DL Question Bank

Neural Network and Deep Learning (Anna University)

Scan to open on Studocu

Studocu is not sponsored or endorsed by any college or university

Downloaded by swati zindani ([email protected])
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024

CCS355 NEURAL NETWORKS AND DEEP LEARNING L T P C

2 0 2 3
COURSE OBJECTIVES:
• To understand the basics in deep neural networks
• To understand the basics of associative memory and unsupervised learning networks
• To apply CNN architectures of deep neural networks
• To analyze the key computations underlying deep learning, then use them to build and train deep neural
networks for various tasks.
• To apply autoencoders and generative models for suitable applications.
UNIT I INTRODUCTION 6
Neural Networks-Application Scope of Neural Networks-Artificial Neural Network: An Introduction- Evolution of
Neural Networks-Basic Models of Artificial Neural Network- Important Terminologies of ANNs-Supervised Learning
Network.
UNIT II ASSOCIATIVE MEMORY AND UNSUPERVISED LEARNING NETWORKS 6
Training Algorithms for Pattern Association-Autoassociative Memory Network-Heteroassociative Memory Network-
Bidirectional Associative Memory (BAM)-Hopfield Networks-Iterative Autoassociative Memory Networks-Temporal
Associative Memory Network-Fixed Weight Competitive Nets-Kohonen Self-Organizing Feature Maps-Learning
Vector Quantization-Counter propagation Networks-Adaptive Resonance Theory Network.
UNIT III THIRD-GENERATION NEURAL NETWORKS 6
Spiking Neural Networks-Convolutional Neural Networks-Deep Learning Neural Networks-Extreme Learning
Machine Model-Convolutional Neural Networks: The Convolution Operation – Motivation – Pooling – Variants of the
basic Convolution Function – Structured Outputs – Data Types – Efficient Convolution Algorithms – Neuroscientific
Basis – Applications: Computer Vision, Image Generation, Image Compression.
UNIT IV DEEP FEEDFORWARD NETWORKS 6
History of Deep Learning- A Probabilistic Theory of Deep Learning- Gradient Learning – Chain Rule and
Backpropagation - Regularization: Dataset Augmentation – Noise Robustness -Early Stopping, Bagging and Dropout -
batch normalization- VC Dimension and Neural Nets.

UNIT V RECURRENT NEURAL NETWORKS 6

Recurrent Neural Networks: Introduction – Recursive Neural Networks – Bidirectional RNNs – Deep Recurrent
Networks – Applications: Image Generation, Image Compression, Natural Language Processing. Complete Auto
encoder, Regularized Autoencoder, Stochastic Encoders and Decoders, Contractive Encoders.
Total Periods :30
TEXT BOOKS:
1. Ian Goodfellow, Yoshua Bengio, Aaron Courville, “Deep Learning”, MIT Press, 2016.
2. Francois Chollet, “Deep Learning with Python”, Second Edition, Manning Publications, 2021.

REFERENCES:
1. Aurélien Géron, “Hands-On Machine Learning with Scikit-Learn and TensorFlow”, Oreilly, 2018.
2. Josh Patterson, Adam Gibson, “Deep Learning: A Practitioner’s Approach”, O’Reilly Media, 2017.

St. Joseph’s Institute of Technology 1

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
COURSE OUTCOMES
Upon completion of the course, students will be able to:
CO1 Apply Convolution Neural Network for image processing.
CO2 Understand the basics of associative memory and unsupervised learning networks.
CO3 Apply CNN and its variants for suitable applications.
CO4 Analyze the key computations underlying deep learning and use them to build and train deep neural
networks for various tasks.
CO5 Apply autoencoders and generative models for suitable applications.

MAPPING BETWEEN CO AND PO, PSO WITH CORRELATION LEVEL1/2/3

COs POs PSOs

PO1 PO2 PO3 PO4 PO5 PO6 PO7 PO8 PO10 PO11 PO12 PSO1 PSO2 PSO3 PSO4

CO1 2 2 1 - 1 - 1 - - 1 2 1 1 1 1
CO2 2 2 1 - 1 - 1 - - 1 2 1 1 1 1
CO3 3 2 2 2 1 - 1 - - 1 2 1 1 1 1
CO4 3 2 2 2 1 - 1 - - 1 2 1 1 1 1
CO5 3 2 2 2 1 - 1 - - 1 2 1 1 1 1

RELATION BETWEEN COURSE CONTENT WITH COs

UNIT I INTRODUCTION
S.No Knowledge level Course Content Course Outcomes
1 R, U & Ap Neural Networks
2 U & Ap Application Scope of Neural Networks
3 An & Ap Artificial Neural Network
4 An & Ap ANN - An Introduction
5 An & Ap Evolution of Neural Networks CO1
6 U & Ap Basic Models of Artificial Neural Network

7 U & Ap Important Terminologies of ANNs

8 U&C Supervised Learning Network

9 U&C Supervised Learning Network

UNIT II ASSOCIATIVE MEMORY AND UNSUPERVISED LEARNING NETWORKS

S.No Knowledge level Course Content Course Outcomes

1 U & Ap Training Algorithms for Pattern Association

Autoassociative Memory Network ,
2 U & Ap
Heteroassociative Memory Network
3 R, U & Ap Bidirectional Associative Memory (BAM) CO2
Hopfield Networks,
4 U & Ap
Iterative Autoassociative Memory Networks
5 An & Ap Temporal Associative Memory Network

St. Joseph’s Institute of Technology 2

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
6 U & Ap Fixed Weight Competitive Nets
Kohonen Self-Organizing Feature Maps,
7 U & An
Learning Vector Quantization
Counter propagation Networks,
8 U & An
Adaptive Resonance Theory Network.
UNIT III THIRD-GENERATION NEURAL NETWORKS
S.No Knowledge level Course Content Course Outcomes
1 U & An Spiking Neural Networks, Convolutional Neural Networks

2 An & Ap Deep Learning Neural Networks

3 U&C Extreme Learning Machine Model

4 U&C Convolutional Neural Networks: The Convolution Operation

5 U&C Motivation, Pooling

Variants of the basic Convolution Function CO3
6 U & An
7 An & Ap Structured Outputs, Data Types
8 U & An Efficient Convolution Algorithms, Neuroscientific Basis
Applications: Computer Vision, Image Generation, Image
9 An & Ap
Compression.
UNIT IV DEEP FEEDFORWARD NETWORKS
S.No Knowledge level Course Content Course Outcomes
1 U&C A Probabilistic Theory of Deep Learning

2 U, An & Ap Gradient Learning

3 U&C Chain Rule and Backpropagation

4 U & Ap Regularization: Dataset Augmentation

5 U & An Noise Robustness CO4
6 U & An Early Stopping,

7 U & Ap Bagging and Dropout

8 U&C batch normalization-

9 U, An & Ap VC Dimension and Neural Nets.

UNIT V RECURRENT NEURAL NETWORKS

S.No Knowledge level Course Content Course Outcomes
1 U, An & Ap Recursive Neural Networks
2 U&E Bidirectional RNNs
3 U&E Deep Recurrent Networks
4 U&E Applications: Image Generation, Image Compression,
5 U&E Natural Language Processing. CO5
6 U, An & Ap Complete Auto encoder,
7 U&E Regularized Autoencoder,
8 U&E Stochastic Encoders and Decoders,
9 U&E Contractive Encoders.
R – Remember; Ap – Apply; An – Analyze; U – Understand, E- Evaluate , C-Create

St. Joseph’s Institute of Technology 3

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
UNIT I INTRODUCTION
Neural Networks-Application Scope of Neural Networks-Artificial Neural Network: An Introduction- Evolution
of Neural Networks-Basic Models of Artificial Neural Network- Important Terminologies of ANNs-Supervised
Learning Network.
UNIT-I / PART-A
1 Define a neural network and provide an example of a real-world application.
A neural network is a computational model inspired by the human brain, consisting of
interconnected nodes organized into layers, used for tasks like pattern recognition and BL1
prediction. Example Application: Convolutional Neural Networks (CNNs) are used for image
classification, such as identifying animals in photos or detecting objects in autonomous vehicles.
2 Explain the purpose of an activation function in a neural network and how it affects the
output.
The purpose of an activation function in a neural network is to introduce nonlinearity
into the model. It enables the network to learn complex relationships in the data by allowing
neurons to activate (fire) based on weighted inputs. Activation functions determine whether a BL2
neuron should be active or not, influencing the output by applying a mathematical
transformation to the weighted sum of inputs. Without activation functions, the entire neural
network would behave as a linear function, limiting its capacity to represent intricate patterns
and relationships in the data.
3 Given the problem of hand writing recognition, describe how a neural network can be
structured (e.g., number of layers, nodes) to address it effectively.
Input Layer: An input layer with 784 nodes (28x28 pixels for each digit image). Each node
represents the intensity of a pixel. BL3
Hidden Layer: A single hidden layer with 128 nodes. This layer helps capture intermediate
features in the images.
Output Layer: An output layer with 10 nodes, each representing a digit (0-9).
4 Compare and contrast feed forward neural networks and recurrent neural networks,
highlighting their respective advantages and use cases.
FNN Advantages: RNN Advantages:
Simplicity. Sequence modeling.
Suitable for static data. Handles variable-length input/output.
Efficient parallel processing. Effective for temporal data. BL4

Use Case: Use Case:

Image classification. Natural language
Structured data analysis. processing. Time series
analysis.
5 How a Neural Network can be used in natural language processing tasks, such as sentiment
analysis or language translation.
Neural networks, including RNNs and CNNs, process text data, tokenize it, and use BL2
architectures suited for sequence modeling or feature extraction. Training on labeled data allows
them to predict sentiment in new text.
6 Compare and contrast the advantages and limitations of using Convolutional Neural
Networks (CNNs) versus Recurrent Neural Networks (RNNs)
CNN RNN
Advantages:Ideal for spatial data like images. Effective with sequential data.
Efficient parallel processing. Handles variable-length inputs BL3
Limitation: Sequential processing can be slow.
Ineffective with sequential data. Vulnerable to vanishing/exploding gradients.
Requires fixed input size.
7 Define an Artificial Neural Network(JAN/2021)
A computational model composed of interconnected nodes, organized into layers,
designed for pattern recognition and data processing tasks. BL1
For instance, it can recognize handwriting in digit recognition tasks or analyze text sentiment in
natural language processing.
8 List the basic components of Artificial Neural Network(JAN/2021)
 Neurons (Nodes): Processing units. BL1
 Layers: Organized neurons.

St. Joseph’s Institute of Technology 4

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
 Weights: Connection strengths.
 Activation Functions: Compute neuron outputs.
 Bias: Shift activation functions.
 Input/Output Data: Information flow.
 Loss Function: Measures prediction errors.
 Backpropagation: Adjusts weights.
 Training Data: Labeled dataset.
 Optimization Algorithm: Minimizes loss.
9 Define the term "loss function" in the context of neural network training
A mathematical function that measures the discrepancy between the predicted output of a
BL1
neural network and the actual target values during training, serving as a guide for adjusting the
network's parameters.
10 What is the concept of overfitting in neural network training
Overfitting in Neural Network Training: Occurs when a neural network learns the training
data too well, capturing noise and irrelevant details, leading to poor generalization on unseen
BL2
data.
it may memorize specific features of individual images, such as unique patterns or noise in the
training set.
11 Give an example of a commonly used activation function. (DEC/2022) (JAN/2022)
ReLU (Rectified Linear Unit): ReLU is a popular activation function used in neural
networks. It computes the output as the maximum of zero and the input value, which means that
BL2
if the input is positive, it remains unchanged, but if it's negative, it becomes zero. This simple yet
effective function is widely used in hidden layers of deep neural networks for tasks like image
classification and natural language processing.
12 Outline the steps involved in training an Artificial Neural Network(DEC/2022)
 Data Preparation
 Initialize Weights
 Forward Propagation BL1
 Loss Calculation
 Backpropagation
 Gradient Descent
13 Compare and contrast supervised learning and unsupervised learning in the context of
Artificial Neural Networks.
Targeted Training: In supervised learning, neural networks are trained using labeled data,
where both input and corresponding output (target) are provided during training. BL3
Unsupervised Learning in Artificial Neural Networks:
Unlabeled Data: Unsupervised learning involves training neural networks with unlabeled data,
meaning that there are no explicit target labels provided during training.
14 How you would choose an appropriate activation function for the output layer.
The activation function should match the desired range of output values. For example, if
you need outputs between 0 and 1, sigmoid is suitable, whereas softmax ensures that the outputs BL2
represent probabilities that sum to 1.
15 Define the Perceptron model in Artificial Neural Networks.
The Perceptron is a simple artificial neural network model consisting of a single layer of
binary classifiers. It takes multiple input features, assigns weights to them, computes a weighted BL1
sum, and passes it through an activation function (usually a step function) to produce a binary
output (0 or 1).
16 Mention key characteristic of Perceptron model
It’s a single-layer neural network used for binary classification. It takes input features,
BL2
assigns weights, computes a weighted sum, applies an activation function (typically a step
function), and produces a binary output.
17 What is the concept of a Multilayer Perceptron (MLP) in Artificial Neural Networks? How
does it differ from a single-layer Perceptron?
An MLP is a type of artificial neural network with multiple layers of neurons, including BL3
an input layer, one or more hidden layers, and an output layer. Unlike a single-layer Perceptron,
MLPs can model complex nonlinear relationships in data.
18 List type of basic Artificial Neural Network model
ANN models include Perceptrons, Multilayer Perceptrons (MLPs), Radial Basis BL1
Function Networks (RBFNs), and Kohonen Self-Organizing Maps (SOMs).

St. Joseph’s Institute of Technology 5

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
19 Define the term activation function. (DEC/2022)
An activation function is a mathematical function used in artificial neural networks to
introduce nonlinearity into the model. It determines whether a neuron should fire (activate) based
on its weighted inputs and biases.
 Tanh BL1
 ReLU
 Sigmoid
 softmax
20 Explain what a "weight" represents in an Artificial Neural Network and how it influences
the output.
In an artificial neural network, a weight represents the strength of the connection BL1
between two neurons. It signifies how much influence the output of one neuron has on the input
of another, affecting the signal's magnitude and direction during information propagation.
21 Describe the role of the bias in an Artificial Neural Network (JAN/2021)
Bias in an artificial neural network is an additional parameter added to each neuron. It
helps shift the activation function's output, allowing the network to model more complex BL1
relationships by introducing an offset.
Ex:y=mx+b,where b is the bias.
22 How bias affects the learning process. (April/2021)
Bias affects the learning process by allowing the neural network to better fit the data. It
BL3
helps fine-tune the decision boundary and adjusts the neuron's activation threshold during
training.
23 What is back propagation? (JAN/2022)
Backpropagation is an iterative optimization algorithm used in training artificial neural
BL1
networks. It involves calculating the gradient of the loss function with respect to the network's
weights and biases, allowing for weight adjustments to minimize the error.
24 Define supervised learning(DEC/2022)
Supervised learning is a machine learning paradigm where the algorithm learns
BL1
from labeled training data to make predictions or classifications on unseen
data.Example:classification,prediction.
25 What is meant by a labeled dataset in supervised learning?
In supervised learning, a labeled dataset consists of input data paired with BL2
corresponding target labels, providing a clear association between input and desired output.
26 What is the primary objective of a supervised learning network?
The primary objective is to train a model that can accurately map input features to BL1
target labels, enabling it to make predictions or classifications for new, unseen data.
26 Explain the difference between input features and target labels in a supervised learning
task.
Input Features: Input features in supervised learning are the variables or attributes that describe
the input data, providing the necessary information for the model to make predictions or BL2
classifications.
Target Labels: Target labels in supervised learning are the desired output values or class labels
associated with the input data, serving as the ground truth for training and evaluation.
27 What is meant by input features in a supervised learning task
Input Features: Input features in supervised learning are the variables or attributes
BL1
that describe the input data, providing the necessary information for the model to make
predictions or classifications.
28 What is meant by target labels in a supervised learning task(JAN/2021)
Target Labels: Target labels in supervised learning are the desired output values or BL1
class labels associated with the input data, serving as the ground truth for training and evaluation.
29 Outline the steps involved in training a supervised learning network.
The steps include data collection and preprocessing, splitting data into training and
BL1
testing sets, selecting an appropriate model architecture, training the model using the training
set, evaluating its performance on the test set, and fine-tuning the model if necessary.
30 Compare and contrast regression and classification tasks in supervised learning,
In supervised learning, regression tasks involve predicting continuous numerical
BL3
values (e.g., predicting house prices), while classification tasks involve assigning data points to
predefined categories or classes.
31 What is meant by regression in supervised learning.list the types.
Linear regression in supervised learning is a method used for modeling the BL1

St. Joseph’s Institute of Technology 6

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024

relationship between input features and a continuous target variable by fitting a linear equation to
the data.Types:Simple ,polynomial,multiple regression
32 What is meant by classification tasks in supervised learning(DEC/2022)
In supervised learning, classification tasks involve assigning input data to discrete
classes or categories based on the learned patterns. Examples include image classification, BL1
sentiment analysis, and spam detection.
Task: naive bayes,decision tree,Random forest,logistic regression
33 What is linear regression in supervised learning(DEC/2022)
Linear regression is a supervised learning technique used to model the relationship
between a dependent variable (also called the target or output) and one or more independent BL2
variables (predictors or input features) by fitting a linear equation to the observed data. It is
primarily used for regression tasks, where the goal is to predict a continuous numerical value.
34 What is decision tree in supervised learning(April/2021)
A decision tree is a supervised learning algorithm used for both classification and BL1
regression tasks. It builds a tree-like model of decisions and their possible consequences.
35 What is neural network in supervised learning
In supervised learning, a neural network is a type of model composed of
BL1
interconnected neurons or nodes that can be used for various tasks, including both regression
and classification. It can capture complex patterns in data through its layered structure.
UNIT I/ PART- B
1. Describe the concept of transfer learning in neural networks, and provide a detailed example
BL2
demonstrating its application in a specific domain.
2. Describe the concept of recurrent neural networks (RNNs) (DEC/2022) BL1
3. Provide a real-world example where CNNs have demonstrated superior performance compared to
traditional neural networks or other machine learning algorithms. BL5
4. Trace the evolution of neural networks from the perceptron model to modern deep learning
BL4
architectures. (JAN/2022)
5. Describe the structure and functioning of a Radial Basis Function Network (RBFN) BL2
6. Define and explain the significance of activation functions in neural networks. (JAN/2021) BL1
7. Describe the concept of weight initialization in neural networks(DEC/2022) BL1
8. Explain the purpose and functioning of dropout regularization in neural networks BL2
9. discuss the impact of different initialization techniques on training convergence and model BL4
performance. Provide examples to illustrate. (JAN/2021)
10 Provide examples of different activation functions and discuss scenarios where specific functions BL5
are more suitable. (JAN/2022)
UNIT I/ PART- C
1. Choose a specific industry or field (e.g., finance, healthcare, and robotics) and elaborate on three
distinct applications of Neural Networks within it, highlighting the unique advantages they bring to BL6
each application.
2. Explain the architecture and training process of Convolutional Neural Network (CNN) in
BL1
detail(JAN/2021)
3. Explain the architecture and training process of a Multilayer Perceptron (MLP) in detail, and BL6
provide a practical example where an MLP is well-suited for a specific task. (JAN/2022)
4. Elaborate on the architecture and training mechanisms of a Self-Organizing Map (SOM), and BL3
provide an example illustrating how SOMs can be applied in unsupervised learning tasks.
5. Explain the process of training a recurrent neural network (RNN) for sequential data analysis. BL2
(JAN/2021)
6. Explain the process of training a convolutional neural network (CNN) for image classification. BL3
Detail the architecture, including convolutional layers, pooling layers, and fully connected layers.
UNIT II
ASSOCIATIVE MEMORY AND UNSUPERVISED LEARNING NETWORKS
Training Algorithms for Pattern Association-Autoassociative Memory Network-Heteroassociative Memory
Network-Bidirectional Associative Memory (BAM)-Hopfield Networks-Iterative Autoassociative Memory
Networks-Temporal Associative Memory Network-Fixed Weight Competitive Nets-Kohonen Self-Organizing
Feature Maps-Learning Vector Quantization-Counter propagation Networks-Adaptive Resonance Theory Network.

St. Joseph’s Institute of Technology 7

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024

UNIT II/PART A
1 Define the term "pattern association" in the context of neural networks BL1
In neural networks, pattern association refers to the ability to link and retrieve specific
patterns or features from a given input, allowing the network to recognize and recall familiar
patterns based on learned associations.
2 Give the difference between Hebbian learning and Perceptron learning in pattern BL3
association.
Hebbian learning is an unsupervised learning rule that strengthens connections between
neurons when they are simultaneously active, promoting associative memory formation.
Perceptron learning, on the other hand, is a supervised learning algorithm used for binary
classification tasks, adjusting weights based on misclassification errors.
3 How does an Auto associative Memory Network differ from a Heteroassociative Memory BL4
Network?
Autoassociative Memory Network retrieves complete patterns from partial or noisy
inputs, essentially completing missing information. Heteroassociative Memory Network
associates one pattern with another, enabling the retrieval of related patterns based on learned
associations.
4 What is Learning Vector Quantization (LVQ)? BL1
LVQ is a supervised learning algorithm where prototype vectors are iteratively adjusted
to better represent input data. It classifies input patterns based on the similarity between
prototypes and input vectors.
5 Explain the principle behind Bidirectional Associative Memory (BAM). BL2
BAM is a neural network architecture that reinforces connections between neurons
activated in pairs. It enables bidirectional pattern retrieval, allowing the network to recognize
patterns in both forward and reverse associations.
6 Define Autoassociative Memory Network. (JAN/2021) BL1
An Autoassociative Memory Network is a type of neural network that can reconstruct
complete patterns from partial or noisy inputs, effectively filling in missing information to
retrieve the original pattern.
7 Define Kohonen Self-Organizing Feature Maps. BL1
Kohonen Maps are unsupervised neural networks that map high-dimensional input data
onto a lower-dimensional grid while preserving the topological relationships between data
points. They're useful for tasks like clustering and dimensionality reduction.
8 Describe the functioning of Hopfield Networks. BL1
Hopfield Networks are recurrent neural networks where each neuron is connected to
every other neuron. They utilize iterative update rules to store and retrieve patterns, allowing
for content-addressable memory and associative recall.
9 What is the significance of Iterative Autoassociative Memory Networks? BL2
Iterative Autoassociative Memory Networks: These networks refine pattern
representations through iterative processing, enhancing the accuracy and robustness of pattern
retrieval, particularly from noisy or incomplete inputs.
10 How does Temporal Associative Memory Network differ from other associative BL3
memory models?
Unlike other associative memory models, Temporal Associative Memory
Network focuses on associating patterns over time. This allows for the sequential recall
of patterns, making it suitable for tasks involving temporal sequences.
11 Define Heteroassociative Memory Network. (April/2021) BL1
Heteroassociative Memory Network associates one pattern with another, enabling the
retrieval of related patterns based on the similarity of their associations. This is distinct from
autoassociative networks which retrieve similar patterns from partial inputs.
12 Explain the concept of Fixed Weight Competitive Nets. BL2
These neural networks have fixed connections between input neurons and output neurons.
Each output neuron competes to respond to a specific pattern, with the strongest response
representing the recognized pattern.

St. Joseph’s Institute of Technology 8

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
13 Describe the process of self-organization in Kohonen Self-Organizing Feature Maps. BL2
Kohonen Self-Organizing Feature Maps organize data in a lower-dimensional grid.
During training, neurons with similar weights become neighbors, preserving the topological
relationships of the input data, enabling tasks like clustering and dimensionality reduction.

14 How does Learning Vector Quantization (LVQ) work? BL4

LVQ is a supervised learning algorithm where prototype vectors are adjusted based on
the similarity between input data and prototypes. It classifies patterns by finding the closest
prototype to the input vector.
15 Explain the training process of Counter propagation Networks. BL2
Counter propagation Networks Training: In this two-stage process, the first stage
(competitive) selects the best neuron, and the second stage (error-correction) refines the weights
based on the difference between the output and target values.
16 What are the adaptive characteristics of Adaptive Resonance Theory Network (ART)? BL1
ART networks adaptively adjust their response thresholds and synaptic weights,
allowing them to categorize patterns while being robust to variations in input patterns.
17 What is Bidirectional Associative Memory (BAM)? BL1
BAM is a neural network that establishes associations between patterns in a bidirectional
manner. When presented with a partial input, it recalls the corresponding associated pattern and
vice versa.
18 Provide an example scenario where an Autoassociative Memory Network can be effectively BL3
used.
In image reconstruction, if an image is partially obscured or noisy, an autoassociative
network can complete the missing parts, restoring the original image.
19 Give an application example where a Heteroassociative Memory Network is more suitable BL3
than an Autoassociative Memory Network.
In language translation, where one language pattern is associated with its corresponding
translation, a heteroassociative network can retrieve the translation based on the input pattern.
20 How can Bidirectional Associative Memory (BAM) be applied in real-world problem- BL4
solving? (JAN/2021)
In cybersecurity, BAM could be used to detect patterns of suspicious network behavior.
It would associate certain behaviors with specific threats, allowing for bidirectional
identification.
21 In what practical situations are Hopfield Networks commonly used? BL5
They're used in tasks like content-addressable memory, optimization problems, and
associative recall, including applications in error correction and combinatorial optimization.
22 Explain the concept of Hopfield Networks. BL2
They are a type of recurrent neural network where neurons are connected to every other
neuron. The network iteratively updates its state to converge towards stored patterns, enabling
associative memory and pattern recognition.
23 Provide an application example where Iterative Autoassociative Memory Networks play a BL4
crucial role.
In medical imaging, they play a crucial role in reconstructing high-quality images from
noisy or incomplete scans, aiding in accurate diagnosis and treatment planning.
24 Describe a real-world scenario where Temporal Associative Memory Network is beneficial. BL3
In speech recognition, a temporal associative memory network helps in understanding
and interpreting the context of spoken words in a sentence, allowing for more accurate and
context-aware transcription.
25 Describe a situation where Learning Vector Quantization (LVQ) is the preferred learning BL5
algorithm. (JAN/2021)
LVQ is preferred in scenarios like character recognition, where prototypes represent
distinct characters. It effectively classifies input patterns into specific character categories.
26 In what domains can Counterpropagation Networks be most effectively employed? BL4
Counter propagation Networks are effectively employed in fields like fault diagnosis in
complex systems (e.g., manufacturing plants) and prediction tasks in financial markets.
27 What are Iterative Autoassociative Memory Networks? BL2
They are neural networks that refine pattern representations through iterative processing,
making them robust for pattern retrieval, especially from noisy or incomplete inputs.

St. Joseph’s Institute of Technology 9

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
28 Give an application domain where Fixed Weight Competitive Nets are employed. BL3
Fixed Weight Competitive Nets, also known as Kohonen Self-Organizing Maps
(SOMs), are artificial neural networks used in various application domains where data
clustering, dimensionality reduction, and visualization are essential. In anomaly detection for
network security, these nets identify unusual patterns in network traffic, where fixed connections
allow for the detection of irregularities.

29 Provide an example scenario where Kohonen Self-Organizing Feature Maps have been BL3
successfully applied.
Kohonen Self-Organizing Feature Maps Example: In customer segmentation for
marketing, these maps group similar customer behaviors, enabling targeted marketing strategies
tailored to specific segments.
30 Explain Counterpropagation Networks. BL2
They have a two-stage training process: competitive selection of neurons in the first
stage, and error-correction in the second stage. This enables them to categorize input patterns
effectively.
31 Provide a practical example of an application where Adaptive Resonance Theory Network BL3
(ART) has shown significant advantage.
In medical diagnosis, ART networks adaptively adjust to variations in patient data,
allowing for accurate and reliable classification of diseases, even with variable symptoms and
test results.
32 Define Temporal Associative Memory Network. BL1
It's an associative memory model that associates patterns over time, making it ideal for
tasks like speech recognition, where understanding the temporal context is crucial.
33 Explain Fixed Weight Competitive Nets. BL2
These neural networks have fixed connections between input and output neurons, and
they're employed in scenarios like face recognition systems, where specific facial features must
be identified reliably.
34 Define Adaptive Resonance Theory Network (ART). BL2
Adaptive Resonance Theory Network (ART) Definition: ART networks
dynamically adjust their response thresholds and synaptic weights, making them excellent
for tasks where patterns may vary or evolve over time, such as adaptive learning systems.
35 Explain an application domain where Fixed Weight Competitive Nets are employed. BL4
In fingerprint recognition systems, they identify unique features by establishing fixed
connections, ensuring consistent and accurate recognition of individuals.
UNIT II/ PART- B
1 Analyze the key characteristics and limitations of Bidirectional Associative Memory (BAM) in BL4
pattern association.
2 Evaluate the effectiveness and limitations of Hopfield Networks in solving specific problem BL5
domains.
3 Explain the significance of Iterative Autoassociative Memory Networks in complex pattern BL2
recognition tasks.
4 Assess the suitability of Temporal Associative Memory Network in scenarios with dynamic data BL4
patterns.
5 Evaluate the strengths and weaknesses of Kohonen Self-Organizing Feature Maps in clustering BL5
and dimensionality reduction tasks.
6 Compare Learning Vector Quantization (LVQ) with other unsupervised learning algorithms, BL4
discussing scenarios where LVQ is most effective.
7 Evaluate the adaptability and learning capabilities of Adaptive Resonance Theory Network BL5
(ART) in dynamic environments. (JAN/2022)
8 Define Autoassociative Memory Network and provide an example of its application. BL1
9 What are the key characteristics of Bidirectional Associative Memory (BAM) and how is it BL2
utilized in pattern association?
10 Explain the concept of Iterative Autoassociative Memory Networks and provide a practical BL2
application example.
UNIT II/ PART- C
1 Compare and contrast Autoassociative Memory Network and Heteroassociative Memory BL4
Network, highlighting their respective applications and advantages.
2 Analyze the function and potential applications of Fixed Weight Competitive Nets in BL4
unsupervised learning.

St. Joseph’s Institute of Technology 10

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
3 Assess the effectiveness of Counterpropagation Networks in tasks involving high-dimensional BL5
data.
4 Describe a real-world application scenario where Bidirectional Associative Memory (BAM) can
be effectively utilized BL6
5 Explain a practical application domain where Fixed Weight Competitive Nets are used for
BL5
unsupervised learning tasks.

UNIT III THIRD-GENERATION NEURAL NETWORKS

Spiking Neural Networks-Convolutional Neural Networks-Deep Learning Neural Networks-Extreme Learning
Machine Model-Convolutional Neural Networks: The Convolution Operation – Motivation – Pooling – Variants of
the basic Convolution Function – Structured Outputs – Data Types – Efficient Convolution Algorithms –
Neuroscientific Basis – Applications: Computer Vision, Image Generation, Image Compression.
UNIT- III /PART –A
1. Define Spiking Neural Networks and provide an example of their application. BL1
These are neural networks that model communication between neurons using spikes
or action potentials. They operate on the principle of timing-based information processing.
Example Application: Neuromorphic computing, where SNNs mimic the behavior of biological
neurons for efficient and low-power processing in hardware.
2. Discuss the importance of Structured Outputs in specialized applications like object BL4
detection and segmentation using CNNs.
In specialized applications like object detection and segmentation, structured outputs
provide detailed information about the spatial arrangement of objects. This is crucial for tasks
that require precise localization and delineation of objects within an image.
3. Analyze the variants of the basic Convolution Function in CNNs and discuss scenarios BL4
where specific variants are preferred.
Variants like dilated convolution are used in tasks requiring a broader context, such as
semantic segmentation. Depthwise separable convolution reduces computational cost and is
valuable in mobile and embedded applications.
4. Explain the core principles of Convolutional Neural Networks (CNNs) and their BL2
significance in modern AI applications. (JAN/2022)
CNNs are designed to automatically and adaptively learn spatial hierarchies of
features from input data. They excel in tasks like image recognition due to their ability to
extract and learn meaningful features directly from raw data.
5. Provide an example of a practical situation where Convolutional Neural Networks (CNNs) BL3
outperform traditional machine learning models.
CNNs Outperforming Traditional Models: In image classification, CNNs
outperform traditional machine learning models because they can automatically learn
hierarchical features, eliminating the need for manual feature engineering. This makes them
highly effective in tasks like image recognition.
6. Describe a real-world application scenario where Deep Learning Neural Networks play a BL4
crucial role in solving a complex problem.
In autonomous driving, Deep Learning Neural Networks play a crucial role in
processing sensor data (like images and LiDAR) to make real-time decisions, enabling vehicles
to navigate safely in complex environments.
7. Provide an example of a specialized application in Computer Vision that heavily relies BL4
on Third-Generation Neural Networks.
In natural language processing, Third-Generation NNs like Transformers are used for
tasks like language translation and sentiment analysis. They excel in tasks requiring
understanding of contextual relationships in text.
8. Evaluate the impact of Pooling in Convolutional Neural Networks and discuss its BL5
contribution to reducing dimensionality.
Pooling reduces spatial dimensions, allowing the network to focus on more important
features. This helps in reducing computational complexity, preventing overfitting, and achieving
translation invariance in CNNs.
9. What is meant by the Extreme Learning Machine Model? BL2
It's a machine learning algorithm where only the output weights are learned, and input
weights are randomly generated. It's known for its simplicity, speed, and ability to handle high-
dimensional data.

St. Joseph’s Institute of Technology 11

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
10 Define Deep Learning Neural Networks and highlight their role in advanced AI tasks. BL1
(Dec/2021)
These are neural networks with multiple hidden layers, allowing them to learn intricate
and abstract features from data. They play a pivotal role in advanced AI tasks like natural
language processing, computer vision, Financial Service,Health care and speech recognition.

11. Explain the rationale behind the Convolution Operation in CNNs and its role in feature BL2
extraction from images.
It scans a filter (also known as kernel) over the input to extract features. This operation
allows the network to recognize local patterns in the input, enabling it to learn hierarchical
representations of features.
12. Analyze the effectiveness of Convolutional Neural Networks in computer vision tasks BL4
compared to other machine learning approaches.
Effectiveness of CNNs in Computer Vision: CNNs outperform other machine learning
approaches in computer vision tasks because they can automatically learn and extract features
directly from raw pixel values, eliminating the need for manual feature engineering and enabling
better performance on complex visual recognition tasks.
13. What is the Extreme Learning Machine Model and in what scenarios is it particularly BL2
effective?
ELM is a machine learning algorithm where only the output weights are learned, while
the input weights are randomly generated. It is particularly effective in scenarios with large
datasets and high-dimensional input spaces, as it offers simplicity, speed, and good
generalization performance.
14. What is the significance of Datatypes in Third-Generation Neural Networks? BL1
Datatypes in Third-Generation Neural Networks determine how information is
represented and processed. For example, in Transformers, the use of float16 or bfloat16 can
significantly reduce memory requirements without sacrificing model accuracy.
15. Explain how different Data Types can affect the performance and memory requirements BL5
of Convolutional Neural Networks.
Using lower precision data types like float16 or int8 can reduce memory requirements and
improve inference speed in CNNs. However, it may also lead to a slight loss in model accuracy
due to reduced precision.
16. Compare and contrast Spiking Neural Networks with traditional artificial neural BL4
networks(JAN/2021)
Spiking Neural Networks model neuron communication with spikes, which is more bio-
realistic but can be computationally expensive. Traditional ANNs use continuous activation
functions. SNNs are efficient for event-based processing, while ANNs are more suitable for
traditional continuous data processing.
17. List Spiking Neural Networks’s advantages and limitations. BL1
 Advantages: Event-driven processing, high energy efficiency, ability to
process asynchronous data, and biological plausibility.
 Limitations: Higher computational requirements, lack of standardized frameworks, and
challenges in training.
18. Explain the key components of Convolutional Neural Networks, including Convolution BL2
Operation, Pooling, and variants of the basic Convolution Function.
 Convolution Operation: It applies a filter over the input to extract features. This
enables the network to recognize local patterns.
 Pooling: Reduces spatial dimensions, focusing on important features and aiding
in translation invariance.
 Variants of Convolution Function: Different functions (e.g., dilated convolution)
modify the convolution operation for specific applications.

St. Joseph’s Institute of Technology 12

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
19. Evaluate the computational efficiency gains of Efficient Convolution Algorithms in BL5
Convolutional Neural Networks.
Efficient Convolution Algorithms play a crucial role in improving the computational
efficiency of Convolutional Neural Networks (CNNs). These algorithms are designed to reduce
the computational complexity of the convolution operation while maintaining or even improving
the model's performance.
 Reduce FLOPS
 Model Performance
20. Compare Spiking Neural Networks with traditional artificial neural networks, BL4
discussing their respective advantages and limitations.
Advantages of Spiking Neural Networks: Mimic biological processes, event-driven
processing, high energy efficiency, and ability to process asynchronous data.
Limitations of Spiking Neural Networks: Higher computational demands, lack of standardized
frameworks, and challenges in training due to spike-based learning.

21. Describe the concept of Structured Outputs in Convolutional Neural Networks and their BL2
importance in specialized tasks.
Structured outputs provide detailed information about the spatial arrangement of objects
in tasks like object detection and segmentation. They are crucial for tasks where precise
localization and delineation of objects within an image are required.
22. Explain how different Datatypes can impact the performance of Third-Generation Neural BL2
Networks.
Different datatypes (e.g., float16, int8) can significantly affect the model's performance.
Choosing the right datatype is crucial for balancing accuracy and computational efficiency.
Floating Point has high precision values when compared to integer.
23. Define Datatypes in the context of Third-Generation Neural Networks. BL1
Datatypes in Third-Generation Neural Networks refer to the precision with which
numerical values are represented in the network's computations. They include options like
float32, float16, int8, etc., each with varying levels of precision and memory requirements.
24 Define Data Types in the context of Convolutional Neural Networks and explain their BL1
relevance.
Data types in CNNs refer to the precision of numerical values used in computations.
For example, using lower precision data types like float16 or int8 can reduce memory
requirements and improve inference speed without sacrificing much accuracy.
25 How do Datatypes affect the memory requirements of Third-Generation Neural BL2
Networks?
Using lower precision datatypes (e.g., float16) reduces memory requirements but may
result in some loss of model accuracy. Higher precision datatypes (e.g., float32) require more
memory but provide greater numerical precision.
26 Provide an example scenario where the choice of Datatypes is crucial for the effectiveness BL3
of a neural network.
In edge computing for IoT devices, choosing lower precision datatypes like int8 is crucial
for reducing memory usage and computational load, allowing the neural network to run
efficiently on resource-constrained devices.
27 Detail Efficient Convolution Algorithms and how they enhance the computational BL2
efficiency of CNNs.
These algorithms optimize the computation of convolutions, reducing the number of
operations required. Techniques like Winograd, FFT-based convolutions, and depthwise
separable convolutions significantly enhance the computational efficiency of CNNs.
28 Evaluate the diverse applications of Third-Generation Neural Networks in Computer BL3
Vision.
Third-Generation NNs, such as Transformers, have been applied in tasks like image
classification, object detection, and image captioning. They excel in tasks requiring an
understanding of contextual relationships within images.
29 Evaluate the diverse applications of Third-Generation Neural Networks in Image BL3
Generation.
Third-Generation NNs like Generative Adversarial Networks (GANs) have found
applications in generating realistic images for tasks like face synthesis, artistic style transfer, and
deepfake creation.

St. Joseph’s Institute of Technology 13

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
30 Evaluate the diverse applications of Third-Generation Neural Networks in Image BL3
Compression.
Using networks like Variational Autoencoders (VAEs), Third-Generation NNs have been
applied to lossy image compression tasks, enabling more efficient storage and transmission of
images.
31 Explain the Neuroscientific Basis of Spiking Neural Networks and its connection to BL2
biological neural systems.
SNNs model neural communication through spikes, mimicking the behavior of
biological neurons. They incorporate the concept of neuron firing thresholds, refractory periods,
and synaptic plasticity, closely resembling the behavior of real neurons.
32 Assess the computational efficiency gains achieved by Efficient Convolution Algorithms in BL4
CNNs and their implications for real-time applications.
Efficient convolution algorithms, like depthwise separable convolution, significantly
reduce computational requirements while maintaining high accuracy. This leads to faster
inference times, making CNNs suitable for real-time applications like autonomous driving and
video processing.

33 Analyze the neuroscientific principles incorporated into Spiking Neural Networks and BL4
their potential benefits in AI research.
SNNs replicate biological neurons' behavior with events based on timing and spikes.
They incorporate concepts like synaptic weight updates based on spike timings and refractory
periods, making them more biologically plausible and potentially useful in modeling complex
neural systems.
34 Provide examples of applications of Third-Generation Neural Networks in Computer BL2
Vision, Image Generation, and Image Compression.
 Computer Vision: Transformers for image captioning, where understanding
image context is crucial.
 Image Generation: GANs for creating photorealistic images in tasks like face synthesis.
 Image Compression: VAEs for lossy compression, reducing storage requirements
while maintaining quality.
35 Evaluate the diverse applications of Third-Generation Neural Networks in Computer BL5
Vision, Image Generation, and Image Compression, providing specific examples.
 Computer Vision: Transformers for fine-grained image classification, GANs for
super- resolution image reconstruction.
 Image Generation: StyleGAN for generating high-quality art and realistic faces.
 Image Compression: VAEs for efficient storage and transmission of medical
imaging data.
UNIT III/ PART- B
1. Define Spiking Neural Networks and explain their fundamental operation. BL1
2. Discuss the importance of Structured Outputs in specialized applications like object detection and BL2
segmentation using CNNs.
3. Define Data Types in the context of Convolutional Neural Networks and discuss their BL1
importance.
4. Define the Extreme Learning Machine Model and provide an example of its application. BL2
5. Analyze the variants of the basic Convolution Function in CNNs and discuss scenarios where BL4
specific variants are preferred.
6. Describe the core principles of Convolutional Neural Networks (CNNs) and their applications. BL2
(JAN/2021)
7. Evaluate the diverse applications of Third-Generation Neural Networks in Computer Vision, BL5
Image Generation, and Image Compression, providing specific examples.
8. Explain the Neuroscientific Basis behind Spiking Neural Networks and how it relates to BL3
biological neural systems.
9. Explain how different Data Types can affect the performance and memory requirements of BL4
Convolutional Neural Networks. (Dec/2022)
10. Examine the effectiveness of Convolutional Neural Networks in computer vision tasks compared BL5
to other machine learning approaches.
UNIT III/ PART- C
1. Analyze the neuroscientific principles incorporated into Spiking Neural Networks and their BL4
potential benefits in AI research.

St. Joseph’s Institute of Technology 14

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
2. Assess the computational efficiency gains achieved by Efficient Convolution Algorithms in BL5
CNNs and their implications for real-time applications.
3. Evaluate the impact of Pooling in Convolutional Neural Networks and discuss its contribution to BL5
reducing dimensionality.
4. Compare and contrast Spiking Neural Networks with traditional artificial neural networks, BL6
highlighting their respective advantages and limitations.
5. Describe Efficient Convolution Algorithms and their impact on the computational efficiency of BL2
CNNs.
UNIT IV DEEP FEEDFORWARD NETWORKS
History of Deep Learning- A Probabilistic Theory of Deep Learning- Gradient Learning – Chain Rule and
Backpropagation - Regularization: Dataset Augmentation – Noise Robustness -Early Stopping, Bagging and
Dropout - batch normalization- VC Dimension and Neural Nets.
UNIT IV/ PART- A
1 Define Deep Learning.
Deep learning is a subfield of machine learning that focuses on artificial neural BL1
networks with multiple layers (deep neural networks) to learn and make predictions from data.

St. Joseph’s Institute of Technology 15

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
It seeks to automatically discover and represent hierarchical patterns in data through the
training of neural networks.
2 List out any four Application of Deep Learning. (Dec/2021)
a. Image Classification and Object Detection
b. Natural Language Processing (NLP) BL2
c. Speech Recognition
d. Autonomous Driving
3 State the use of hidden layers. (JAN/2021)
Hidden layers in a neural network are used to learn and represent complex, hierarchical
BL1
features and patterns in data. They enable the network to model non-linear relationships and
capture abstract information from the input.
4 Why deep learning models are called feedforward?
Deep learning models are called feedforward because the data flows through the BL1
network from input to output without any feedback loops or connections that create cycles.
5 What is gradient Descent? (April/2021)
Gradient descent is an optimization algorithm used to minimize the cost function of a
machine learning model by iteratively adjusting the model's parameters in the direction of BL1
steepest descent of the cost function.
6 What is cost function?give the formula.
The cost function (also known as loss function) quantifies the error or mismatch
between the model's predictions and the actual target values. A common cost function for
regression problems is Mean Squared Error (MSE), represented as: BL1

7 List the component of hidden layers.

Hidden layers consist of nodes (neurons), each with weights, biases, and an BL4
activation function.
8 What is sigmoid function?
The sigmoid function is an activation function commonly used in neural networks. It
transforms the weighted sum of inputs into a range between 0 and 1.
Its formula is:
BL1

9 Difference between forward propagation and backward Propagation.

 Forward Propagation: It involves passing input data through the network to
generate predictions. BL2
 Backward Propagation: It is the process of calculating gradients of the cost function
with respect to the model's parameters, which is used for optimization during training.
10 What is dataset augmentation?
Dataset augmentation is a technique in deep learning where new training examples
BL1
are created by applying various transformations (e.g., rotations, translations, flips) to the
existing data. It helps improve model generalization.
11 What is forward propagation?
Forward Propagation: It involves passing input data through the network to generate BL1
predictions.
12 What is backward Propagation?
Backward Propagation: It is the process of calculating gradients of the cost function BL1
with respect to the model's parameters, which is used for optimization during training.
13 Define Regularization. (JAN/2022)
Regularization is a technique to prevent overfitting by adding a penalty term to the
BL1
cost function. It discourages the model from fitting the noise in the data.
Types: L1 and L2
14 Difference between L1 and L2 regularization.
 L1 Regularization adds the absolute values of the weights to the cost function. BL1
 L2 Regul
arization adds the squared values of the weights to the cost
function.
St. Joseph’s Institute of Technology 15
Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
15 State few points on early stopping.
Early stopping is a technique in training deep learning models where training is halted BL1
when the model's performance on a validation set starts to degrade, preventing overfitting.
16 What are the types of output units available in deep learning?
In deep learning, common output units include softmax (for multiclass classification), BL1
linear (for regression), and sigmoid (for binary classification).
17 What is the drawback of deep learning?
Deep learning models often require large amounts of data and computing power, BL1
making them resource-intensive. They can also be prone to overfitting with small datasets.
18 List out the applications of Deep learning.
 Healthcare (e.g., medical image analysis)
 Recommendation systems BL1
 Gaming (e.g., AlphaGo)
 Financial forecasting
19 Draw the relation between A.I,M.L,D.L and Data Science .

BL1

20 How is the posterior probability between two random variables represented?

The posterior probability between two random variables is often represented using
BL3
conditional probability notation, such as P(X | Y), which represents the probability of event X
given event Y.
21 Why was manifold learning introduced?
Manifold learning was introduced to find a lower-dimensional representation of high- BL3
dimensional data while preserving its underlying structure. It helps in dimensionality reduction.
22 List out any five point for using Deep Learning.
 Large datasets
 Complex, non-linear relationships in data
BL1
 High computational resources
 Need for automatic feature learning
 Availability of labeled data
23 What are the factors that influence the backpropagation training? (April/2021)
 Learning rate
 Network architecture
BL1
 Initialization methods
 Batch size
 Regularization techniques
24 Differentiate Bagging from Boosting
Bagging (Bootstrap Aggregating) creates multiple parallel models by resampling the
dataset with replacement and averaging their predictions. BL1
Boosting creates an ensemble of models sequentially, with each model giving more weight to
misclassified samples from the previous model.
25 What are the advantages of Adaboosting?
 Improved accuracy by focusing on difficult examples.
BL1
 Robustness to overfitting.
 Simplicity of implementation.
26 What is VC-Dimensions.
VC-dimensions are a theoretical measure of the capacity of a machine learning model or
BL1
hypothesis space.They provide insights into a model's ability to fit various patterns in data.

St. Joseph’s Institute of Technology 16

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
27 Why is understanding the VC dimension important for addressing overfitting in neural
networks?
It provides a theoretical framework for characterizing the capacity or complexity of a
BL1
model's hypothesis space. The VC dimension helps us analyze a model's ability to fit noise in
the training data, which is a common cause of overfitting.

28 In what layers of a neural network is batch normalization typically applied?

Batch normalization is typically applied to the hidden layers of a neural network to
BL1
improve training stability and accelerate convergence.It will improve the performance of
the model.
29 What are the key parameters in batch normalization, and how do they affect the
normalization process?
BL1
 Scale and shift parameters
 Moving averages and variances
30 Explain the potential benefits of using an ensemble of models created through bagging.
(Dec/2021)
Bagging can improve model robustness and accuracy by reducing variance. It combines
BL1
multiple models trained on different subsets of data. It does this by taking random subsets of
an original dataset, with substitute, and fits either a classifier (for classification) or regressor
(for regression) to each subset.
31 How do weight updates occur in back propagation-based learning?
In backpropagation, weight updates occur by computing the gradient of the cost
BL3
function with respect to the model's parameters and adjusting the weights in the opposite
direction of the gradient.
32 Name two key historical milestones in the development of deep learning.
1957: The invention of the first artificial neural network by Frank Rosenblatt. BL5
2012: The breakthrough of deep learning in computer vision with AlexNet
33 What is the relationship between neural networks and the history of deep learning.
Neural networks are the foundational models used in deep learning. Deep learning BL1
extends neural networks by adding more hidden layers.
34 How can uncertainty estimation be beneficial in deep learning applications?
Uncertainty estimation in deep learning provides confidence intervals or uncertainty BL1
measures for model predictions, which can be critical in safety-critical applications.
35 What are some advantages of using probabilistic models in deep learning?
 They can provide probabilistic predictions, which include uncertainty information.
BL1
 They are suitable for handling data with inherent uncertainty.
 They allow Bayesian reasoning for model parameters.
UNIT IV/ PART- B
1 Explain the probabilistic foundations of deep learning and how they differ from traditional
BL4
machine learning approaches.
2 Provide a step-by-step explanation of the backpropagation algorithm, emphasizing the role
of the chain rule in computing gradients. BL3
3 How does dataset augmentation contribute to regularization in deep learning? Provide
BL2
examples of data augmentation techniques.
4 Discuss methods for making deep learning models more robust to noisy data. How does
BL2
regularization play a role in this context?
5 Compare and contrast early stopping, bagging (bootstrap aggregating), and dropout as
BL3
techniques for improving neural network training and generalization.
6 Explain the concept and benefits of batch normalization in deep neural networks. How does it
BL3
address issues like internal covariate shift?
7 Define VC dimension and discuss its relevance to neural networks. How does VC dimension
BL5
relate to model capacity and generalization?
8 What are the gradient vanishing and exploding problems in deep learning? Provide insights
BL6
into their causes and potential solutions.
9 Explain the importance of activation functions in neural networks. Compare and contrast
BL6
common activation functions like ReLU, Sigmoid, and Tanh.
10 Differentiate Machine learning and Deep learning along with the architectural flow. BL6
UNIT IV/PART C BL2
1 Provide an overview of generative adversarial networks (GANs) and their applications in
BL2
generating synthetic data.

St. Joseph’s Institute of Technology 17

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
2 Describe the architecture and applications of convolutional neural networks (CNNs) in deep
BL1
learning. How do they handle spatial hierarchies?
3 Discuss the architecture and use cases of recurrent neural networks (RNNs) in deep learning.
BL2
What are the challenges in training RNNs?
4 Explain the principles behind autoencoders and variational autoencoders (VAEs) in
BL4
unsupervised learning. How do they differ?
5 Describe the concept of transfer learning in deep learning. How does it enable the reuse
BL3
of pre-trained models for new tasks, and what are the benefits?
UNIT V RECURRENT NEURAL NETWORKS
Recurrent Neural Networks: Introduction – Recursive Neural Networks – Bidirectional RNNs – Deep Recurrent
Networks – Applications: Image Generation, Image Compression, Natural Language Processing. Complete Auto
encoder, Regularized Autoencoder, Stochastic Encoders and Decoders, Contractive Encoders.
UNIT V/ PART-A
1 What is the fundamental characteristic that distinguishes RNNs from feedforward
neural networks?
The fundamental characteristic that distinguishes Recurrent Neural Networks (RNNs)
from feedforward neural networks is their ability to handle sequential data by maintaining BL1
hidden states that capture temporal dependencies. Unlike feedforward networks, where
information flows in one direction, RNNs have recurrent connections that allow information to
cycle back into the network, making them suitable for tasks involving sequences.
2 State the vanishing gradient problem in the context of training RNNs.
The vanishing gradient problem in RNNs occurs during training when gradients become
extremely small as they are back-propagated through time. This makes it difficult for the BL2
network to learn long-range dependencies in sequences, causing issues with capturing and
preserving information over extended time steps.
3 How does an RNN handle sequential data? Provide the architecture.
RNNs handle sequential data by maintaining a hidden state vector that evolves with
each time step. The architecture includes an input layer, hidden layer (with recurrent BL1
connections), and an output layer. The hidden state at each time step captures information from
previous steps and combines it with the current input to make predictions.
4 What are the primary limitations of vanilla RNNs when dealing with long sequences?
Vanilla RNNs struggle with handling long sequences due to the vanishing gradient
BL1
problem. They also have difficulty in capturing complex dependencies and suffer from the
"exploding gradient" issue when gradients become too large during training.
5 Describe the purpose of the hidden state in an RNN. (JAN/2021)
The hidden state in an RNN serves as a memory that encodes information about the
BL2
past sequence elements. It allows the network to maintain context and capture dependencies
between elements in the sequence, making it crucial for sequential data tasks.
6 What are Recursive Neural Networks, and how do they differ from traditional RNNs?
Recursive Neural Networks are a type of RNN that handle hierarchical structures by
recursively applying neural network units to subparts of the input. They differ from traditional BL4
RNNs in their ability to model hierarchical relationships, making them useful for tasks like
parsing natural language sentences.
7 Provide an example of a task where Recursive Neural Networks are particularly useful.
(April/2021)
RecNNs are particularly useful for tasks involving syntactic or semantic parsing of BL2
sentences, where the hierarchical structure of language needs to be understood. For example, in
constituency parsing, RecNNs can recursively build parse trees for sentences.
8 How can Recursive Neural Networks handle hierarchical structures in data?
RecNNs handle hierarchical structures by breaking down the input into smaller
BL1
components and recursively processing them with neural network units. This enables them to
capture dependencies at various levels of abstraction within the data.
9 What are some advantages of using recursive structures in natural language processing
tasks?
Using recursive structures in natural language processing tasks allows for better modeling BL1
of hierarchical relationships, improved understanding of context, and enhanced performance in
tasks like parsing, sentiment analysis, and machine translation.
10 Explain the concept of bidirectional RNNs and their significance in sequence modeling.
Bidirectional RNNs process input sequences in both forward and backward directions BL5

St. Joseph’s Institute of Technology 18

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
simultaneously. This bidirectional information flow enhances the model's understanding of
context by considering past and future context at each time step.
11 In what scenarios would you use bidirectional RNNs over unidirectional RNNs?
Bidirectional RNNs are useful when the prediction at a given time step depends on both
BL1
past and future context. They are commonly employed in tasks like speech recognition,
sentiment analysis, and named entity recognition.
12 How does bidirectional information flow affect the hidden state computation in an RNN?
In bidirectional RNNs, the hidden state computation combines information from both the
BL2
forward and backward directions, resulting in a more comprehensive representation of the input
sequence.
13 Can bidirectional RNNs be applied to image processing tasks? If so, how?
Yes, Bidirectional RNNs can be applied to image processing tasks, such as image
BL1
captioning, where understanding context in both directions (left-to-right and right-to-left) helps
generate more accurate descriptions for images.
14 What are Deep Recurrent Networks, and how do they differ from shallow RNNs?
Deep Recurrent Networks have multiple layers of recurrent units stacked on top of each
BL1
other. They differ from shallow RNNs by allowing for more complex hierarchical
representations.
15 What advantages does a deep architecture provide in sequence modeling?
Deep RNNs can capture intricate dependencies and hierarchical features in data, making
BL4
them more suitable for tasks with complex sequences. They can learn hierarchical abstractions
and representations.
16 Discuss the challenges of training deep recurrent networks.
Training deep recurrent networks can be challenging due to the vanishing and exploding
BL1
gradient problems, increased computational complexity, and the need for careful initialization
and regularization.
17 How can you mitigate the vanishing gradient problem in deep recurrent
networks?
To mitigate the vanishing gradient problem in deep recurrent networks, techniques like BL1
gradient clipping, using different activation functions (e.g., LSTM or GRU units), and careful
weight initialization are employed.
18 Explain how Recurrent Neural Networks can be used for image generation.
RNNs can be used for image generation by training them to generate sequences of data
BL1
representing pixels or other image elements. This is commonly seen in tasks like generating
handwritten characters or artistic images.
19 What are the advantages of using RNNs in image compression compared to traditional
methods? (Dec/2022)
RNNs offer advantages in image compression by learning efficient representations of BL1
images. They can capture complex patterns and reduce data size while preserving essential
information.
20 Give an example of a natural language processing task where RNNs have shown
significant success.
RNNs have shown significant success in NLP tasks like machine translation, where BL1
they model the sequential nature of language and capture dependencies between words,
enabling accurate translation between languages.
21 How do RNNs help in language modeling for tasks like text generation and machine
translation? (JAN/2021)
RNNs are crucial for language modeling in tasks like text generation and machine BL3
translation. They learn to predict the next word in a sequence based on the context provided by
the previous words.
22 What is the primary purpose of an autoencoder in neural network architectures?
An autoencoder's primary purpose is to learn efficient representations of data by encoding
BL1
it into a lower-dimensional space and then decoding it back to its original form. It can be used
for tasks like data compression, denoising, and feature learning.
23 Describe the structure of a complete autoencoder and its training process.
A complete autoencoder consists of an encoder that maps input data to a latent
BL1
representation and a decoder that reconstructs the input from the latent representation. Both the
encoder and decoder are neural networks trained together.
24 How do regularized autoencoders prevent overfitting during training? (JAN/2022)
Regularized autoencoders add penalty terms to the loss function to prevent overfitting BL4

St. Joseph’s Institute of Technology 19

Downloaded by swati zindani
 CCS355- Neural Networks and Deep Learning Department of CSE 2023-2024
during training. Techniques like dropout, weight decay, and sparsity constraints are commonly
used for regularization.
25 What is the role of stochastic encoders and decoders in variational autoencoders?
VAEs introduce stochasticity in the encoder and decoder to generate probabilistic latent
BL2
representations. This allows for sampling from the latent space, making them useful for tasks
like generating new data samples.
26 Define how contractive encoders are used to impose structure in the latent space of
autoencoders.
Contractive encoders are used to impose structure in the latent space by penalizing changes BL3
in the encoder's output with respect to the input data. This helps in creating a structured and
meaningful latent space.
27 Define Sparse autoencoder.
A sparse autoencoder encourages the learned representations to be sparse, meaning that
BL5
only a few units in the latent representation are activated for a given input. This can help in
feature selection and better understanding of data.
28 List out any four disadvantage of RNN
 Difficulty in capturing long-range dependencies.
 Vulnerability to vanishing and exploding gradient problems. BL1
 Limited parallelization due to sequential processing.
 Computationally expensive for deep architectures.
29 What are the advantages of RNN.
a. Effective for sequential data tasks.
b. Maintain memory of past inputs with hidden states.
BL5
c. Suitable for tasks like time series prediction, natural language processing, and speech
recognition.
d. Can capture temporal dependencies.
30 Why RNN is called as Recurrent?
RNNs are called recurrent because they use feedback connections, allowing information to BL5
cycle through the network, making them well-suited for sequential and temporal data modelling.
UNIT V/PART B
1. Differentiate RNN and Feedforward Neural Networks along with architecture. (April/2021) BL2
2. a. What are the applications of NLP
BL2
b. State the two phases of NLP.
3. Explain the four types of RNN with example. (JAN/2021) BL6
4. List out and explain the four hyper parameters that are set before training the autoencoder. BL3
5 a. Describe anamoly detection with example.
BL2
b. Discuss how Image denoising is implemented.
6 Differentiate RNN and CNN(April/2021) BL6
7 List out the advantages and disadvantages of sparse auto encoders. BL6
8 Describe the common application of NLP. (JAN/2022) BL6
9 Discuss the two major obstacles of RNN:vanishing gradient and exploding gradient. BL3
10 Draw the framework of CNN which helps in Natural language Processing.
UNIT V/PART C
1 Briefly explain the architecture of Autoencoder along with different autoencoders. BL2
2 What is RNN?What is the Need for RNN?Explain Bi-directional RNN. BL2
3 Explain the structure of stochastic auto encoder with step by step process. BL4
4 Describe Contractive Encoders with penalty term and the slope for the reconstruction of data.
(JAN/2021) BL2
5 Discuss in detail how the input is encoded and decoded along with loss function. BL4

St. Joseph’s Institute of Technology 20

Downloaded by swati zindani