3 1 2 1
C310.3 3 2 2 3 3 - - - - 2 2 3 3 3 3 2
C310.4 3 2 2 3 3 - - - - 2 2 3 3 3 3 2
C310.5 3 2 2 2 3 - - - - 2 3 3 3 1 2 1

MAPPING BETWEEN CO AND PO, PSO WITH CORRELATION LEVEL 1/2/3

RELATION BETWEEN COURSE CONTENT WITH COs
UNIT I BASICS OF NEURAL NETWORKS
Sl. Course Content Knowledge level
Books No. of hours
No Referred to be handled
1 Basic Concept of Neurons BL2 T1 1
2 Perception Algorithm BL2,B L1 T1 1
3 MultiLayer Perception Algorithm BL2, BL4 T1 1
4 Linear Classifiers BL2, BL4 T1 1
5 Gradient Descent BL2,B L4 T1 1
6 Types of Gradient Descent BL2,B L4 T1 1
7 Feed Forward BL2,B L4 T1 1
8 Back propagation Networks BL2,B L4 T1 1
9 Loss Function BL2,B L4 T1 1
UNIT II INTRODUCTION TO DEEP LEARNING
Sl. Course Content Knowledge Books No. of hours
No. level referred to be handled
1 Deep Feed-Forward Neural Networks BL2, BL1 T1 1
2 Gradient Descent- Back-Propagation BL2 T1 1
3 and Other Differentiation Algorithms BL2,BL1 T1 1
4 Vanishing Gradient Problem BL2 T1 1
5 Mitigation – Rectified Linear Unit(ReLU) BL2 T1 1
6 Heuristics for Avoiding Bad Local Minima BL2, BL1 T1 1
7 Regularization for Deep Learning BL2 T1 1
8 Adversial Training BL4 T1 1
9 Optimization for Training Deep Models. BL4 T1 1
UNIT III CONVOLUTIONAL NEURAL NETWORKS
Sl. Course Content Knowledge Books No. of hours to
No. level Referred be handled
1 CNN Architectures BL2, BL3 T1 1
2 Convolution Layers BL2, BL3 T1 1
3 Pooling Layers BL2, BL3 T1 1
4 Transfer Learning BL2, BL3 T1 1
5 Image Classification using Transfer BL2, BL3 T1 1
Learning
6 Recurrent and Recursive Nets BL2, BL3 T1 1

[Link]’s Institute of Technology

1
 IT4653 – Deep Learning
Department of IT and ADS 2024-2025
7 Recurrent Neural Networks BL2, BL1 T1 1
8 Deep Recurrent Networks B L4 T1 1
9 Recursive Neural Networks–Applications BL4 T1 1
UNIT IV UNSUPERVISED DEEP LEARNING
[Link]. Course Content Knowledge level Books Referred No. of hours to
be handled
1 Autoencoders BL2, BL1, BL3 T2 1
2 Autoencoders in unsupervised BL2 T2 1
learning
3 Parse Denoising BL, BL1 T2 1
4 Standard -Parse Denoising BL2 T2 1
5 Variational Autoencoders BL2, BL3 T2 1
6 Contractive Variational BL2, BL3 T2 1
Autoencoders
7 Adversarial Generative Networks BL1, BL3 T2 1
8 Boltzmann Machine. BL1, BL4 T2 1
9 Deep Boltzmann Machine (DBM). BL1, BL3 T2 1
UNIT V APPLICATIONS OF DEEP LEARNING
Sl. Knowledge Books No. of
No. Course Content
level Referred hours to
be handled
1 Images segmentation – Object Detection BL3, BL4 T2 1

2 Multi class Object Detection -Object BL2, BL3 T2 1

Classification and Localization
3 Automatic Image Captioning BL2, BL3, BL4 T2 1
4 Image generation with Generative BL2, BL3, BL4 T2 1
adversarial networks
5 Opinion Mining using Recurrent Neural BL2, BL3, BL4 T2 1
Networks
6 Parsing and Sentiment Analysis BL2, BL3, BL4 T2 1
7 Recursive Neural Networks BL2, BL3,BL4 T2 1
8 Sentence Classification using BL2, BL3,BL4 T2 1
Convolutional Neural Networks
9 Dialogue Generation with LSTMs BL2, BL3,BL4 T2 1
BL1- Remembering, BL2- Understanding, BL3 – Applying, BL4 –Analyzing, BL5 – Evaluating,
BL6 - Creating
TEXT BOOKS:
1. Ian Goodfellow, Yoshua Bengio, Aaron Courville, “Deep Learning”, MIT Press,2018
[Link] Chollet, “Deep Learning with Python”, Manning Publications,2018
REFERENCE BOOKS:
[Link] Kim, ”MATLAB Deep Learning: With Machine Learning, Neural Networks and Artificial
Intelligence”,Apress,2017.
2. Ragav Venkatesan, Baoin Li, Convolutional Neural Networks in Visual computing”, CRCPress,2018
[Link] Kumar Manaswi,”Deep Learning with Applications Using Python”,Apress,2018
[Link] [Link],”R Deep Learning Essentials”,Packt Publications,2016
5. Deng & Yu, Deep Learning: Methods and Applications, Now Publishers, 2013.
[Link]’s Institute of Technology
2
 IT4653 – Deep Learning
Department of IT and ADS 2024-2025
UNIT I BASICS OF NEURAL NETWORKS
Basic Concept of Neurons – Perceptron Algorithm - Linear Classifiers and Gradient Descent– Feed
Forward and Back propagation Networks.
UNIT-I/PART-A (C310.1)
1 What is a neuron in the context of deep learning, and how does it function? C310.1 BL1
2 What is the role of an activation function in a deep learning neuron? C310.1 BL1
3 How are weights and biases used in a deep learning neuron C310.1 BL1
4 How are weights and biases used in a deep learning neuron? C310.1 BL1
5 Define the term activation function C310.1 BL2
6 What a weight represents in an Artificial Neural Network and how it influences the output. C310.1 BL1
7 How bias affects the learning process. C310.1 BL1
8 What is the difference between a perceptron and a multi-layer neural network? C310.1 BL1
9 How does a deep learning neuron learn during the training process? C310.1 BL1
10 Define the Perceptron model in Artificial Neural Networks.(Mar 2022) C310.1 BL2
11 Mention key characteristic of Perceptron model. C310.1 BL1
12 Write the short notes on basic working principle of the Perceptron Algorithm. C310.1 BL1
13 What is the role of the learning rate in the Perceptron algorithm? C310.1 BL1
14 How you would choose an appropriate activation function for the output layer. C310.1 BL1
15 What is a linear classifier in machine learning? C310.1 BL1
16 How a linear classifier works in binary classification? C310.1 BL1
17 Outline the steps involved in training an Artificial Neural Network C310.1 BL2
18 What is Gradient Descent and how is it used in training machine learning models? C310.1 BL1
19 What is the role of the learning rate in the Gradient Descent algorithm? C310.1 BL1
20 What is Stochastic Gradient Descent?(Nov 2023) C310.1 BL1
21 How does Gradient Descent apply to a linear classifier during training? C310.1 BL1
22 What is a feed forward neural network? C310.1 BL1
23 Write the concept of a forward pass in a feed forward neural network. C310.1 BL1
24 What is back propagation in neural networks?(Nov 2023) C310.1 BL1
25 How does the back propagation algorithm update the weights in a neural network? C310.1 BL1
26 What is the role of the activation function in a feed forward neural network? C310.1 BL1
Unit – I / PART- B
1 Explain the concept of a neuron in the context of deep learning. Describe its components C310.1 BL3
and how it works.
2 Describe the Perceptron Algorithm and explain its working principle. Discuss its strengths C310.1 BL3
and limitations.
3 What are linear classifiers in machine learning? Explain how they work and provide C310.1 BL4
examples of their applications.
4 Explain the concept of Gradient Descent. How does it optimize the loss function during the C310.1 BL3
training of a neural network?
5 Compare and contrast the various types of gradient descent (batch, stochastic, and mini- C310.1 BL4
batch). Discuss their advantages and disadvantages.
6 Explain the process of training a neural network using feed forward propagation and back C310.1 BL4
propagation. How are the weights updated during training?
[Link]’s Institute of Technology
3
IT4653 – Deep Learning
Department of IT and ADS 2024-2025
7 Discuss the concept of activation functions in neural networks. Why are they important, and C310.1 BL3
what are the differences between popular activation functions like ReLU, Sigmoid, and
Tanh?
8 What is the role of the loss function in training a neural network? Explain how the loss C310.1 BL3
function is related to optimization and model performance.
9 Describe how a linear classifier like the Perceptron can be extended to multi-class C310.1 BL4
classification problems. Discuss the challenges and methods to handle them.
10 How does the concept of overfitting relate to neural networks? Discuss how regularization C310.1 BL3
techniques like L2 regularization, dropout, and early stopping can help mitigate over fitting.
UNIT II INTRODUCTION TO DEEP LEARNING
Deep Feed Forward Neural Networks – Gradient Descent – Back Propagation and other Differentiation
Algorithms- Vanishing Gradient Problem- Mitigation – Rectified Linear Unit(ReLU)- Heuristics for
Avoiding Bad Local Minima – Heuristics for Faster Training – Nestors Accelerated Gradient Descent –
Regularization for Deep Learning – Dropout – Adersial Training – Optimization for Training Deep models.
UNIT-II/PART-A(C310.2)
1 Define a deep feedforward neural network. C310.2 BL1
2 What is the main difference between a shallow and deep neural network? C310.2 BL1
3 Define gradient descent. C310.2 BL1
4 What is the purpose of gradient descent in neural networks? C310.2 BL1
5 What is backpropagation? C310.2 BL2
6 What is the vanishing gradient problem? C310.2 BL1
7 How does the vanishing gradient problem affect training in deep networks? C310.2 BL1
8 Name a technique used to mitigate the vanishing gradient problem. C310.2 BL1
9 Differentiate ReLU from the sigmoid activation function. C310.2 BL1
1 What is the mathematical function of ReLU? C310.2 BL1
0
1 Why is ReLU commonly used in deep neural networks? C310.2 BL1
1
1 List one heuristic used to avoid getting trapped in bad local minima. C310.2 BL2
2
1 How does random initialization help in avoiding bad local minima? C310.2 BL1
3
1 Why is batch normalization used in deep learning? C310.2 BL1
4
1 Define momentum in the context of gradient descent. C310.2 BL1
5
1 How does using a learning rate schedule improve training speed? C310.2 BL1
6
1 What is the purpose of Nesterov’s accelerated gradient descent? C310.2 BL2
7
1 Define regularization in deep learning. C310.2 BL1
8
1 Why is regularization important in training deep neural networks? C310.2 BL2
9
2 How does dropout help prevent overfitting? C310.2 BL2
0
2 What is dropout in neural networks? C310.2 BL1
1
[Link]’s Institute of Technology
4
 IT4653 – Deep Learning
Department of IT and ADS 2024-2025
2 Define adversarial training. C310.2 BL1
2
2 Describe the role of adversarial training in neural networks. C310.2 BL1
3
2 What is an adversarial example in the context of neural networks? C310.2 BL1
4
2 List two optimization algorithms commonly used for training deep models.s C310.2 BL2
5
2 What is Optimization for Training Deep models? C310.2
6
UNIT-II/PART-B C310.2
1 Discuss the advantages and potential drawbacks of very deep networks and analyze how C310.2 BL5
increasing the depth of a neural network impacts both its representational capacity and
computational complexity.
2 Explain how a neural network learns from data. Illustrate the steps involved in updating C310.2 BL6
weights and biases during training.
3 Illustrate the process of back propagation through an example. Explain each step involved in C310.2 BL4
calculating the gradients and updating weights.
4 Describe the limitations of back propagation, particularly with respect to vanishing C310.2 BL4
gradients. Suggest modifications or enhancements that could improve the stability of
gradient flow through deep layers.
5 Examine the role of activation functions in addressing the vanishing gradient problem. C310.2 BL5
Compare ReLU with other activation functions and discuss how ReLU improves the
gradient flow in deep networks.
6 Explain the operations of deep feedforward network with neat diagram.(Nov 2023) C310.2 BL3
7 Describe in detail about the impact of dropout and L2 regularization on over fitting in deep C310.2 BL3
networks
8 Describe a scenario where adversarial training could be essential and discuss how to C310.2 BL3
implement it.
9 Develop an optimization strategy using Nesterov’s accelerated gradient method. C310.2 BL4
10 Explain why optimization algorithms are critical for training deep networks. C310.2 BL3
UNIT III CONVOLUTIONAL NEURAL NETWORKS
CNN Architectures-Convolution- Pooling Layers-Transfer Learning-Image classification using Transfer
Learning – recursion and recursive Net-Recurrent Neural networks -Deep recurrent Networks- Recursive
Neural Networks- Applications
UNIT-III/PART-A(C310.3)
1 What is the primary purpose of Convolutional Neural Networks (CNNs) in deep learning? C310.3 BL1
2 Name two commonly used CNN architectures for image classification. C310.3 BL1
3 How does a CNN differ from a fully connected neural network? C310.3 BL1
4 What is the function of a convolutional layer in a CNN? C310.3 BL2
5 Define "filter" or "kernel" in the context of convolutional layers. C310.3 BL2
6 Explain the purpose of stride in a convolutional layer. C310.3 BL1
7 What is the purpose of a pooling layer in a CNN?(Nov 2023) C310.3 BL1
8 Differentiate between max pooling and average pooling. C310.3 BL2
9 How does pooling affect the size of feature maps in CNNs? C310.3 BL2
10 What is transfer learning in the context of deep learning? C310.3 BL1

[Link]’s Institute of Technology

5
IT4653 – Deep Learning
Department of IT and ADS 2024-2025
11 Why is transfer learning beneficial for tasks with limited labeled data? C310.3 BL1
12 Mention two popular pre-trained models commonly used for transfer learning in image classification.
C310.3 BL2
13 How does transfer learning improve accuracy in image classification? C310.3 BL1
14 What is fine-tuning in transfer learning, and how is it performed? C310.3 BL1
15 Why are the initial layers of pre-trained models often "frozen" during transfer learning? C310.3 BL1
16 Define recursion in the context of neural networks. C310.3 BL1
17 What differentiates recursive networks from recurrent networks? C310.3 BL2
18 How are recursive neural networks used in natural language processing (NLP)? C310.3 BL1
19 What is the main characteristic of Recurrent Neural Networks (RNNs)?(Nov 2023) C310.3 BL1
20 How do RNNs handle sequential data differently than CNNs? C310.3 BL2
21 What is padding, and why is it used in convolutional layers? C310.3 BL1
22 How do residual connections in ResNet help to improve training in deep networks? C310.3 BL1
23 What is the purpose of using deep recurrent networks instead of a single-layer RNN? C310.3 BL1
24 Explain how vanishing gradients affect the training of deep recurrent networks. C310.3 BL1
25 How do recursive neural networks process hierarchical structures? C310.3 BL2
26 Give one example of an application where recursive neural networks are useful. C310.3 BL1
Unit –III/PART- B
1 Discuss the architecture and working of a Convolutional Neural Network (CNN) with a C310.3 BL3
focus on the convolutional and pooling layers. Explain how these layers contribute to
feature extraction and dimensionality reduction in image classification.
2 Compare and contrast different CNN architectures such as VGGNet, ResNet, and Inception. C310.3 BL3
Describe their unique features, advantages, and limitations, and explain how these
architectures have advanced the field of deep learning for image classification.
3 Describe in detail about the different convolutional architectures (e.g., AlexNet, VGG, C310.3 BL4
ResNet) differ in handling feature extraction and classification, and what are the strengths
and limitations of each?
4 Analyze the role of pooling layers in CNNs, and how do techniques like max pooling, C310.3 BL3
average pooling, and global pooling affect feature maps and model performance?
5 How does transfer learning work in CNNs, particularly in image classification tasks, and C310.3 BL3
what are the trade-offs in using a pre-trained model vs. training a model from scratch?
6 In what ways does fine-tuning the deeper layers of a pre-trained model during transfer C310.3 BL4
learning impact the performance and specialization of a model for a specific task?
7 How does recursion operate within Recursive Neural Networks (RNNs), and in what types C310.3 BL3
of data structures (e.g., tree-like data) are they most effective?
8 Describe about the differences between traditional Recurrent Neural Networks (RNNs) and C310.3 BL4
deep recurrent networks, and how does depth impact their ability to learn long-term
dependencies?
9 Explain Recursive Neural Networks (RNNs) handle hierarchical data representations and C310.3 BL3
how it differently from other RNN architectures, and what are their primary applications?
10 Elaborate the memory cells, such as those in LSTM and GRU units, play in enhancing the C310.3 BL4
capabilities of recurrent neural networks?
UNIT IV- UNSUPERVISED DEEP LEARNING
Auto encoders-Standard -Sparse-De noising -Contractive – Variational Auto encoders- Adversarial
Generative Networks- Deep Boltzmann Machine (DBM)
UNIT-IV/PART-A(C310.4)

[Link]’s Institute of Technology

6
IT4653 – Deep Learning
Department of IT and ADS 2024-2025
1 What is an autoencoder, and how is it used in unsupervised learning? C310.4 BL1
2 What are the main components of a standard autoencoder? C310.4 BL1
3 How does a sparse autoencoder differ from a standard autoencoder? C310.4 BL1
4 What is regularized autoencoder? (Nov 2023) C310.4 BL1
5 What is the purpose of adding noise to the input of a de-noising autoencoder? C310.4 BL2
6 Define a contractive autoencoder and its main goal in feature learning. C310.4 BL1
7 How does a variational autoencoder (VAE) differ from a standard autoencoder? C310.4 BL1
8 What is the role of the latent space in a variational autoencoder? C310.4 BL1
9 Explain the purpose of the Kullback-Leibler divergence in VAEs. C310.4 BL1
10 What is meant by "reconstruction loss" in autoencoders? C310.4 BL1
11 Define a Stochastic encoder. (Nov 2023) C310.4 BL2
12 What is the purpose of the discriminator in a generative adversarial network (GAN)? C310.4 BL1
13 How do the generator and discriminator interact in a GAN? C310.4 BL1
14 What is mode collapse in GANs, and why is it a problem? C310.4 BL2
15 What is the main goal of a generative adversarial network (GAN)? C310.4 BL1
16 What is the vanishing gradient problem in the context of GANs? C310.4 BL1
17 What are deep Boltzmann machines (DBMs), and how do they differ from deep C310.4 BL2
belief networks?
18 What role do hidden layers play in a deep Boltzmann machine? C310.4 BL1
19What is the process of pre-training in DBMs. C310.4 BL1
20What is the purpose of Gibbs sampling in training a DBM? C310.4 BL2
21How is feature extraction achieved in a sparse autoencoder? C310.4 BL1
22Why are contractive autoencoders used for learning robust representations? C310.4 BL1
23What is the advantage of using de-noising autoencoders for image reconstruction? C310.4 BL1
24How does the regularization term in a contractive autoencoder impact its behavior? C310.4 BL1
25What is a Restricted Boltzmann Machine (RBM), and how does it relate to DBMs? C310.4 BL2
26What are the applications of variational autoencoders in anomaly detection? C310.4 BL1
Unit –IV/PART- B C310.4
1 Define and describe the role of each component in a standard autoencoder model. C310.4 BL3
2 Explain the differences between a sparse autoencoder and a contractive autoencoder, and C310.4 BL4
discuss how each type is designed to learn different aspects of data representations.
3 Given a dataset with noisy images, outline the steps to train a de-noising autoencoder to C310.4 BL3
restore the images. What parameters or settings would you prioritize?
4 Analyze the objective functions of the generator and discriminator in a GAN. How do they C310.4 BL4
work together, and what challenges might arise in optimizing these objectives concurrently?
5 Compare the strengths and weaknesses of deep Boltzmann machines (DBMs) versus C310.4 BL5
variational autoencoders (VAEs) in learning latent data representations. Under what
circumstances would each model be preferable?
6 Design a modified GAN architecture that could potentially reduce mode collapse. Justify C310.4 BL6
each of the architectural changes and explain how they address the problem.
7 Describe how a contractive autoencoder uses regularization to limit the sensitivity of the C310.4 BL2
encoder to small input changes, and discuss the effect on feature learning.
8 Suppose you want to use a variational autoencoder for anomaly detection in a C310.4 BL3
medical dataset. Outline the approach, including pre-processing, training, and
evaluation.
9 Compare and contrast the training methods of deep Boltzmann machines (DBMs) and C310.4 BL3
[Link]’s Institute of Technology
7
 IT4653 – Deep Learning
Department of IT and ADS 2024-2025
restricted Boltzmann machines (RBMs). How does the layer-wise pre-training in DBMs
enhance their learning capability?
10 Critically evaluate the usefulness of Wasserstein GANs (WGANs) in stabilizing GAN C310.4 BL5
training. What trade-offs might exist when using WGANs over traditional GANs?
UNIT V APPLICATIONS OF DEEP LEARNING
Images Segmentation-Object detection – Multi class object Detection – Object classification and
Localization – Automatic Image Captioning – Image generating with Generative adversarial networks –
Opinion Mining using Recurrent Neural Networks -Parsing and sentiment analysis using Recursive Neural
Networks -Sentence classification using Convolutional Neural networks -Dialogue Generation with LSTMs.
UNIT-V/PART-A(C310.5)
1 What is image segmentation, and how is it useful in computer vision applications? C310.5 BL1
2 Define object detection and list its main components. C310.5 BL1
3 What is multi-class object detection, and how does it differ from binary object detection? C310.5 BL1
4 What are the difference between object classification and object localization? C310.5 BL1
5 What is the purpose of automatic image captioning, and which neural networks are C310.5 BL2
commonly used for this task?
6 How do generative adversarial networks (GANs) generate new images? C310.5 BL1
7 What role does the generator play in a GAN model? C310.5 BL1
8 Describe the purpose of the discriminator in a GAN model. C310.5 BL1
9 What is opinion mining, and why are recurrent neural networks (RNNs) suitable for this C310.5 BL1
task?
10 What is sentiment analysis, and how can recursive neural networks (RecNNs) be used for C310.5 BL1
it?
11 How does sentence classification benefit from convolutional neural networks (CNNs)? C310.5 BL1
12 What is the role of Long Short-Term Memory (LSTM) networks in dialogue generation? C310.5 BL1
Nov.2023
13 What is the difference between semantic segmentation and instance segmentation in image C310.5 BL1
processing.
14 What is the Intersection over Union (IoU) metric, and why is it important in object C310.5 BL1
detection?
15 What are anchor boxes in object detection, and how do they help in identifying objects? C310.5 BL1
16 What is the process of region proposal in object detection frameworks like Faster R-CNN. C310.5 BL1
17 What is a fully convolutional network (FCN), and why is it effective for image C310.5 BL2
segmentation?
18 How does single-shot multi-box detection (SSD) differ from region-based approaches like C310.5 BL1
Faster R-CNN?
19 What are the key components of a YOLO (You Only Look Once) model in object detection? C310.5 BL1
20 How attention mechanisms improve automatic image captioning models. C310.5 BL2
21 What is a sequence-to-sequence model, and how is it used in image captioning? C310.5 BL1
22 Define adversarial training and its purpose in training GANs. C310.5 BL1
23 What are the primary challenges in training GANs for image generation? C310.5 BL1
24 What are the benefits of using RNNs over traditional machine learning methods for opinion C310.5 BL2
mining?
25 What is the concept of sentiment polarity in sentiment analysis? C310.5 BL1
26 How do recursive neural networks parse sentence structures for sentiment analysis? C310.5 BL1
Unit –V/PART- B C310.5
1 What are the primary tasks involved in object detection, and how do object detection models C310.5 BL3
like YOLO and Faster R-CNN perform these tasks?
[Link]’s Institute of Technology
8
 IT4653 – Deep Learning
Department of IT and ADS 2024-2025
2 Explain how semantic segmentation differs from instance segmentation. Provide examples C310.5 BL4
of applications where each type of segmentation would be most beneficial.
3 Given a dataset of labeled images, outline the steps and key considerations for implementing C310.5 BL3
a multi-class object detection model. Which techniques would you prioritize for effective
training?
4 Analyze the advantages and limitations of using attention mechanisms in automatic image C310.5 BL4
captioning models. How does attention improve caption accuracy?
5 Evaluate the effectiveness of LSTM-based models versus transformer-based models for C310.5 BL5
dialogue generation. Which approach would you recommend for generating natural-
sounding responses and why?
6 Design a GAN-based system for generating images of new product designs. Specify how C310.5 BL6
you would structure the generator and discriminator, and justify the choice of architecture
for each component.
7 Describe how CNNs are modified to handle sentence classification in NLP applications. C310.5 BL3
Discuss how these modifications help CNNs capture contextual information in text.
8 Consider a sentiment analysis application using RNNs. Identify key steps in C310.5 BL3
preprocessing text data for this model, and explain how different tokenization
methods might impact performance.
9 Compare the role of feature maps in CNNs for object classification versus their role in C310.5 BL4
semantic segmentation. What differences in processing are required to support these
different applications?
10 Critically evaluate the impact of pre-trained word embeddings on the performance of C310.5 BL5
sentiment analysis models. When might custom embeddings be more effective than pre-
12
trained ones?

[Link]’s Institute of Technology

9