A.
Deep Feedforward Neural Networks (DNNs)
1. What is a deep feedforward neural network?
2. What is the difference between a shallow and a deep network?
3. What is vanishing gradient problem?
4. What is exploding gradient problem?
5. How does weight initialization affect training in deep networks?
6. What is the role of hidden layers in DNNs?
7. Why is ReLU preferred over sigmoid in deep networks?
8. What are the disadvantages of ReLU?
9. Explain backpropagation in a deep network.
10. Why do deeper networks have higher representational power?
11. What is layer normalization and when is it used?
12. How does adding more layers affect bias and variance?
13. When to use softmax activation function in output layer? Justify.
B. Hyperparameter Tuning & Optimization
1. What are hyperparameters? Give examples.
2. Higlight on possible approaches to identify the range values for each
hyperparameters?
3. Difference between parameters and hyperparameters.
4. What is learning rate? What happens if it is too small or too large?
5. What are batch size and epoch?
6. What is the difference between SGD and mini-batch gradient descent?
7. What is early stopping?
8. Compare Grid Search and Random Search.
9. What is Bayesian Optimization in hyperparameter tuning?
10. Why is tuning batch size considered important?
C. Convolutional Neural Networks (CNNs)
1. What is convolution?
2. What is the purpose of filters/kernels?
3. What is padding? Why is it used?
4. What is stride in convolution?
5. How does max pooling differ from average pooling?
6. Why does CNN work better than DNNs for image inputs?
7. What is a feature map, and how is it generated during convolution?
8. Explain 1×1 convolution and its use.
9. What is depthwise separable convolution?
10. Explain difference between 2D and 3D convolution.
11. What is dilated convolution?
12. How does CNN extract hierarchical features?
13. Why do deeper CNN layers detect more abstract features?
14. Explain the concept of CNN overfitting. What techniques are commonly used to reduce
overfitting in CNNs?
15. How to calculate the number of parameters of any CNN model.
16. Discuss the relationship between deep layers and its parameters.
17. Discuss the different methods of upsampling in Convolutional Neural Networks,
providing an example for each method.
18. Given the AlexNet architecture, calculate the feature map size, weights, bias, and total
parameters for each layer, considering the kernel size, stride, padding, and filter
information.
19. Perform a 2D convolution operation on the following 5 × 5 input matrix using the given
3 × 3 kernel. Assume stride = 1 and no padding (valid convolution)
Tasks:
a) Compute the convolution output step-by-step.
b) Write the final 3 × 3 output feature map.
c) Explain how the output dimension is calculated.
D. Transfer Learning using Pretrained CNN Models
1. What is transfer learning?
2. Why is transfer learning useful?
3. What is fine-tuning in transfer learning?
4. What happens when you freeze layers in a pretrained model?
5. Name any 5 popular pretrained CNN models.
6. What is feature extraction vs. fine-tuning?
7. Why is ImageNet commonly used for pretraining?
8. What is domain mismatch problem in transfer learning?
9. Why does fine-tuning only last layers give good results for small datasets?
E. Recurrent Neural Networks (RNNs)
1. What is a recurrent neural network?
2. What is the vanishing gradient problem in RNNs and how do LSTM/GRU solve it?
3. What are the limitations of a simple RNN?
4. What is sequence-to-sequence model?
5. What is the role of hidden state in RNNs?
6. What is long-term dependency problem?
7. Compare RNN, LSTM, and GRU.
8. Why are RNNs slower to train than CNNs?
Deep Learning
1. What is the capacity of a deep learning model, and why is it important?
2. List any four activation functions and Why activation functions required in Deep
Learning models?
3. List any five techniques that can be used to enhance the generalization capability of
deep learning models. Justify how each technique helps in improving the model’s
performance on unseen data.
4. Explain the significance and importance of deep learning models in contemporary
artificial intelligence and machine learning.
5. Discuss key advantages and applications that make deep learning an influential and
transformative paradigm in solving complex problems.
6. Elaborate challenges associated with deep learning and how ongoing research
addresses these issues.