1. What does the term "GAN" stand for in deep learning?
- Generalized Artificial Neural networks
- Generative Adversarial Networks
- Graphical Activation Nodes
- Gradient Approximation Network
Generative Adversarial Networks (GANs) AI and Deep Learning MCQ Exam
Explore key concepts, architecture and applications of GANs in image generation and deep learning. Perfect for students and AI professionals.
📌 Important Exam Instructions
- ✅ This is a free online test. Do not pay anyone claiming otherwise.
- 📋 Total Questions: 40
- ⏳ Time Limit: 40 minutes
- 📝 Marking Scheme: +1 for each correct answer. No negative marking.
- ⚠️ Avoid page refresh or closing the browser tab to prevent loss of test data.
- 🔍 Carefully read all questions before submitting your answers.
- 🎯 Best of Luck! Stay focused and do your best. 🚀
Time Left (min): 00:00
2. The goal of the Generator in a GAN is to:
- Optimize hyperparameters automatically
- Classify data into categories
- Reduce the computational complexity of models
- Generate data that is indistinguishable from real data
3. What is the primary role of the Discriminator in a GAN?
- To generate synthetic data
- To differentiate between real and generated data
- To increase the accuracy of classification
- To reduce loss in backpropagation
4. GANs are primarily used in which of the following applications?
- Data encryption
- Image synthesis and enhancement
- Statistical regression
- Signal compression
5. Which loss function is commonly used in the training of GANs?
- Mean Squared Error
- Cross-Entropy Loss
- Binary Cross-Entropy Loss
- Hinge Loss
6. The training of GANs is often described as a:
- Zero-sum game
- Cooperative learning process
- Multi-label classification task
- Optimization problem with fixed weights
7. What is "mode collapse" in GANs?
- The generator produces limited diversity in outputs
- The discriminator fails to learn patterns
- The training process halts prematurely
- The generator stops generating any outputs
8. Conditional GANs (cGANs) allow:
- Training on unlabeled data
- Controlling the output based on input labels
- Generating only text data
- Higher accuracy in classification tasks
9. What is the key difference between a Vanilla GAN and a Wasserstein GAN (WGAN)?
- WGAN uses the Wasserstein distance metric for training
- Vanilla GANs require less computational power
- WGAN uses multiple discriminators
- Vanilla GANs do not use a loss function
10. Which activation function is commonly used in the output layer of a GAN's generator?
- ReLU
- Sigmoid
- Tanh
- Softmax
11. Which of the following is a real-world use case of GANs?
- Fraud detection in financial systems
- Generating realistic faces from random noise
- Optimizing supply chain logistics
- Enhancing the performance of linear regression models
12. GAN training is computationally expensive because:
- Both the generator and discriminator are updated simultaneously
- It requires a vast amount of labeled data
- It relies heavily on manual tuning of hyperparameters
- It involves solving a multi-objective optimization problem
13. What is the primary challenge in training GANs?
- Balancing the generator and discriminator performance
- Lack of sufficient training data
- High variance in model predictions
- Over-reliance on GPU hardware
14. What does "latent space" refer to in the context of GANs?
- The feature representation learned by the generator
- The training data distribution
- The hyperparameter space of the model
- The output space of the discriminator
15. StyleGAN is a specialized GAN architecture used for:
- Data compression
- Generating highly realistic images with customizable features
- Optimizing neural network weights
- Creating video content
16. CycleGAN is primarily used for:
- Enhancing audio signals
- Generating text data
- Translating images from one domain to another without paired examples
- Improving video resolution
17. What is the typical input to a GAN's generator?
- A sequence of tokens
- Labeled data
- Pre-trained embeddings
- Random noise vector
18. GANs are unsuitable for tasks requiring:
- Supervised classification
- Data augmentation
- Image restoration
- Video frame prediction
19. Pix2Pix GAN requires:
- Paired training data
- Unlabeled data
- Pre-trained weights
- Real-time feedback during training
20. What is the primary advantage of Wasserstein GANs (WGANs)?
- Lower memory usage
- Faster inference speed
- Improved stability during training
- Simplified architecture
21. Which optimization algorithm is commonly used in GAN training?
- Gradient Descent
- Adam Optimizer
- Stochastic Gradient Descent
- Newton's Method
22. What is the key feature of Progressive GANs?
- Incremental training with growing image resolution
- Training with reduced computational cost
- Use of hybrid neural networks
- Faster training on small datasets
23. The discriminator loss in a GAN measures:
- The ability to distinguish between real and fake data
- The quality of generated data
- The performance of the generator
- The variance in training data
24. BigGAN improves on traditional GANs by:
- Reducing model complexity
- Employing unsupervised training techniques
- Using larger batch sizes and higher-capacity models
- Using fewer layers in the generator
25. What does "mode balancing" in GANs aim to achieve?
- Equal representation of all modes in generated data
- Reduction in model complexity
- Improved discriminator loss
- Faster convergence during training
26. DCGAN stands for:
- Deep Convolutional Generative Adversarial Network
- Distributed Cognitive GAN
- Dynamic Convolution GAN
- Differentiable Convolution GAN
27. Which dataset is often used for benchmarking GANs?
- MNIST
- ImageNet
- CIFAR-10
- All of the above
28. GANs have been used in deepfake creation because:
- They can generate realistic images and videos
- They require minimal training data
- They rely solely on text data
- They use supervised learning
29. Which of these is NOT a type of GAN?
- SRGAN
- DCGAN
- RNN-GAN
- WGAN
30. What role does the ReLU activation function play in GANs?
- It helps in non-linear transformations in the generator
- It calculates discriminator loss
- It standardizes inputs
- It minimizes overfitting
31. GANs struggle with generating:
- Realistic images
- Sequential text data
- Complex image transformations
- High-resolution videos
32. What does the term "generator loss" indicate in GAN training?
- The quality of data generated by the generator
- The failure of the discriminator
- Overfitting during training
- Latent space optimization
33. Transfer learning can be used in GANs to:
- Simplify discriminator architecture
- Train from scratch on large datasets
- Enhance performance on related tasks
- Reduce memory requirements
34. Which of the following represents a real-world application of GANs?
- Art generation and style transfer
- Text classification
- Signal processing optimization
- Web scraping
35. Why are spectral normalization techniques used in GANs?
- To increase the generator’s capacity
- To stabilize training by controlling discriminator weight updates
- To reduce memory usage during training
- To simplify the network architecture
36. The inception score is a metric for evaluating GANs based on:
- The quality and diversity of generated images
- Training efficiency
- Model complexity
- Computational cost
37. What is the key benefit of using batch normalization in GANs?
- It accelerates the convergence of the generator
- It reduces the need for labeled data
- It helps prevent overfitting in the discriminator
- It stabilizes training by normalizing layer inputs
38. What type of loss function is typically used for the discriminator in a traditional GAN?
- Mean Squared Error Loss
- Binary Cross-Entropy Loss
- Hinge Loss
- Categorical Cross-Entropy Loss
39. Which of the following is a commonly used evaluation metric for the performance of a GAN?
- F1-score
- Precision
- Inception score
- Mean Squared Error
40. In the context of GANs, what is meant by "adversarial training"?
- The process of training the generator to be adversarial to the discriminator
- The simultaneous training of two models (generator and discriminator) to compete against each other
- Using adversarial attacks to improve model robustness
- Training the models using only unsupervised learning