Font size
WorksheetsGenerative AI Quiz
Total questions: 125
Worksheet time: 1hrs 3mins
How do researchers evaluate the performance of Generative AI models?
By comparing generated outputs to predefined templates
By conducting user surveys and feedback sessions
By measuring similarity to real-world data distributions
By applying subjective judgments and opinions
What role does data play in the effectiveness of Generative AI models?
No role, as Generative AI does not rely on data
Limited role, with data serving as a minor input
Significant role, with data influencing model performance
Dominant role, with data determining model outcomes
How might Generative AI contribute to addressing societal challenges?
By exacerbating existing inequalities and injustices
By promoting transparency and fairness in decision-making
By fostering collaboration and community engagement
By limiting access to generative technologies
What ethical considerations are associated with the deployment of Generative AI?
Bias, privacy, and accountability
Transparency, fairness, and interpretability
Inclusivity, diversity, and accessibility
Innovation, creativity, and progress
What impact might Generative AI have on traditional employment roles?
No impact, as Generative AI is not suitable for real-world applications
Negative impact, leading to job displacement in certain sectors
Positive impact, creating new job opportunities across industries
Limited impact, confined to specific job functions
How does Generative AI contribute to innovation in various fields?
By restricting access to creative tools
By limiting experimentation and exploration
By fostering creativity and generating novel ideas
By promoting conformity and stagnation
What role does creativity play in Generative AI?
No role, as Generative AI relies solely on predefined rules
Limited role, with creativity primarily driven by human input
Significant role, with algorithms capable of creative output
Minimal role, with creativity confined to specific tasks
Which industries can benefit from the application of Generative AI techniques?
Healthcare, finance, and education
Manufacturing, agriculture, and transportation
Entertainment, marketing, and fashion
All of the above
What advantages does Generative AI offer in comparison to traditional content creation methods?
Faster production time and reduced costs
Higher quality output with minimal human intervention
Limited creativity and innovation
Dependence on manual labor for each creation
Which technique is used for generating new images based on artistic styles?
Variational Autoencoders (VAEs)
Deep Dream
Neural Style Transfer
Reinforcement Learning
What distinguishes generative models from discriminative models?
Generative models learn the underlying data distribution
Discriminative models generate new data samples
Generative models require labeled data for training
Discriminative models predict probabilities
Which of the following is NOT an application of Generative AI?
Image classification
Style transfer
Music generation
Text generation
What is the main principle behind Generative Adversarial Networks (GANs)?
Supervised learning
Unsupervised learning
Reinforcement learning
Adversarial training
What is a common application of generative chatbots?
Image captioning
Language translation
Conversation generation
Sentiment analysis
Which type of neural network is commonly used for generating sequences of data?
Convolutional Neural Networks (CNNs)
Recurrent Neural Networks (RNNs)
Generative Adversarial Networks (GANs)
Reinforcement Learning Networks (RLNs)
What is the primary focus of Generative AI?
Classifying data
Generating new data
Predicting outcomes
Optimizing algorithms
In which domain is generative AI NOT commonly applied?
Image synthesis
Music generation
Medical diagnosis
Text generation
Which type of neural network is commonly used in Generative AI?
Convolutional Neural Networks (CNNs)
Recurrent Neural Networks (RNNs)
Generative Adversarial Networks (GANs)
Reinforcement Learning Networks (RLNs)
What distinguishes Variational Autoencoders (VAEs) from other generative models?
They use reinforcement learning
They are unsupervised learning models
They learn a continuous latent space
They are based on convolutional networks
What does Generative AI focus on?
Replicating existing data
Creating new data
Analyzing data patterns
Optimizing algorithms
What role does the generator play in a Generative Adversarial Network (GAN)?
It discriminates between real and fake data
It generates new data samples
It optimizes the loss function
It creates noise for training
Which technique is used for transferring artistic styles onto images?
Variational Autoencoders (VAEs)
Deep Dream
Neural Style Transfer
Reinforcement Learning
What distinguishes generative models from discriminative models?
Generative models predict probabilities
Discriminative models generate new data
Generative models learn the joint probability distribution
Discriminative models create realistic images
Which of the following is NOT a common application of Generative AI?
Image generation
Music composition
Speech recognition
Text generation
What is the main goal of Generative Adversarial Networks (GANs)?
Image classification
Data compression
Generating realistic data
Language translation
What is the primary goal of an autoencoder?
Classify input data into labels
Compress and reconstruct input data
Generate adversarial examples
Predict time-series data
In the context of latent space, which of the following is true?
Latent space is the noise added during training
It represents a compressed representation of input data
It is the raw pixel space of images
It always has a higher dimension than input data
Which of the following is a common loss function used in training autoencoders?
Cross-entropy loss
Hinge loss
Mean squared error (MSE)
Huber loss
What is the primary purpose of KL Divergence in Variational Autoencoders (VAEs)?
To enforce data reconstruction
To increase training speed
To regularize the distribution of latent variables
To increase model capacity
What does the decoder do in an autoencoder architecture?
Maps input data to latent space
Reconstructs the input from latent space
Computes the loss
Adds noise to the input
In PyTorch, which module is commonly used to define a custom neural network?
torch.nn.Module
torch.utils.data
torch.TensorBoard
torch.nn.functional
Which of the following best describes the limitation of basic autoencoders in generative tasks?
They can't compress data
They do not learn a smooth latent space
They overfit easily
They use too much memory
What does VAE sampling involve?
Using deterministic latent variables
Sampling from a learned distribution (mean and variance)
Applying dropout to the encoder
Selecting the maximum activation in each layer
Which arithmetic operation is demonstrated in latent vector arithmetic?
Convolution
Addition and subtraction of vectors in latent space
Multiplication of feature maps
Max-pooling of features
What is the role of the Denoising Autoencoder (DAE)?
To generate new images
To classify noisy input data
To reconstruct clean data from noisy inputs
To enforce sparsity in outputs
In a VAE, the encoder outputs two vectors. What do they represent?
Loss and accuracy
Mean and log variance of the latent distribution
Input and output labels
Activation maps
Which library in PyTorch is used for automatic differentiation?
torch.optim
torch.autograd
torch.nn.functional
torch.utils.data
Which task is NOT typically handled by autoencoders?
Denoising images
Data compression
Classification
Feature extraction
What is one application of VAEs in the medical field?
Diagnosing infections
Synthesizing medical images for augmentation
Replacing doctors
Performing surgeries
Autoencoders can help reduce:
Learning rate
Memory usage
Dimensionality of data
Training epochs
Autoencoders are often used in recommendation systems for:
Label encoding
User embedding generation
Predicting weather
Data normalization
A use of autoencoders in NLP includes:
Syntax correction
Tokenization
Sentence embedding
Part-of-speech tagging
Which property is desirable in a good latent space?
Sparsity
Semantic smoothness
Noise amplification
High dimensionality
What happens when two points in latent space are interpolated?
Noise is reduced
Reconstruction is skipped
A meaningful transition between representations is created
Dropout is applied
Latent space helps with:
Overfitting
Model compression
Understanding data structure
Enhancing activation functions
Which model learns a latent distribution explicitly?
CNN
GAN
VAE
LSTM
Which characteristic is true for continuous latent variables?
They are binary
They allow smooth interpolation
They cannot be visualized
They need labels
Tied weights in an autoencoder mean:
Encoder and decoder use different parameters
Decoder reuses encoder weights
Latent space is doubled
Gradients are shared across epochs
Which component is NOT essential in an autoencoder?
Encoder
Decoder
Classifier
Bottleneck
Undercomplete autoencoders:
Have fewer hidden units than input size
Overfit the training data
Do not use backpropagation
Encode categorical variables only
The purpose of sparsity constraints in autoencoders is:
Increase accuracy
Encourage feature selectivity
Reduce training time
Eliminate noise
Overcomplete autoencoders:
Compress input
Use dropout aggressively
Have more hidden units than input size
Fail to generalize
L1 loss promotes:
Large errors
Smooth outputs
Sparsity in activations
Overfitting
Which loss is more sensitive to outliers?
L2 loss
Huber loss
MSE
Cross-entropy
A good reconstruction loss should:
Maximize overfitting
Minimize the difference between input and output
Add regularization noise
Ignore small differences
Custom loss functions in PyTorch require:
torch.nn.Linear
autograd.backward
Subclassing nn.Module
Defining forward logic with gradients
In PyTorch, backward() is used to:
Run forward pass
Update weights manually
Compute gradients
Stop training
Which PyTorch module is used to define custom neural networks?
torch.nn.Module
torch.nn.functional
torch.optim
torch.tensor
What does model.train() do in PyTorch?
Freezes weights
Switches model to training mode
Starts the optimizer
Evaluates the model
What is the purpose of torch.no_grad()?
Enforces gradient computation
Enables dropout
Disables gradient calculation
Trains the model
What is the optimizer used for in PyTorch?
Creating datasets
Saving the model
Updating model weights
Building layers
Which function is used to calculate loss in PyTorch?
loss()
compute_loss()
criterion()
nn.()
One limitation of autoencoders is:
They require supervised labels
They cannot compress data
They may memorize instead of generalizing
They are only usable on text
Why do autoencoders struggle with generative tasks?
They lack latent space constraints
They use too many layers
They are too simple
They rely on labels
What issue occurs when the latent space is not regularized?
Increased accuracy
Clear class boundaries
Poor generalization
Faster training
Autoencoders perform poorly when:
The latent space is small
The input data is low-dimensional
They are trained with labels
Reconstruction loss is high for training data
Vanilla autoencoders lack:
A decoder network
A latent space
A probabilistic framework
Activation functions
In a VAE, the encoder outputs:
A reconstructed image
One latent vector
Mean and log-variance vectors
Softmax probabilities
The decoder in a VAE uses:
Raw input data
Random noise
Samples from latent distribution
Training labels
Why do VAEs use the reparameterization trick?
To increase randomness
To compute gradients through stochastic sampling
To speed up decoding
To convert data to tensors
What is a key advantage of VAEs over standard autoencoders?
Higher compression
Better reconstruction
Structured latent space suitable for sampling
Faster training time
VAEs treat the latent space as:
Discrete
Randomized embeddings
A probability distribution
A loss function
Which two distributions are compared using KL Divergence in VAEs?
Input and output
Prior and posterior latent distributions
Noise and signal
Labels and predictions
A high KL Divergence value suggests:
Latent distribution is similar to the prior
Latent distribution diverges from the prior
Model is regularized
Loss is minimal
KL Divergence is zero when:
The decoder fails
Prior equals posterior
Input is noise
Output is binary
In VAE sampling, the latent vector is drawn from:
A uniform distribution
A one-hot vector
A Gaussian distribution
A categorical distribution
The reparameterization trick enables:
Manual weight updates
Computation of gradients through sampling
Saving models during training
Using categorical data
The random noise added during sampling in a VAE comes from:
The decoder
A learned distribution
A standard normal distribution
The output labels
VAE sampling enables:
Reconstruction without input
Classification of data
Generation of new samples
Data shuffling
Which operation is used in VAE sampling?
z = mean + noise
z = mean + std * epsilon
z = decoder(input)
z = encoder(output)
Latent vector arithmetic involves:
Normalizing pixel values
Adding and subtracting embeddings
Sorting datasets
Random dropout
In generative models, latent vector arithmetic can reveal:
Training loss
Hyperparameter tuning
Semantic relationships
Overfitting issues
What kind of structure is expected in a good latent space?
Linear separability
Sparsity
Discrete tokens
Constant variance
Latent vector arithmetic demonstrates that:
Decoders ignore noise
Latent space captures useful semantic features
Models overfit easily
Training data must be labeled
An example of vector arithmetic in VAEs is:
Adding noise to training labels
Walking in the latent space to interpolate between concepts
Removing layers from decoder
Increasing KL divergence
A Denoising Autoencoder is trained to:
Add noise to data
Learn a compressed vector
Reconstruct the original input from noisy input
Classify noisy inputs
DAEs help models to:
Memorize input features
Overfit noisy labels
Learn robust representations
Ignore noise completely
What kind of noise is commonly added in DAEs?
Gaussian noise
One-hot noise
Binary noise
Softmax noise
The goal of DAEs is to:
Increase dimensionality
Remove dropout
Generalize better by learning invariants to noise
Increase activation sparsity
DAEs differ from standard autoencoders by:
Using no encoder
Training with noisy inputs
Removing loss functions
Only working on text data
Which of the following fields uses autoencoders for anomaly detection?
E-commerce
Network security
Weather prediction
Game design
Autoencoders are useful in image colorization because they:
Reduce color space
Learn grayscale compression
Reconstruct missing chromatic channels
Encode RGB values
In fraud detection, autoencoders help by:
Predicting transaction amounts
Labeling fraud directly
Detecting outliers in encoded space
Encrypting user data
Which application uses VAE for data augmentation?
Text summarization
Image synthesis
Sorting algorithms
Language translation
Autoencoders can be applied in sound processing to:
Track speaker age
Denoise audio signals
Translate music to text
Predict lyrics
Which loss function is common for binary classification?
MSE
Categorical cross-entropy
Binary cross-entropy
L1 loss
Reconstruction loss in an autoencoder should:
Be high for regular inputs
Minimize the difference between input and output
Maximize the KL divergence
Penalize sparsity
Which loss function is most commonly used with VAEs?
MSE only
BCE + KL Divergence
L2 norm only
L1 norm + Dropout loss
Why is KL Divergence added to the loss in VAEs?
To penalize random noise
To regularize the latent space
To increase output sharpness
To avoid vanishing gradients
Huber loss is ideal when:
You want to ignore all outliers
You expect binary classification
You need a compromise between L1 and L2 losses
Your output is probabilistic
In PyTorch, which function performs a forward pass?
.loss()
.step()
.forward()
.train()
How are gradients zeroed in PyTorch?
optimizer.zero_grad()
model.zero_grad()
loss.backward()
torch.reset()
Which module provides activation functions like ReLU in PyTorch?
torch.nn.functional
torch.optim
torch.cuda
torch.init
What does model.eval() do in PyTorch?
Trains the model
Resets weights
Turns off dropout and batchnorm
Compiles the model
What is required to perform backpropagation in PyTorch?
.compile()
.eval()
.zero_grad()
.backward()
Which problem is common when autoencoders are too powerful?
Underfitting
Memorizing training data
Reducing input noise
Over-regularization
Autoencoders are not ideal for:
Dimensionality reduction
Data generation without structure
Anomaly detection
Denoising
What is a key drawback of deterministic autoencoders?
They are slow to train
They cannot reconstruct images
They lack stochastic sampling ability
They do not use loss functions
Why might an autoencoder fail in generalization?
Excessive KL divergence
Too small batch size
Lack of latent space constraints
Use of ReLU activation
A potential issue with autoencoders in practice is:
Latent space being interpretable
Over-regularization from KL
Poor interpolation between samples
Limited hardware support
What does the encoder in a VAE learn?
Deterministic latent code
A probability distribution over latent variables
A one-hot label
The decoder function
What does KL divergence prevent in a VAE?
Overfitting by decoder
Irregularity in latent space
Sampling variability
Softmax overfitting
The decoder in a VAE learns to:
Reconstruct the noise
Match samples to true labels
Map latent samples to data space
Compress data into a bottleneck
What improves sampling in VAEs?
Randomly shuffling weights
Encouraging the latent distribution to match a known prior
Adding dropout to latent layer
Removing reconstruction loss
A sign that VAE training is failing is:
Latent vectors are Gaussian
KL loss is near zero constantly
Encoder uses two outputs
Reconstruction loss decreases
KL Divergence is asymmetric, meaning:
D(P‖Q) = D(Q‖P)
D(P‖Q) ≠ D(Q‖P)
It does not measure distributions
It is always zero
KL Divergence is minimized when:
P(x) = 0 for all x
P and Q are independent
P and Q are identical
Q has lower entropy
In VAE training, KL divergence penalizes:
Output variance
Overfitting to the decoder
Deviation from prior distribution
Noise in the dataset
KL Divergence measures the information lost when:
Using one distribution to approximate another
Compressing data
Encoding inputs as one-hot
Discarding outliers
A key mathematical property of KL Divergence is:
It's bounded above by 1
It's negative when distributions overlap
It's always non-negative
It works only for discrete data
Latent space refers to:
A visible layer in a neural network
Intermediate compressed representation of input
Raw input data
Fully connected output layer
In a VAE, points in latent space correspond to:
Random noise
One-hot vectors
Encoded representations of data
Class labels
A good latent space should:
Be unstructured
Contain only zero vectors
Allow smooth interpolation between data points
Be trained using supervised learning only
Which of the following helps regularize latent space?
Reconstruction loss
KL Divergence
Dropout
Batch normalization
Latent space enables:
Labeling unstructured data
Visualizing class probabilities
Generative capabilities in VAEs
Overfitting reduction in CNNs
