Deep Learning Viva Questions and Simple Answers
1. What is Deep Learning? How is it different from Machine Learning?
Answer: Deep Learning is a type of Machine Learning using neural networks with many layers. It
learns features automatically.
2. What are neurons and perceptrons?
Answer: Neurons process inputs in a neural network. A perceptron is a simple single-layer neural
network.
3. Explain the architecture of a neural network.
Answer: It has input, hidden, and output layers. Each layer contains neurons connected with
weights.
4. What are activation functions? Name a few commonly used ones.
Answer: They decide if a neuron activates. Examples: ReLU, Sigmoid, Tanh, Softmax.
5. What is the role of weights and biases in a neural network?
Answer: Weights show connection strength. Bias helps shift the output of the activation function.
6. What is forward and backward propagation?
Answer: Forward gives output from input. Backward updates weights to reduce error.
7. What is the cost/loss function? Give examples.
Answer: It measures error. Examples: MSE, Cross-Entropy.
8. What is the difference between Batch Gradient Descent and Stochastic Gradient Descent
(SGD)?
Answer: Batch uses all data; SGD uses one sample; Mini-batch uses a few samples.
9. Why is normalization important in deep learning?
Answer: It makes training faster and stable by scaling inputs.
10. What is overfitting and how can it be prevented?
Answer: Model works well on training data but not new data. Prevent using dropout, early stopping.
11. What are the types of neural networks?
Answer: Feedforward, CNN, RNN, Autoencoders, GANs.
12. What is a Convolutional Neural Network (CNN)? Explain its layers.
Answer: CNNs are for image data. Layers: Convolution, Pooling, Fully Connected.
13. What are filters/kernels in CNNs?
Answer: Filters detect features like edges by sliding over images.
14. What is pooling? Explain Max Pooling and Average Pooling.
Answer: Pooling reduces size. Max: largest value; Average: average value.
15. What is a ReLU function and why is it preferred?
Answer: ReLU = max(0,x). It is fast and avoids vanishing gradients.
16. What is dropout and how does it help in training?
Answer: Randomly turns off neurons to prevent overfitting.
17. What are vanishing and exploding gradients?
Answer: Vanishing: too small; Exploding: too large. Both hurt learning.
18. What is an epoch, batch size, and iteration?
Answer: Epoch: full pass; Batch size: samples per update; Iteration: one update.
19. What is the role of the learning rate?
Answer: It controls how fast weights are updated.
20. Explain the concept of transfer learning.
Answer: Using a trained model for a new task to save time.
21. What are RNNs and how are they different from CNNs?
Answer: RNNs work on sequences, CNNs on images.
22. Explain the problem of vanishing gradient in RNNs and how LSTM solves it.
Answer: RNNs forget long-term info. LSTM remembers using gates.
23. What are LSTM and GRU? How are they different?
Answer: Both remember sequences. GRU is simpler than LSTM.
24. What is attention mechanism in deep learning?
Answer: It helps focus on important input parts.
25. Explain the Transformer architecture briefly.
Answer: It uses attention, not recurrence, for sequence data.
26. What are Generative Adversarial Networks (GANs)?
Answer: Two networks: Generator makes fake data; Discriminator checks it.
27. What is the role of the discriminator and generator in GANs?
Answer: Generator creates fake data, Discriminator checks if it's real.
28. What are autoencoders? What are their applications?
Answer: They compress and rebuild data. Used in noise removal, detection.
29. How do you choose the number of layers and neurons in a deep network?
Answer: Trial and error based on problem and tuning.
30. How do you evaluate the performance of a deep learning model?
Answer: Use accuracy, precision, recall for classification; MSE for regression.
31. Which deep learning frameworks have you used?
Answer: Common ones: TensorFlow, Keras, PyTorch.
32. How do you handle imbalanced data in deep learning?
Answer: Use class weights, oversampling, or SMOTE.
33. What are pretrained models? Name a few.
Answer: Trained on large data. Examples: VGG, ResNet, BERT.
34. What are some common metrics used for classification and regression in DL?
Answer: Classification: Accuracy, F1. Regression: MSE, MAE.
35. What is the difference between softmax and sigmoid functions?
Answer: Softmax: multi-class; Sigmoid: binary classification.
36. What is backpropagation through time (BPTT)?
Answer: It trains RNNs by unrolling them through time.
37. What is the role of regularization (L1, L2) in deep learning?
Answer: They reduce overfitting by adding penalty terms.
38. How do you tune hyperparameters in a DL model?
Answer: Use grid search, random search, or tuning libraries.
39. How is a CNN used in image classification tasks?
Answer: CNN extracts features and classifies them.
40. What are the limitations of deep learning?
Answer: Needs lots of data, compute; hard to explain results.
