Deep Learning & Neural Networks (NPMCE06) -
Question Paper with Answers
1. Define Machine Learning.
Machine Learning is a branch of AI that enables systems to learn from data and improve
performance without explicit programming.
2. What is Deep Learning?
Deep Learning is a subset of machine learning that uses multi-layer neural networks to learn
complex patterns.
3. What is a Perceptron?
A perceptron is the simplest artificial neuron used for binary classification.
4. What is Forward Propagation?
It is the process of passing input data through network layers to generate output.
5. What is Backpropagation?
It is an algorithm that updates weights by propagating error backward through the network.
6. Purpose of Activation Function?
It introduces non-linearity so neural networks can learn complex relationships.
7. Define ReLU.
ReLU(x)=max(0,x); it outputs zero for negative inputs and the input itself for positive values.
8. What is Overfitting?
Overfitting occurs when a model memorizes training data and performs poorly on unseen data.
9. What is CNN?
CNN is a neural network specialized for image and spatial data processing.
10. What is RNN?
RNN is a neural network designed for sequential data with memory of previous inputs.
11. History and Evolution of Deep Learning
Deep learning evolved from artificial neural networks. Key milestones include perceptrons,
backpropagation, deep belief networks, CNNs, RNNs, and modern transformer models.
12. Advantages and Challenges of Deep Learning
Advantages: automatic feature extraction, high accuracy, scalability. Challenges: overfitting, large
data requirements, computational cost, and explainability.
13. Architecture of a Deep Neural Network
A DNN consists of an input layer, multiple hidden layers, and an output layer. Hidden layers learn
hierarchical features.
14. Sigmoid vs Tanh vs ReLU
Sigmoid outputs 0-1, Tanh outputs -1 to 1, and ReLU outputs max(0,x). ReLU is most commonly
used in hidden layers.
15. Forward Propagation
Input data passes through neurons where weighted sums and activation functions are applied until
final output is produced.
16. Backpropagation Algorithm
The algorithm computes loss, calculates gradients using the chain rule, and updates weights using
gradient descent.
17. CNN Architecture
CNN contains convolution, activation, pooling, and fully connected layers for feature extraction and
classification.
18. RNN Applications and Limitations
Applications include language modeling and speech recognition. Limitations include vanishing
gradients and slow training.
Question 19 - Detailed Answer
Biological neurons inspire perceptrons. Perceptrons can implement AND, OR and NOT gates using
weights and bias. XOR cannot be solved by a single-layer perceptron because it is not linearly
separable; it requires hidden layers.
Question 20 - Detailed Answer
Hyperparameters include learning rate, batch size, epochs, hidden units and regularization settings.
Learning rate controls update size. Regularization prevents overfitting using techniques such as
dropout and L1/L2 penalties.
Question 21 - Detailed Answer
CNN architecture includes convolution layers for feature extraction, ReLU for non-linearity, pooling
for dimensionality reduction and fully connected layers for classification.
Question 22 - Detailed Answer
RNNs process sequential data using recurrent connections. They maintain hidden states across
time steps and are used in NLP, speech recognition and time-series analysis.