UNIT-1
1. Explain the McCulloch-Pitts (MCP) neuron model, detailing its components and how it
operates with thresholding logic. Discuss its historical significance in the development of
artificial neural networks.
2. Describe the Perceptron Learning Algorithm (PLA). Outline the steps involved in its iterative
learning process, including the weight update rule.
3. Define linear separability in the context of classification. Illustrate with examples of both
linearly separable and non-linearly separable datasets.
4. State and explain the Perceptron Convergence Theorem. What are the conditions under which
the Perceptron Learning Algorithm is guaranteed to converge?
5. Compare and contrast the structure and function of a biological neuron with that of an
artificial computational unit like the Linear Perceptron.
6. Briefly introduce the concept of Image Classification. How do the foundational ideas of
computational units and linear separability relate to basic approaches in image classification?
2MARKS
1. 1.What is the primary function of the thresholding logic in a McCulloch-
Pitts unit?
2. Briefly define the concept of a "computational unit" in the context of
artificial neural networks.
3. What is the key difference between a linearly separable and a non-
linearly separable dataset?
4. In the Perceptron Learning Algorithm, when does a weight update occur?
5. What is the significance of the Perceptron Convergence Theorem?
6. List two essential components of a Linear Perceptron.
UNIT-2
1. Explain the architecture of a Multilayer Perceptron (MLP), detailing the roles
of its different layers (input, hidden, output) and how they contribute to its
functionality.
2. Describe the Backpropagation algorithm in detail, explaining how it efficiently
computes gradients for weight updates in a feedforward neural network.
3. Discuss the problem of vanishing/exploding gradients in deep neural networks
and explain how greedy layer-wise pre-training can mitigate this issue.
4. Explain the concept of regularization in neural networks. Describe at least two
common regularization techniques (e.g., L1/L2 regularization, dropout) and
their benefits.
5. What is an Autoencoder? Describe its architecture and explain its primary
purpose, providing an example of its application.
6. Explain the process of Gradient Descent as an optimization algorithm in the
context of training neural networks. Discuss its role in minimizing the loss
function.
2MARKS
1. What is the primary function of a hidden layer in a Multilayer Perceptron?
2. Briefly define the role of Gradient Descent in training neural networks.
3. What problem does Backpropagation primarily solve?
4. Name two common regularization techniques used in neural networks.
5. What is the main objective of an Autoencoder?
6. Why are deep neural networks generally more difficult to train than shallower
ones?
7. What is the core idea behind greedy layer-wise training?
8. What is the purpose of "architecture design" in feedforward networks?
UNIT-3
1. Explain the working of the Adam optimizer and compare it with RMSProp
and Adagrad.
2. Discuss the significance of second-order optimization methods in training
neural networks. How do they differ from first-order methods?
3. What is the saddle point problem in neural network training? How does it
affect the convergence of optimization algorithms?
4. Describe the concept of Nesterov Accelerated Gradient (NAG) and how it
improves over standard momentum-based methods.
5. What are the key differences between dropout and drop connect regularization
methods? How do they help prevent overfitting?
6. Explain the role of Batch Normalization in neural networks and discuss its
impact on training speed and performance.
2MARKS
1. What is the main idea behind the Adagrad optimization algorithm?
2. How does RMSProp handle the learning rate differently from Adagrad?
3. What is the purpose of using dropout in neural networks?
4. Define the saddle point problem in the context of neural network training.
5. What is the key advantage of using batch normalization?
6. Name one difference between first-order and second-order optimization methods.
UNIT-4
1. Explain the Backpropagation Through Time (BPTT) algorithm used for
training Recurrent Neural Networks.
2. Describe the architecture and working of Long Short-Term Memory (LSTM)
networks. How do they overcome the vanishing gradient problem?
3. Compare and contrast Gated Recurrent Units (GRUs) with LSTMs in terms of
structure and performance.
4. What are Bidirectional RNNs and Bidirectional LSTMs? Discuss their
advantages and applications.
5. Explain the architecture of LeNet and its significance in the development of
Convolutional Neural Networks (CNNs).
6. Describe the key features and innovations introduced by AlexNet that led to its
success in image classification.
7. What is a Restricted Boltzmann Machine (RBM)? Explain its structure
and training process.
8. Discuss the role of Gibbs Sampling and MCMC methods in training
RBMs and Deep Boltzmann Machines.
2 MARKS
What is the purpose of Backpropagation Through Time (BPTT) in RNNs?
Name the three gates used in an LSTM cell.
What is the main difference between a GRU and an LSTM?
Why are Bidirectional RNNs useful in sequence modeling tasks?
Who proposed the LeNet architecture and for what task was it originally designed?
List two major innovations introduced by AlexNet.
What does the term 'Restricted' refer to in Restricted Boltzmann Machines?
What is the role of Gibbs Sampling in training RBMs?
Unit-5
1. Explain the Transformer architecture and its role in Natural Language Processing (NLP).
2. Describe the application of Variational Autoencoders (VAEs) in image synthesis and their
benefits.
3. What are the applications of Transformers in speech processing? Explain their advantages.
4. How can Transformers and GPT be used for Multimodal AI tasks? Provide examples.
5. What are the ethical challenges associated with large generative AI models? Suggest
mitigation strategies.
6. Explain the generative capabilities of GPT models and their applications in NLP.
What is the primary function of Variational Autoencoders (VAEs) in deep learning?
What is the key innovation of Transformers in natural language processing?
What is the primary application of GPT models in natural language processing?
How do VAEs differ from traditional autoencoders?
What are some applications of Transformers in computer vision?
What role do attention mechanisms play in Transformers?
