Font size
WorksheetsGenerative AI MCQ Quiz
Total questions: 75
Worksheet time: 48mins
What is the primary objective of a GAN's generator?
To classify images
To detect anomalies
To generate realistic data that can fool the discriminator
To calculate loss functions
Which of the following is NOT a role of the discriminator in a GAN?
Distinguish real data from fake data
Provide gradient feedback to the generator
Generate new data samples
Improve over time to detect better fakes
What kind of loss is used in the original GAN paper for the generator?
Mean Squared Error
Binary Cross-Entropy
Hinge Loss
Wasserstein Loss
What is 'mode collapse' in GANs?
The generator learns the entire data distribution
The discriminator becomes too weak
The generator produces limited types of outputs
The training fails due to overfitting
Which of the following helps in improving GAN training stability?
Decreasing the number of epochs
Using ReLU in the discriminator
One-sided label smoothing
Removing the discriminator
What architecture does DCGAN use in place of fully connected layers?
RNN layers
LSTM layers
Convolutional layers
Transformer blocks
What does a transposed convolution layer do in a GAN generator?
Reduces spatial resolution
Acts as a pooling layer
Upsamples feature maps
Computes gradients
In a Conditional GAN, how is the class label typically fed into the generator?
As an output
Ignored
Concatenated with the input noise
Used to initialize weights
What is the key idea behind CycleGAN?
Supervised learning from labeled pairs
Learning mappings using conditional labels
Unsupervised image-to-image translation using cycle consistency
Reducing GAN parameters
Pix2Pix is best described as:
A style transfer GAN without supervision
A paired image-to-image translation GAN
A text-to-image generator
A classifier based on adversarial loss
What is one benefit of using Wasserstein loss in GANs?
It reduces training time
It eliminates the need for a discriminator
It provides more stable gradients for training
It increases mode collapse
Which component of StyleGAN is primarily responsible for style control?
Discriminator
Generator output layer
Mapping network
Latent vector z
What does the term 'scalability' in BigGAN mainly refer to?
Reducing model parameters
Increasing batch size and model capacity
Shrinking dataset size
Using transformers
How does latent space interpolation work in GANs?
By averaging discriminator outputs
By linearly combining latent vectors
By modifying loss values
By reducing input noise
Which GAN variant is most suited for unpaired image translation tasks?
DCGAN
CycleGAN
Pix2Pix
BigGAN
What is the primary objective of a GAN's generator?
Classify real vs. fake data
Minimize reconstruction error
Generate data that can fool the discriminator
Encode data into latent space
What does GAN stand for?
Generalized Adversarial Network
Generative Adversarial Network
Generative Analysis Network
General Artificial Network
Which two components make up a GAN?
Encoder and Decoder
Generator and Discriminator
Classifier and Regressor
Generator and Encoder
In a GAN, what type of learning is primarily used?
Supervised Learning
Unsupervised Learning
Reinforcement Learning
None of the above
In a GAN, what type of learning is primarily used?
Supervised Learning
Unsupervised Learning
Reinforcement Learning
Self-supervised Learning
What loss function is commonly used in GANs?
Cross-entropy loss
Mean squared error
Hinge loss
Triplet loss
What is the primary objective of a GAN's generator?
To classify images
To detect anomalies
To generate realistic data that can fool the discriminator
To calculate loss functions
Which of the following is NOT a role of the discriminator in a GAN?
Distinguish real data from fake data
Provide gradient feedback to the generator
Generate new data samples
Improve over time to detect better fakes
What kind of loss is used in the original GAN paper for the generator?
Mean Squared Error
Binary Cross-Entropy
Hinge Loss
Wasserstein Loss
What is 'mode collapse' in GANs?
The generator learns the entire data distribution
The discriminator becomes too weak
The generator produces limited types of outputs
The training fails due to overfitting
Which of the following helps in improving GAN training stability?
Decreasing the number of epochs
Using ReLU in the discriminator
One-sided label smoothing
Removing the discriminator
What architecture does DCGAN use in place of fully connected layers?
RNN layers
LSTM layers
Convolutional layers
Transformer blocks
What does a transposed convolution layer do in a GAN generator?
Reduces spatial resolution
Acts as a pooling layer
Upsamples feature maps
Computes gradients
In a Conditional GAN, how is the class label typically fed into the generator?
As an output
Ignored
Concatenated with the input noise
Used to initialize weights
What is the key idea behind CycleGAN?
Supervised learning from labeled pairs
Learning mappings using conditional labels
Unsupervised image-to-image translation using cycle consistency
Reducing GAN parameters
Pix2Pix is best described as:
A style transfer GAN without supervision
A paired image-to-image translation GAN
A text-to-image generator
A classifier based on adversarial loss
Which technique helps mitigate mode collapse?
Batch Normalization
Gradient Penalty
Dropout
Data Augmentation
What is a common cause of instability in GAN training?
Learning rate mismatch
Overfitting of the discriminator
Insufficient data
All of the above
Which architecture uses convolutional layers for image generation?
DCGAN
CycleGAN
StyleGAN
All of the above
Which GAN variant is known for high-resolution image synthesis?
BigGAN
CycleGAN
Pix2Pix
StyleGAN
What is the main advantage of StyleGAN over traditional GANs?
Faster training
Better image resolution
Improved control over generated features
Simpler architecture
Which task is CycleGAN particularly suited for?
Image classification
Image-to-image translation without paired data
Text generation
Time-series forecasting
What is the primary use case of Pix2Pix?
Image super-resolution
Image-to-image translation with paired data
Text-to-image generation
Anomaly detection
In StyleGAN, what does the mapping network do?
Maps latent vectors to styles
Generates images
Classifies images
None of the above
What is latent space manipulation in GANs?
Modifying the training data
Altering the generator's weights
Interpolating between latent vectors to generate new images
Adjusting the discriminator's parameters
What is the primary advantage of using Deep Convolutional GANs (DCGANs) over traditional GANs?
Faster training convergence
Improved image resolution
Enhanced stability during training
Ability to generate 3D images
In CycleGAN, what is the purpose of the cycle consistency loss?
To ensure the generator learns to produce diverse outputs
To prevent mode collapse
To ensure that an image translated to another domain can be translated back to the original domain
To penalize the discriminator for making incorrect classifications
Which of the following is a key feature of StyleGAN's generator architecture?
Use of residual blocks for feature refinement
Progressive growing of layers during training
Mapping of latent vectors to an intermediate latent space before image generation
Incorporation of attention mechanisms for feature selection
Which of the following techniques is commonly used to address mode collapse in GANs?
Batch normalization
Two time-scale update rule (TTUR)
Data augmentation
Early stopping
Which loss function is commonly used in Wasserstein GANs (WGANs) to improve training stability?
Binary cross-entropy loss
Mean squared error loss
Earth Mover's Distance (Wasserstein loss)
Hinge loss
Which metric is commonly used to evaluate the quality of images generated by GANs?
Mean squared error (MSE)
Fréchet Inception Distance (FID)
Precision and recall
Cross-entropy loss
What is the primary difference between a standard GAN and a Conditional GAN (cGAN)?
What is the primary difference between a standard GAN and a Conditional GAN (cGAN)?
cGANs use a different loss function
cGANs incorporate additional information (like labels) into both the generator and discriminator
cGANs do not use a discriminator
cGANs generate images without any input noise
Which of the following is a characteristic of BigGAN?
It uses a smaller network architecture for faster training
It is designed for high-resolution image generation with improved scalability
It operates on a single GPU for training
It focuses on generating text data rather than images
Which of the following is a key difference between CycleGAN and Pix2Pix?
CycleGAN requires paired images for training, while Pix2Pix does not
CycleGAN uses a cycle consistency loss to ensure bidirectional translation, while Pix2Pix uses paired images for supervised learning
CycleGAN is used for text generation, while Pix2Pix is used for image generation
CycleGAN uses a single generator, while Pix2Pix uses multiple generators
Which GAN variant introduces a Wasserstein loss to improve training stability?
CGAN
WGAN
DCGAN
CycleGAN
In Conditional GANs (cGANs), what additional information is fed to both the generator and discriminator?
Noise
Labels
Latent vectors
Loss values
What innovation does CycleGAN introduce to perform image-to-image translation without paired data?
True or False: The Wasserstein distance is more stable than JS-divergence in the context of GANs.
What is the purpose of using gradient penalty in WGAN-GP?
Which GAN variant is best suited for high-resolution image synthesis?
CycleGAN
WGAN
Progressive Growing GAN
InfoGAN
InfoGAN is designed to:
Translate images from one domain to another
Learn disentangled and interpretable latent representations
Improve training stability
Enhance discriminator performance
Which loss function does StyleGAN use to improve perceptual quality?
MSE Loss
Hinge Loss
Perceptual Loss
Fréchet Inception Distance (FID)
In WGAN, what key mathematical property is enforced to ensure the Wasserstein distance is valid?
Symmetry of the generator output
Lipschitz continuity of the discriminator
Orthogonality of latent vectors
Convexity of the loss function
Which technique is used in WGAN-GP to enforce the 1-Lipschitz constraint without weight clipping?
Spectral Normalization
Gradient Penalty
Batch Normalization
Dropout Regularization
What is a common symptom of mode collapse in GAN training?
Discriminator outputs near 0.5 for all inputs
Generator produces diverse but inaccurate samples
Generator produces identical or very similar outputs
Training loss of the generator increases steadily
The Fréchet Inception Distance (FID) is used to evaluate GANs by comparing:
Pixel-wise distances between real and generated images
Log-likelihood of real vs generated data
Feature distributions of real and generated images in a pretrained network
Number of distinct modes captured by the generator
Which of the following GAN models is best known for high-quality face generation and style transfer between image features?
CycleGAN
WGAN
StyleGAN2
InfoGAN
In StyleGAN, the intermediate latent space W is introduced to:
Increase randomness of the generated samples
Improve disentanglement of semantic features
Reduce training time
Enforce mode collapse
What is the role of the mapping network in StyleGAN?
Projects images to labels
Converts random noise z to intermediate latent code w
Calculates perceptual loss
Discriminates real vs fake images
In Progressive Growing of GANs (PGGAN), what is the primary advantage of growing the network over time?
Lower computation cost during inference
Reduced number of training epochs
Stability and quality improvement for high-resolution image generation
Elimination of discriminator collapse
Which GAN model introduces cycle-consistency loss to allow translation between domains without paired data?
Pix2Pix
CycleGAN
BigGAN
DCGAN
The discriminator in a GAN may become too powerful during training, leading to:
Mode expansion
Generator gradient vanishing
Faster convergence
Improved diversity
Spectral Normalization in GANs is used to:
Enforce sparsity in weights
Limit the Lipschitz constant of the discriminator
Reduce mode collapse
Normalize generator output
Which of the following loss functions is used in Hinge GANs?
Binary Cross Entropy
L2 Loss
Hinge Loss
Wasserstein Loss
In BigGAN, which of the following is a major enhancement for high-quality image generation?
Unpaired image translation
Conditional Batch Normalization
Cycle-consistency loss
Transformer blocks
One of the goals of InfoGAN is to:
Reduce training time
Improve adversarial loss
Learn interpretable and disentangled latent variables
Transfer image style between two domains
Why is training GANs considered a non-convex optimization problem?
It uses dropout in both networks
The objective function is linear in the generator
The loss surface has multiple saddle points and local minima
Discriminator is always convex
Which of the following techniques is commonly used to stabilize GAN training?
Dropout
Batch Normalization
Label Smoothing
Both b and c
What distinguishes GANformer from traditional GANs?
Use of transformers for hierarchical generation
Progressive upscaling
Mutual information maximization
Graph-based latent encoding
