DEEP LEARNING

Part of a biological neuron that receives signals from other neurons is called

1. Axon
2. Dendrite
3. Soma
4. Synapse

In MLP training, which algorithm computes gradients layer by layer?

1. Forward propagation
2. Back-propagation
3. Hebbian learning
4. K-means clustering

____________ helps back-propagation avoid overfitting?

1. Increasing learning rate
2. Early stopping and regularization
3. Large momentum only
4. Removing hidden layers

Optimization technique that uses a moving average of the gradient to smooth updates and navigate ravines faster is

1. RMSProp
2. Momentum
3. Batch Gradient Descent
4. AdaGrad

What is the fundamental input to the Generator network in a GAN?

1. Real training data
2. Labeled class information
3. Random noise vector
4. Output from the Discriminator

When the generator produces limited variety of samples, the problem is called

1. Overfitting
2. Mode collapse
3. Gradient vanishing
4. Underfitting

Which mechanism allows the decoder to access encoder outputs?

1. Masked Self-Attention
2. Positional Encoding
3. Cross-Attention
4. Feed-Forward Networks

Which model is famous for being "Bidirectional"

1. GPT
2. BERT
3. RNN
4. LSTM

In CNNs, what is the effect of using a stride greater than 1 during convolution?

1. Increases output feature map size
2. Reduces output feature map size
3. Adds more channels
4. Normalizes the input

GoogLeNet introduced ______ module to improve computational efficiency

1. Residual connection
2. Pooling pyramid
3. Inception module
4. Dense block

Explain the mathematical model of an artificial neuron with commonly used activation functions.

Describe the architecture and need of a Multilayer Perceptron. Discuss the role of hidden layers in solving complex non-linear problems.

Explain the working principle of the back-propagation algorithm in training multilayer neural networks.

Design a suitable multilayer neural network architecture to solve the XOR problem.

Consider training a simple linear regression model y = wx + b (with no bias, b = 0, for simplicity, so y = wx) using Mean Squared Error (MSE) loss. Given the dataset as (x, y)={(1, 2), (2, 4), (3, 5)}, initial parameter w₀ = 0 and learning rate η = 0.1. Define the Batch MSE Cost Function J(w). Calculate the gradient dJ/dw at w₀ = 0. Perform one update of the weight w using Batch Gradient Descent.

Explain the concept of parameter norm penalties in deep learning and derive the modified cost function when L2 regularization is applied.

Describe various image augmentation techniques to preserve class labels while increasing dataset diversity.

Explain the convolution operation in CNNs with mathematical formulation. Discuss the purpose of pooling layers in CNN architectures.

Describe the architecture of LeNet in detail. Explain the contribution of LeNet to early successes in handwritten digit recognition.

Explain the concept of residual learning in ResNet and derive the idea behind skip connections.

Analyze the vanishing gradient problem in RNNs using mathematical reasoning.

Describe the role of the generator and discriminator in a Generative Adversarial Networks.

Explain the architecture of BERT and its use of masked language modeling and next sentence prediction during pre-training.

Discuss the relationship between Transformer architectures and autoencoders in representation learning.

Back Propagation and Differentiation

RMSProp

Bidirectional RNNs

Image Augmentation