(Prepare maximum Questions possible, covering all areas of the modules

assigned )

1 (a) Illustrate the limitation of a single layer perceptron with an example (3) CO I K2 T S

(b) Specify the advantages of ReLU over sigmoid activation function. (3) CO I K2 T A

(c) Explain back propagation algorithm for neural network training. (9) CO I K2 T A

(d) How does bias and variance trade-off affect machine learning algorithms? (5) CO I K2 T A

2 (a) Explain the basic concept of a perceptron. (3) CO I K3 T A

(b) Explain the importance of choosing the right step size in neural networks. (3) CO I K2 T S

(c) With an example classification problem, explain the following terms: (8) CO I K2 T D

a) Hyper parameters b) Training set c) Validation sets d) Bias e)

Variance

3 (d) Compare overfitting and underfitting. How can it affect model (6) CO I K4 T A

(a) generalization ? Explain the concept of loss functions in neural networks. (3) CO I K2 T A

(b) Explain the concept of overfitting in neural network training. (3) CO I K2 T A

(c) Calculate the output of the following neuron Y if the activation function is a binary (4) CO I K3 P A
sigmoid

IQAC
 (d) Draw the architecture of a multi-layer perceptron. Derive update rules for (10) CO I K3 T A
parameters in the multi-layer neural network through the gradient descent

4 (a) Given a perceptron with two input features (x1 and x2), weights (w1 and w2), and (3) CO I K3 P
bias (b), calculate the output of the perceptron for a specific input (x1 = 0.5, x2 =
0.8, w1 = 0.4, w2 = 0.6, b = -0.2) using the step (sign) activation function.

(b) Explain the significance of validation sets in training neural networks. (3) CO I K2 T S

(c) Suppose you have a perceptron with two input features, x1 and x2, and weights w1 (4) CO I K3 P A
and w2. The perceptron uses a step function as its activation function. If the
perceptron's decision boundary is defined by w1x1+w2x2+b=0, where b is the bias
term, answer the following:
a) If w1=2, w2=−3, and b=1, what output would the perceptron produce for the
input (x1,x2)=(3,4)?
b) What does the value of the bias term (b) represent in the context of the
perceptron's decision boundary?

(d) Consider a simple feed forward neural network with a single hidden layer, as shown (10) CO I K3 P D
below: Input Layer (2 neurons) --> Hidden Layer (3 neurons) --> Output Layer (1
neuron).Perform one iteration of the back propagation algorithm for the given
sample to update the weights and biases using the sigmoid activation function and
mean squared error (MSE) loss. Given w1 = 0.3, w2 = 0.5, w3 = 0.2, w4 = -0.4,
 IQAC
5 (a) w5= -0.1, w6 = 0.2, w7=0.4,w8 = -0.3, w9 = -0.2, learning rate (α) of (3) CO I K4 T A

(b) (3) CO I K4 T A

(c) (6) CO I K2 T S

(d) (8) CO I K2 T S
0.1.

You have a dataset with 100 samples for a binary classification task. Your model
achieves 100% accuracy on the training data but performs poorly on a validation
set. Explain why this might be happening and suggest potential solutions.

Compare and contrast the activation functions: sigmoid and ReLU.

Explain the differences between hard tanh and soft tanh activation functions.
Provide examples of scenarios where each might be preferred.

Describe the architecture of a deep feedforward neural network (deep feedforward

network). What distinguishes it from a shallow neural network?

MODULE II 1* 2* 3* 4*
 IQAC
20 Marks Questions Marks Course Knowled Theory(T)/ Difficul
Outcome ge Level ty
Problem( Level
Each question can have maximum four sub division P)/ (S/A/D/T)
Design(D
(Prepare maximum Questions possible, covering all areas of the modules assigned) )

1 (a) Explain the importance of proper weight initialization in neural networks. (3) CO II K2 T S

(b) Describe the vanishing and exploding gradient problems in deep neural networks. (4) CO II K2 T A
How do these problems affect training, and what are the consequences for network
convergence?

(c) Discuss various optimization techniques used in training neural networks, (6) CO II K2 T A

(d) Compare and contrast the adaptive optimization methods AdaGrad, RMSProp, and (7) CO II K4 T D
Adam. Explain the principles behind these optimization algorithms, including how
they adapt learning rates during training. Discuss scenarios where each of these
methods is most effective, and highlight any potential issues or drawbacks.

2 (a) Define regularization in the context of neural networks. (3) CO II K2 T S

(b) Explain the purpose of L1 and L2 regularization techniques. How do they prevent (4) CO II K2 T S
overfitting, and what is the key difference between them?

(c) Explore the advantages of dataset augmentation in deep learning. Provide examples (6) CO II K2 T S
of common data augmentation techniques and explain how they can improve model
performance and generalization.

(d) Discuss parameter tying and sharing as a regularization technique. Explain when (7) CO II K2 T A
and why parameter tying and sharing can be beneficial. Provide examples of
architectures where this technique is commonly applied and its impact on model
complexity.
 IQAC
3 (a) Explain the fundamental idea behind ensemble learning and how it can improve (3) CO II K2 T S
model performance

(b) Explain the motivation behind batch normalization and how it addresses common (4) CO II K2 T S
training challenges.

(c) Describe various types of ensemble methods and its importance. (6) CO II K2 T A

(d) Initializing the weights of a neural network with very small or large random (7) CO II K2 T D
numbers is not advisable. Justify.

4 (a) Explain how dropout layers work and why they are effective in preventing (3) CO II K2 T S
overfitting.

(b) Explain how bagging and boosting differ in terms of their approaches to combining (4) CO II K4 T S
multiple base models.

(c) Discuss the advantages of dropout and batch normalization when applied together (5) CO II K2 T A
in a neural network architecture.

(d) Differentiate gradient descent with and without momentum. Give equations for (8) CO II K2 T A
weight updation in GD with and without momentum. Illustrate plateaus, saddle
points and slowly varying gradients.

5 (a) Derive weight updating rule in gradient descent when the error function is a) mean (3) CO II K2 T A
squared error b) cross entropy

(b) List any three methods to prevent overfitting in neural networks (6) CO II K2 T S

(c) Compare and contrast Kaiming (He) initialization and Xavier initialization methods. (4) CO II K2 T S

(d) Explain how L2 regularization improves the performance of deep feed forward (7) CO II K2 T A
neural networks.
 MODULE III 1* 2* 3* 4*

IQAC

20 Marks Questions Marks Course Knowled Theory(T)/ Difficul

Outcome ge Level ty
Problem( Level
Each question can have maximum four sub division P)/ (S/A/D/T)
Design(D
(Prepare maximum Questions possible, covering all areas of the modules )
assigned )

1 (a) Give two benefits of using convolutional layers instead of fully connected ones for (3) CO III K2 T S
visual tasks.

(b) Illustrate the strengths and weaknesses of convolutional neural networks. (3) CO III K2 T A

(c) Consider an activation volume of size 13×13×64 and a filter of size 3×3×64. (6) CO III K2 P D
Discuss whether it is possible to perform convolutions with strides 2, 3 and 5.
Justify your answer in each case.

2 (d) Suppose that a CNN was trained to classify images into different categories. It (8) CO III K2 T A
performed well on a validation set that was taken from the same source as the
trainingset but not on a testing set. What could be the problem with the training of
such a CNN? How will you ascertain the problem? How can those problems be
solved?
(a) (3) CO III K2 T A
What happens if the stride of the convolutional layer increases? What can be the
maximum stride? Justify your answer
(b) (3) CO III K2 T S
Describe the convolution operation in CNNs.

(c) What is the motivation behind convolution neural networks? (4) CO III K2 T A
 (d) Explain the following convolution functions a)tensors b) kernel flipping c) down (10) CO III K2 T S
sampling d) strides e) zero padding.

IQAC
3 (a) Describe common pooling methods like max-pooling and average pooling. Discuss (3) CO IV K2 T S
how pooling layers contribute to reducing spatial dimensions while preserving
essential features.

(b) Explain the architecture of Convolutional Neural Networks (CNNs). How do they (6) CO III K2 T A
differ from traditional fully connected neural networks, and what are the key
components that make CNNs suitable for tasks like image recognition?

(c) How do structured output models differ from traditional CNNs with single scalar (4) CO III K4 T S
outputs?

(d) Explain different data types commonly used in CNNs, such as grayscale images, (7) CO IV K2 T A
RGB images, and multi-channel inputs. Discuss the impact of data types on the
design and performance of CNN architectures.

4 (a) Explain variants of convolution functions used in CNNs. (3) CO III K2 T A

(b) Describe some key applications of Convolutional Neural Networks (CNNs) beyond (3) CO IV K2 T A
image classification.

(c) Take one pre-trained CNN architecture (e.g., VGGNet-19) and describe its key (7) CO III K2 T A
features, including the architecture's depth, use of convolutional layers, and pooling
layers. Explain the advantages of using pre-trained models as a starting point for
new projects.

(d) Illustrate the transfer learning process using a pre-trained CNN architecture. (7) CO III K2 T S
Provide a step-by-step guide on how to adapt and fine-tune a pre-trained model for
a new task, including data preparation, architecture modification, and training.

5 (a) Discuss the significance of pre-trained CNN architectures in deep learning. (3) CO III K2 T S
(b) Explain the concept of structured outputs in CNNs and how it applies to tasks like (3) CO III K2 T S
image segmentation and object detection.

IQAC
(c) Explain efficient convolution algorithms, their principles and how they optimize the (6) CO III K2 T S
convolution operation for deep networks.

(d) Briefly explain pre-trained convolutional architectures: AlexNet, and ResNet-50. (8) CO III K2 T S
Describe the key characteristics of each architecture, including the number of
layers, the use of convolutional and pooling layers, and any notable architectural
innovations.

MODULE IV 1* 2* 3* 4*

20 Marks Questions Course Knowled Theory(T)/ Difficul

Marks Outcome ge Level ty
Problem( Level
P)/ (S/A/D/
Each question can have maximum four sub division Design(D T)
)
(Prepare maximum Questions possible, covering all areas of the modules assigned )
1 (a) Explain the concept of computational graphs in the context of Recurrent Neural (3) CO IV K2 T S
Networks (RNNs).

(b) How does a recursive neural network work? (3) CO IV K2 T S

The vanishing gradient problem is more pronounced in RNN than in traditional
(c) neural networks. Give reason. Discuss a solution for the problem. (6) CO IV K2 T S
If we have a recurrent neural network (RNN), we can view it as a different type
(d) of network by "unrolling it through time". Briefly explain what that means. (8) CO IV K2 T D

2 (a) List the differences between LSTM and GRU (3) CO IV K1 T S

IQAC
(b) Explain the architecture of GRU. (3) CO IV K2 T S

(c) Show the steps involved in an LSTM to predict stock prices. Give one advantage (7) CO IV K2 T A
of using an RNN rather than a convolution network.

(d) Explain the architectures of Long Short-Term Memory (LSTM) and Gated Recurrent (7) CO IV K2 T A
Unit (GRU). Compare and contrast LSTM and GRU in terms of their design, memory
capabilities, and advantages.

3 (a) Explain the significance of BERT in the field of natural language processing (NLP). (3) CO IV K2 T S

(b) Discuss the key considerations in designing an RNN architecture. (3) CO IV K2 T S

(c) Describe the encoder-decoder sequence-to-sequence architecture. Explain how it is (6) CO IV K2 T A

used for tasks like machine translation or text summarization.

(d) Illustrate the concept of language modeling using an RNN. Describe the structure of an (8) CO IV K2 T S
RNN-based language model and how it generates probabilistic predictions for
sequences of words.

4 (a) Explain the challenges associated with training Recurrent Neural Networks (RNNs) (3) CO IV K2 T D
 (c)
(b) Describe techniques used to address the vanishing gradient problem in RNNs. (4) CO IV K2 T S

Explore advanced optimization algorithms suitable for training RNNs, such as Adam (6) CO IV K2 T A
and RMSProp.

(d) Discuss regularization techniques for RNNs, including dropout and weight tying. (6) CO IV K2 T S
Explain how these techniques can prevent overfitting and improve generalization.

5 (a) How does deep RNN architecture differ from traditional, shallow RNNs in terms of (3) CO IV K2 T S
network depth and capabilities?

IQAC
(b) Describe the architecture of a sentiment analysis model based on RNNs. Explain how (4) CO IV K2 T S
such a model can be designed to classify social media text into positive, negative, or
neutral sentiments.

(c) Explain the concept of bidirectional RNNs (Bi-RNNs). How can they be used to (6) CO IV K2 T S
capture both past and future context in sequential data?

(d) Compare and contrast the architectural differences between deep recurrent neural (7) CO IV K2 T S
networks (DRNNs) and convolutional neural networks (CNNs). Explain the strengths
and weaknesses of each architecture for sequence-based and image-based tasks.
 MODULE V 1* 2* 3* 4*

20 Marks Questions Course Knowled Theory(T)/ Difficul

Marks Outcom ge Level ty
e Problem( Level
P)/ (S/A/D/
Each question can have maximum four sub division Design(D T)
)
(Prepare maximum Questions possible, covering all areas of the modules assigned ) 1 (a)Generative
K2
Adversarial Networks(GANs) include a generator and a discriminator. (10)
CO V
Sketch a basic GAN using those elements, a source of real images, and a source of T D
randomness. K2
CO V
(b)The word “adversarial” in the acronym for GANs suggests a two-player game. What are the T D

two players, and what are their respective goals? (4)

2 (a) How does the variational auto-encoder(VAE) architecture allow it to generate new data (3) CO V K2 T S
points, compared to auto-encoder, which cannot generate new data points?

(b) List the difference between Boltzmann Machine and Deep Belief Network. (3) CO V K2 T D

IQAC
(c) Define Generative Adversarial Networks (GANs) and explain their fundamental (6) CO V K2 T S
structure, including the roles of the generator and discriminator networks. Highlight
the primary objective of GANs in generating realistic data.

(d) Discuss how autoencoders can be used for tasks such as text generation, (8) CO V K2 T A
summarization, and language translation.

3 (a) Is an autoencoder for supervised learning or for unsupervised learning? Explain briefly. (3) CO V K2 T S
4 (b) Define Deep Belief Networks (DBNs) and their significance in the field of deep (3) CO V K2 T S
learning.

(c) Explain auto encoder with an example. (7) CO V K2 T D

(d) Explain Generative Adversarial Networks using suitable diagram. (7) CO V K2 T A

(a) How do autoencoders encode and decode data? (3) CO V K2 T S

(b) Describe the structure and working principles of a variational autoencoder (VAE). (4) CO V K2 T S

(c) Explain the unsupervised pre-training and fine-tuning process of DBNs. (6) CO V K2 T S

5 (a) Explain how generative models differ from discriminative models. (3) CO V K2 T S

(b) Describe the architecture and working principles of Boltzmann Machines. (3) CO V K2 T S

(c) Discuss the challenges and limitations associated with Generative Adversarial (8) CO V K2 T A
Networks (GANs), such as mode collapse, training instability, and the generation of
biased or offensive content.

IQAC
(d) Describe the training process of GANs, including the adversarial training loop and the (6) CO V K2 T A
minimax game between the generator and discriminator.