data/Generative_Models.json

{
  "title": "Generative Models Mastery: 100 MCQs",
  "description": "A complete 100-question set covering fundamental concepts, algorithms, architectures, optimization techniques, and applications of Generative Models.",
  "questions": [
    {
      "id": 1,
      "questionText": "What is the primary goal of a generative model?",
      "options": [
        "To classify input data into categories",
        "To cluster data points",
        "To reduce dimensionality of data",
        "To generate new data samples similar to the training data"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Generative models aim to learn the underlying distribution of the data and generate new samples that resemble the training data."
    },
    {
      "id": 2,
      "questionText": "Which of the following is a type of generative model?",
      "options": [
        "Random Forest",
        "K-Means Clustering",
        "Variational Autoencoder (VAE)",
        "Support Vector Machine (SVM)"
      ],
      "correctAnswerIndex": 2,
      "explanation": "VAEs are generative models that learn a latent representation and can generate new data samples."
    },
    {
      "id": 3,
      "questionText": "In generative modeling, what does the term 'latent space' refer to?",
      "options": [
        "The output prediction space",
        "A lower-dimensional representation capturing the underlying factors of variation",
        "A space for storing training labels",
        "The input feature space"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Latent space encodes the hidden factors that capture important structure in the data, which can be used for generation."
    },
    {
      "id": 4,
      "questionText": "Which of the following models uses a game-theoretic approach to generate data?",
      "options": [
        "Naive Bayes",
        "Variational Autoencoder (VAE)",
        "Generative Adversarial Network (GAN)",
        "Principal Component Analysis (PCA)"
      ],
      "correctAnswerIndex": 2,
      "explanation": "GANs consist of a generator and a discriminator competing in a minimax game to generate realistic samples."
    },
    {
      "id": 5,
      "questionText": "What distinguishes a generative model from a discriminative model?",
      "options": [
        "Generative models learn the data distribution, discriminative models learn decision boundaries",
        "Discriminative models are always unsupervised",
        "Generative models only classify data, discriminative models generate data",
        "Generative models cannot be probabilistic, discriminative models can"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Generative models learn P(x) or P(x, y), whereas discriminative models learn P(y|x) to classify data."
    },
    {
      "id": 6,
      "questionText": "Which of the following is a probabilistic generative model?",
      "options": [
        "Naive Bayes",
        "Decision Tree",
        "SVM",
        "K-Nearest Neighbors"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Naive Bayes models the joint probability of features and class labels, making it a probabilistic generative model."
    },
    {
      "id": 7,
      "questionText": "What is a key application of generative models in computer vision?",
      "options": [
        "Color quantization",
        "Edge detection",
        "Image synthesis and inpainting",
        "Classification of handwritten digits"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Generative models can produce realistic images, fill missing parts, or create new images from learned distributions."
    },
    {
      "id": 8,
      "questionText": "Which loss function is commonly used in Variational Autoencoders (VAE)?",
      "options": [
        "Mean squared error only",
        "Hinge loss",
        "Reconstruction loss + KL divergence",
        "Cross-entropy loss"
      ],
      "correctAnswerIndex": 2,
      "explanation": "VAE optimizes reconstruction loss to reconstruct input data and KL divergence to regularize the latent space."
    },
    {
      "id": 9,
      "questionText": "In GANs, what is the role of the discriminator?",
      "options": [
        "To calculate reconstruction error",
        "To generate new data samples",
        "To distinguish real data from generated data",
        "To compress data into latent space"
      ],
      "correctAnswerIndex": 2,
      "explanation": "The discriminator evaluates whether a given sample is real or generated, providing feedback to the generator."
    },
    {
      "id": 10,
      "questionText": "What is a common challenge in training GANs?",
      "options": [
        "Vanishing gradient in VAE encoder",
        "Mode collapse",
        "Lack of reconstruction loss",
        "Overfitting on the latent space"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Mode collapse occurs when the generator produces limited variety, failing to cover the full data distribution."
    },
    {
      "id": 11,
      "questionText": "Which model can both encode and generate data samples?",
      "options": [
        "K-Means",
        "SVM",
        "Decision Tree",
        "Autoencoder"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Autoencoders compress data into a latent space (encoding) and reconstruct it (decoding), enabling generation."
    },
    {
      "id": 12,
      "questionText": "In a VAE, why is the latent space regularized?",
      "options": [
        "To ensure maximum likelihood estimation",
        "To allow smooth sampling and meaningful interpolation",
        "To prevent overfitting on labels",
        "To increase reconstruction error"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Regularizing the latent space with KL divergence ensures that points sampled from the prior produce realistic outputs."
    },
    {
      "id": 13,
      "questionText": "Which generative model is non-probabilistic?",
      "options": [
        "Hidden Markov Model",
        "Gaussian Mixture Model",
        "Variational Autoencoder (VAE)",
        "Vanilla Autoencoder"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Vanilla autoencoders learn deterministic mappings and do not model probability distributions explicitly."
    },
    {
      "id": 14,
      "questionText": "Which of the following is an example of a sequential generative model?",
      "options": [
        "Decision Tree",
        "Convolutional Neural Network (CNN)",
        "K-Means",
        "Recurrent Neural Network (RNN) based language models"
      ],
      "correctAnswerIndex": 3,
      "explanation": "RNN-based models can generate sequences like text or music by learning sequential dependencies."
    },
    {
      "id": 15,
      "questionText": "Which generative model explicitly models the joint probability distribution of the data?",
      "options": [
        "K-Nearest Neighbors",
        "Feedforward Neural Network Classifier",
        "Gaussian Mixture Model (GMM)",
        "PCA"
      ],
      "correctAnswerIndex": 2,
      "explanation": "GMM models P(x) as a mixture of Gaussian distributions, capturing the underlying data distribution."
    },
    {
      "id": 16,
      "questionText": "In GAN training, what does the generator aim to maximize?",
      "options": [
        "The KL divergence",
        "The reconstruction loss",
        "The classification accuracy",
        "The probability of the discriminator being mistaken"
      ],
      "correctAnswerIndex": 3,
      "explanation": "The generator tries to produce samples that fool the discriminator into classifying them as real."
    },
    {
      "id": 17,
      "questionText": "What is a key difference between VAE and GAN?",
      "options": [
        "GAN cannot generate images",
        "VAE is probabilistic and uses reconstruction loss; GAN uses adversarial loss",
        "Both are deterministic autoencoders",
        "VAE uses adversarial loss; GAN uses reconstruction loss"
      ],
      "correctAnswerIndex": 1,
      "explanation": "VAE models a probabilistic latent space and reconstruction loss, while GANs use a generator-discriminator game with adversarial loss."
    },
    {
      "id": 18,
      "questionText": "Which type of generative model is suitable for clustering mixed continuous and categorical data?",
      "options": [
        "RNN",
        "Gaussian Mixture Model (GMM)",
        "SVM",
        "Convolutional Autoencoder"
      ],
      "correctAnswerIndex": 1,
      "explanation": "GMMs can model continuous data probabilistically, and extensions exist for mixed data types."
    },
    {
      "id": 19,
      "questionText": "What is the primary evaluation metric for generative models in image synthesis?",
      "options": [
        "Classification accuracy",
        "Mean squared error on labels",
        "Confusion matrix",
        "Inception Score (IS) or FID"
      ],
      "correctAnswerIndex": 3,
      "explanation": "IS and FID measure the quality and diversity of generated images compared to real data."
    },
    {
      "id": 20,
      "questionText": "Which of the following can generative models be used for in NLP?",
      "options": [
        "Word classification only",
        "Sentence segmentation",
        "Text generation and language modeling",
        "Named entity recognition exclusively"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Generative models like RNNs or Transformers can generate coherent text sequences or predict next words."
    },
    {
      "id": 21,
      "questionText": "Which approach is commonly used to stabilize GAN training?",
      "options": [
        "Using deterministic latent space",
        "Increasing KL divergence weight",
        "Removing the generator",
        "Label smoothing and batch normalization"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Label smoothing and normalization techniques help prevent instability in the generator-discriminator game."
    },
    {
      "id": 22,
      "questionText": "Which generative model can model complex, multi-modal distributions explicitly?",
      "options": [
        "Linear Regression",
        "Standard Autoencoder",
        "Decision Tree",
        "Normalizing Flows"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Normalizing flows model complex distributions by transforming a simple base distribution via invertible functions."
    },
    {
      "id": 23,
      "questionText": "Which of these models learns by minimizing divergence between true data distribution and model distribution?",
      "options": [
        "Variational Autoencoder",
        "K-Means",
        "Random Forest",
        "Decision Tree"
      ],
      "correctAnswerIndex": 0,
      "explanation": "VAE minimizes reconstruction loss plus KL divergence, aligning the latent distribution with a prior."
    },
    {
      "id": 24,
      "questionText": "Which type of generative model is based on a chain of conditional probabilities?",
      "options": [
        "GANs",
        "Autoregressive models",
        "Feedforward Neural Networks",
        "VAEs"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Autoregressive models predict each element conditioned on previous elements, modeling the joint distribution sequentially."
    },
    {
      "id": 25,
      "questionText": "What is a limitation of simple autoencoders as generative models?",
      "options": [
        "They are deterministic and cannot sample new points smoothly",
        "They overfit the discriminator",
        "They cannot reduce dimensionality",
        "They require adversarial loss"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Simple autoencoders do not model probability distributions, so sampling new latent points may not generate realistic data."
    },
    {
      "id": 26,
      "questionText": "Which of the following is a key property of a probabilistic generative model?",
      "options": [
        "It performs clustering only",
        "It estimates P(x) or P(x, y)",
        "It maximizes classification accuracy",
        "It does not use probability distributions"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Probabilistic generative models explicitly model probability distributions of data or data-label pairs."
    },
    {
      "id": 27,
      "questionText": "In conditional GANs (cGANs), what is provided to the generator additionally?",
      "options": [
        "Only random noise",
        "The discriminator's parameters",
        "Conditioning information such as class labels",
        "Reconstruction loss"
      ],
      "correctAnswerIndex": 2,
      "explanation": "cGANs use additional conditioning variables to control the type of data generated, e.g., generating specific class images."
    },
    {
      "id": 28,
      "questionText": "Which generative model is best for sequence-to-sequence data?",
      "options": [
        "Autoencoders without temporal structure",
        "Gaussian Mixture Models",
        "CNNs only",
        "RNN-based or Transformer models"
      ],
      "correctAnswerIndex": 3,
      "explanation": "RNNs or Transformers can handle sequential dependencies, making them suitable for generating sequences."
    },
    {
      "id": 29,
      "questionText": "Which of the following is a key challenge in training VAEs?",
      "options": [
        "Mode collapse",
        "Label smoothing",
        "Vanishing discriminator gradient",
        "Balancing reconstruction loss and KL divergence"
      ],
      "correctAnswerIndex": 3,
      "explanation": "VAEs need to trade off reconstruction accuracy with latent space regularization using KL divergence."
    },
    {
      "id": 30,
      "questionText": "Which scenario demonstrates the use of generative models in practice?",
      "options": [
        "Generating realistic human faces from learned distributions",
        "Sorting a list of numbers",
        "Clustering sensor data without generation",
        "Computing shortest path in a graph"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Generative models can synthesize new data samples, e.g., realistic faces, by learning underlying distributions."
    },
    {
      "id": 31,
      "questionText": "Which loss function is typically used in GANs?",
      "options": [
        "Mean squared error",
        "Reconstruction loss only",
        "KL divergence only",
        "Adversarial loss (minimax)"
      ],
      "correctAnswerIndex": 3,
      "explanation": "GANs are trained using adversarial loss in a minimax game between generator and discriminator."
    },
    {
      "id": 32,
      "questionText": "In a VAE, what is the purpose of the reparameterization trick?",
      "options": [
        "To reduce mode collapse",
        "To normalize input images",
        "To improve discriminator accuracy",
        "To allow backpropagation through stochastic sampling"
      ],
      "correctAnswerIndex": 3,
      "explanation": "The reparameterization trick expresses the sampled latent variable as a differentiable function, enabling gradient-based optimization."
    },
    {
      "id": 33,
      "questionText": "Which type of GAN explicitly conditions on auxiliary information?",
      "options": [
        "Wasserstein GAN",
        "Vanilla GAN",
        "Conditional GAN (cGAN)",
        "DCGAN"
      ],
      "correctAnswerIndex": 2,
      "explanation": "cGANs incorporate additional conditioning variables, such as class labels, to control generation."
    },
    {
      "id": 34,
      "questionText": "What is the main purpose of a discriminator in a GAN?",
      "options": [
        "To cluster data points",
        "To reconstruct input data",
        "To encode input into latent space",
        "To distinguish real data from generated data"
      ],
      "correctAnswerIndex": 3,
      "explanation": "The discriminator evaluates the authenticity of samples, guiding the generator to produce realistic outputs."
    },
    {
      "id": 35,
      "questionText": "Which architecture is commonly used for image generation in GANs?",
      "options": [
        "Recurrent layers only",
        "SVM classifier",
        "Convolutional layers (DCGAN)",
        "Fully connected only"
      ],
      "correctAnswerIndex": 2,
      "explanation": "DCGANs leverage convolutional layers to capture spatial hierarchies for high-quality image generation."
    },
    {
      "id": 36,
      "questionText": "What is mode collapse in GANs?",
      "options": [
        "When latent space is regularized",
        "When reconstruction error increases",
        "When discriminator overfits",
        "When the generator produces limited variety of outputs"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Mode collapse occurs when the generator maps different latent vectors to similar outputs, reducing diversity."
    },
    {
      "id": 37,
      "questionText": "Which generative model is based on sequential factorization of joint probability?",
      "options": [
        "PCA",
        "VAEs",
        "Autoregressive models",
        "GANs"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Autoregressive models predict each variable conditioned on previous ones, effectively factorizing P(x) sequentially."
    },
    {
      "id": 38,
      "questionText": "Which of the following is a key advantage of Normalizing Flows?",
      "options": [
        "Works only for discrete data",
        "No need for latent space regularization",
        "Automatic mode collapse prevention",
        "Exact likelihood computation and invertibility"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Normalizing flows provide invertible mappings from latent to data space, allowing exact likelihood evaluation."
    },
    {
      "id": 39,
      "questionText": "Which metric evaluates both quality and diversity of generated images?",
      "options": [
        "KL divergence only",
        "Mean squared error",
        "Fréchet Inception Distance (FID)",
        "Cross-entropy loss"
      ],
      "correctAnswerIndex": 2,
      "explanation": "FID compares statistics of generated and real images to assess quality and diversity."
    },
    {
      "id": 40,
      "questionText": "In a GAN, what does the generator network learn?",
      "options": [
        "To map latent vectors to realistic samples",
        "To encode samples into latent vectors",
        "To classify images into categories",
        "To minimize KL divergence only"
      ],
      "correctAnswerIndex": 0,
      "explanation": "The generator transforms random noise (latent vectors) into data samples resembling the true distribution."
    },
    {
      "id": 41,
      "questionText": "Which technique can stabilize GAN training?",
      "options": [
        "Maximizing reconstruction error",
        "Wasserstein loss with gradient penalty",
        "Reducing latent vector size to 1",
        "Removing the discriminator"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Wasserstein GAN (WGAN) loss with gradient penalty improves convergence and reduces mode collapse."
    },
    {
      "id": 42,
      "questionText": "Which generative model uses latent variables to represent data probabilistically?",
      "options": [
        "Autoregressive model",
        "Decision Tree",
        "Variational Autoencoder (VAE)",
        "GAN"
      ],
      "correctAnswerIndex": 2,
      "explanation": "VAE learns a probabilistic latent space with parameters (mean and variance) to generate samples."
    },
    {
      "id": 43,
      "questionText": "Which generative model is particularly suitable for text generation?",
      "options": [
        "DCGAN",
        "GMM",
        "Convolutional Autoencoder",
        "RNN-based or Transformer-based models"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Sequential models like RNNs and Transformers capture temporal dependencies in text."
    },
    {
      "id": 44,
      "questionText": "What is the main difference between explicit and implicit generative models?",
      "options": [
        "Implicit models cannot generate samples",
        "Implicit models compute exact likelihood; explicit models approximate it",
        "Explicit models estimate data likelihood; implicit models do not",
        "Explicit models always use GANs; implicit models use VAEs"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Explicit models model the probability distribution (e.g., VAE, Normalizing Flows), whereas implicit models (GANs) learn to sample without computing likelihood."
    },
    {
      "id": 45,
      "questionText": "Which model is most suitable for generating high-resolution images?",
      "options": [
        "Progressive GAN or StyleGAN",
        "Vanilla Autoencoder",
        "Gaussian Mixture Model",
        "RNN"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Progressive GANs and StyleGANs can synthesize high-resolution images by progressively increasing image size during training."
    },
    {
      "id": 46,
      "questionText": "Which of the following is an autoregressive generative model?",
      "options": [
        "K-Means",
        "PixelRNN or PixelCNN",
        "GAN",
        "VAE"
      ],
      "correctAnswerIndex": 1,
      "explanation": "PixelRNN/CNN model images pixel by pixel, conditioning each on previous pixels."
    },
    {
      "id": 47,
      "questionText": "Which of the following is a limitation of VAEs?",
      "options": [
        "Cannot encode data",
        "Generated samples may be blurry",
        "Cannot model probability distributions",
        "Require adversarial loss"
      ],
      "correctAnswerIndex": 1,
      "explanation": "VAEs optimize a trade-off between reconstruction and regularization; this can result in less sharp images compared to GANs."
    },
    {
      "id": 48,
      "questionText": "Which type of generative model is most suitable for density estimation?",
      "options": [
        "Normalizing Flows",
        "GANs",
        "RNN for sequence generation",
        "Vanilla Autoencoders"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Normalizing flows allow exact likelihood computation, making them ideal for density estimation."
    },
    {
      "id": 49,
      "questionText": "Which technique improves diversity of generated samples in GANs?",
      "options": [
        "Minibatch discrimination",
        "Removing the discriminator",
        "Increasing KL divergence only",
        "Reducing latent space size"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Minibatch discrimination introduces dependencies between samples to prevent mode collapse."
    },
    {
      "id": 50,
      "questionText": "Which loss is used in Wasserstein GANs (WGAN)?",
      "options": [
        "Mean squared error",
        "Cross-entropy loss",
        "Earth-Mover (Wasserstein) distance",
        "KL divergence only"
      ],
      "correctAnswerIndex": 2,
      "explanation": "WGAN minimizes the Wasserstein distance between real and generated distributions for better training stability."
    },
    {
      "id": 51,
      "questionText": "Which model learns a mapping from a simple distribution to a complex distribution using invertible functions?",
      "options": [
        "Normalizing Flows",
        "Autoregressive model",
        "GAN",
        "VAE"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Normalizing flows use invertible transformations to map simple distributions (like Gaussian) to complex target distributions."
    },
    {
      "id": 52,
      "questionText": "Which approach allows VAEs to generate smooth interpolations between samples?",
      "options": [
        "Sequential sampling without regularization",
        "Random noise addition",
        "Regularized latent space with Gaussian prior",
        "Discriminator feedback"
      ],
      "correctAnswerIndex": 2,
      "explanation": "A regularized latent space ensures nearby points correspond to similar outputs, enabling smooth interpolation."
    },
    {
      "id": 53,
      "questionText": "Which generative model can perform style transfer effectively?",
      "options": [
        "RNN",
        "GANs (e.g., CycleGAN)",
        "Normalizing Flows",
        "Vanilla Autoencoders"
      ],
      "correctAnswerIndex": 1,
      "explanation": "CycleGAN and other GAN variants can transfer style between domains without paired data."
    },
    {
      "id": 54,
      "questionText": "Which metric is used to compare distributions of generated and real data?",
      "options": [
        "MSE",
        "KL divergence or JS divergence",
        "Accuracy",
        "F1-score"
      ],
      "correctAnswerIndex": 1,
      "explanation": "KL and JS divergence measure how similar the generated distribution is to the true distribution."
    },
    {
      "id": 55,
      "questionText": "Which technique improves training of deep GANs for image synthesis?",
      "options": [
        "Removing convolutional layers",
        "Progressive growing of generator and discriminator",
        "Reducing latent space dimension to 1",
        "Only using MSE loss"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Progressively increasing image resolution during training stabilizes GANs and produces high-quality images."
    },
    {
      "id": 56,
      "questionText": "Which generative model is suitable for multi-modal outputs?",
      "options": [
        "Linear regression",
        "Mixture density networks or VAEs with flexible priors",
        "K-Means",
        "Decision tree"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Flexible priors or mixture models allow generating diverse outputs representing multiple modes in the data."
    },
    {
      "id": 57,
      "questionText": "Which of the following can be used to improve latent space disentanglement in VAEs?",
      "options": [
        "Using only adversarial loss",
        "Increasing discriminator size",
        "Removing the encoder",
        "β-VAE with adjustable KL weight"
      ],
      "correctAnswerIndex": 3,
      "explanation": "β-VAE introduces a weight on KL divergence to encourage disentangled latent representations."
    },
    {
      "id": 58,
      "questionText": "Which GAN variant is designed to reduce mode collapse?",
      "options": [
        "VAE",
        "Vanilla GAN",
        "Autoregressive GAN",
        "Unrolled GAN"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Unrolled GAN simulates discriminator updates ahead of time to prevent mode collapse."
    },
    {
      "id": 59,
      "questionText": "Which generative model is best for audio waveform synthesis?",
      "options": [
        "Vanilla Autoencoder",
        "WaveNet (autoregressive model)",
        "GMM",
        "DCGAN"
      ],
      "correctAnswerIndex": 1,
      "explanation": "WaveNet uses autoregressive convolutions to generate realistic audio waveforms."
    },
    {
      "id": 60,
      "questionText": "Which model is best for text-to-image generation?",
      "options": [
        "PixelCNN",
        "RNN only",
        "Conditional GANs or Diffusion Models",
        "Standard VAE"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Conditional GANs and diffusion-based models can generate images conditioned on text descriptions."
    },
    {
      "id": 61,
      "questionText": "Which of the following is a major challenge in generative modeling?",
      "options": [
        "Reducing reconstruction only",
        "Maximizing classification accuracy",
        "Balancing diversity and quality of generated samples",
        "Minimizing clustering error"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Generative models must produce realistic and diverse samples, which is often challenging to balance."
    },
    {
      "id": 62,
      "questionText": "Which of the following models can model conditional distributions directly?",
      "options": [
        "Vanilla Autoencoder",
        "Conditional VAE or cGAN",
        "PCA",
        "Unsupervised GAN without labels"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Conditional generative models can generate samples based on input conditions like labels or attributes."
    },
    {
      "id": 63,
      "questionText": "Which generative model is capable of exact likelihood evaluation?",
      "options": [
        "RNN language model",
        "GANs",
        "Vanilla Autoencoders",
        "Normalizing Flows"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Invertible transformations in Normalizing Flows allow computing the exact probability of generated samples."
    },
    {
      "id": 64,
      "questionText": "Which technique improves GAN convergence?",
      "options": [
        "Reducing latent dimension to 1",
        "Removing batch normalization",
        "Using only fully connected layers",
        "Spectral normalization of discriminator weights"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Spectral normalization stabilizes discriminator updates and prevents gradient explosion."
    },
    {
      "id": 65,
      "questionText": "Which generative model is suitable for semi-supervised learning?",
      "options": [
        "Vanilla Autoencoder",
        "RNN autoregressive model",
        "Normalizing Flow",
        "GANs with auxiliary classifier (AC-GAN)"
      ],
      "correctAnswerIndex": 3,
      "explanation": "AC-GAN incorporates labels in training, enabling semi-supervised learning by generating labeled data."
    },
    {
      "id": 66,
      "questionText": "Which generative model can combine multiple modalities (e.g., text and image)?",
      "options": [
        "PixelCNN",
        "Multimodal VAEs or GANs",
        "Standard Autoencoder",
        "RNN language model"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Multimodal generative models can generate data conditioned on multiple input types."
    },
    {
      "id": 67,
      "questionText": "Which GAN variant improves gradient flow and reduces training instability?",
      "options": [
        "Vanilla GAN without batch normalization",
        "Conditional GAN without discriminator",
        "Wasserstein GAN with gradient penalty",
        "VAE with KL divergence only"
      ],
      "correctAnswerIndex": 2,
      "explanation": "WGAN-GP ensures smoother gradient updates, improving training stability."
    },
    {
      "id": 68,
      "questionText": "Which evaluation metric measures similarity of feature distributions between real and generated images?",
      "options": [
        "Accuracy",
        "KL divergence only",
        "MSE",
        "Fréchet Inception Distance (FID)"
      ],
      "correctAnswerIndex": 3,
      "explanation": "FID computes the distance between feature distributions of real and generated images, assessing both quality and diversity."
    },
    {
      "id": 69,
      "questionText": "Which model is most suitable for density estimation in high-dimensional continuous data?",
      "options": [
        "Standard Autoencoders",
        "GANs",
        "Normalizing Flows",
        "PixelCNN"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Normalizing flows can handle high-dimensional continuous data with exact likelihood computation."
    },
    {
      "id": 70,
      "questionText": "Which technique can improve VAE image sharpness?",
      "options": [
        "Removing decoder",
        "Using fully connected layers only",
        "Reducing KL divergence weight to zero",
        "Combining VAE with GAN (VAE-GAN)"
      ],
      "correctAnswerIndex": 3,
      "explanation": "VAE-GAN combines reconstruction with adversarial loss, producing sharper and more realistic images."
    },
    {
      "id": 71,
      "questionText": "You are training a GAN for high-resolution image generation, but the generator produces blurry outputs. What is the most likely cause?",
      "options": [
        "Training data is too small for a VAE",
        "Mode collapse in the discriminator",
        "Latent space regularization is too strong",
        "The model architecture or loss function is not suitable for high-resolution outputs"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Blurry outputs often result from inadequate generator architecture or loss function for high-resolution images. Solutions include using DCGAN, Progressive GAN, or VAE-GAN architectures."
    },
    {
      "id": 72,
      "questionText": "During VAE training, the KL divergence term dominates the reconstruction loss. What effect does this have?",
      "options": [
        "Causes mode collapse in the generator",
        "Reduces latent space smoothness",
        "The model may produce outputs similar to the prior but poorly reconstruct inputs",
        "Improves image sharpness"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Excessive KL weight forces latent variables to match the prior, reducing reconstruction fidelity."
    },
    {
      "id": 73,
      "questionText": "You are using a conditional GAN to generate labeled images. The generator only produces images of a single class. What is happening?",
      "options": [
        "Underfitting of VAE",
        "Overfitting of discriminator",
        "Latent space regularization",
        "Mode collapse"
      ],
      "correctAnswerIndex": 3,
      "explanation": "The generator collapses to producing limited outputs, ignoring class conditioning, which is classic mode collapse."
    },
    {
      "id": 74,
      "questionText": "You want to generate diverse text sequences. Which generative model is most appropriate?",
      "options": [
        "Gaussian Mixture Model",
        "Transformer-based autoregressive model",
        "DCGAN",
        "Vanilla Autoencoder"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Transformers model sequence dependencies well and can generate diverse, coherent text sequences."
    },
    {
      "id": 75,
      "questionText": "You need exact likelihood evaluation for high-dimensional continuous data. Which model should you choose?",
      "options": [
        "RNN autoregressive model",
        "GAN",
        "Normalizing Flows",
        "VAE"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Normalizing flows provide invertible mappings allowing exact likelihood computation, suitable for density estimation."
    },
    {
      "id": 76,
      "questionText": "You want to combine VAE reconstruction with realistic image quality. Which approach is best?",
      "options": [
        "PixelCNN",
        "DCGAN only",
        "VAE-GAN",
        "Vanilla VAE"
      ],
      "correctAnswerIndex": 2,
      "explanation": "VAE-GAN combines reconstruction loss with adversarial loss to generate sharp images while maintaining latent structure."
    },
    {
      "id": 77,
      "questionText": "While training a GAN, gradients vanish and the generator fails to improve. Which technique helps?",
      "options": [
        "Use only MSE loss",
        "Increase KL divergence",
        "Remove the discriminator",
        "Use Wasserstein loss with gradient penalty"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Wasserstein GAN with gradient penalty stabilizes training and prevents vanishing gradients."
    },
    {
      "id": 78,
      "questionText": "You are generating multimodal data (images + text). Which generative approach is suitable?",
      "options": [
        "Normalizing Flows for text only",
        "Multimodal VAE or GAN",
        "PixelRNN only",
        "Standard Autoencoder"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Multimodal generative models can handle multiple types of inputs and generate data conditioned on both modalities."
    },
    {
      "id": 79,
      "questionText": "Your GAN produces high-quality images but only from a limited subset of the data distribution. What is this issue called?",
      "options": [
        "Underfitting",
        "Mode collapse",
        "Overfitting",
        "Latent space regularization"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Mode collapse occurs when the generator ignores parts of the data distribution and produces limited variety."
    },
    {
      "id": 80,
      "questionText": "Which evaluation metric can detect mode collapse in image generation?",
      "options": [
        "Fréchet Inception Distance (FID)",
        "Accuracy",
        "MSE",
        "Cross-entropy"
      ],
      "correctAnswerIndex": 0,
      "explanation": "FID measures distributional similarity; poor FID often indicates lack of diversity or mode collapse."
    },
    {
      "id": 81,
      "questionText": "Which approach encourages disentangled latent representations in VAEs?",
      "options": [
        "PixelCNN",
        "Standard GAN",
        "β-VAE",
        "Autoregressive model"
      ],
      "correctAnswerIndex": 2,
      "explanation": "β-VAE adds weight to the KL divergence term, encouraging independent latent factors."
    },
    {
      "id": 82,
      "questionText": "You are training a text-to-image model. Which generative architecture is suitable?",
      "options": [
        "Autoregressive flow model",
        "PixelRNN only",
        "Conditional GAN or diffusion-based model",
        "VAE only"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Conditional GANs and diffusion models can generate images conditioned on text input."
    },
    {
      "id": 83,
      "questionText": "In a VAE, if latent space dimension is too small, what is likely to happen?",
      "options": [
        "Mode collapse",
        "Poor reconstruction quality due to information bottleneck",
        "Gradient explosion in discriminator",
        "Overfitting on test set"
      ],
      "correctAnswerIndex": 1,
      "explanation": "A small latent dimension limits information storage, reducing reconstruction fidelity."
    },
    {
      "id": 84,
      "questionText": "Which GAN variant allows semi-supervised learning?",
      "options": [
        "PixelCNN",
        "Normalizing Flow",
        "Standard VAE",
        "AC-GAN (Auxiliary Classifier GAN)"
      ],
      "correctAnswerIndex": 3,
      "explanation": "AC-GAN uses an auxiliary classifier in the discriminator to incorporate labeled data for semi-supervised learning."
    },
    {
      "id": 85,
      "questionText": "Which method helps prevent mode collapse by making discriminator aware of multiple samples?",
      "options": [
        "KL divergence scaling",
        "Gradient clipping",
        "Latent space regularization",
        "Minibatch discrimination"
      ],
      "correctAnswerIndex": 3,
      "explanation": "Minibatch discrimination introduces dependencies among samples, encouraging generator diversity."
    },
    {
      "id": 86,
      "questionText": "You are training a GAN on limited data, but it overfits. Which technique can help?",
      "options": [
        "Increase latent space dimension",
        "Data augmentation and regularization",
        "Remove the discriminator",
        "Reduce batch size to 1"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Data augmentation expands the dataset, helping the generator and discriminator generalize better."
    },
    {
      "id": 87,
      "questionText": "Which generative model is best for continuous sequence prediction (e.g., speech waveforms)?",
      "options": [
        "PixelCNN",
        "DCGAN",
        "VAE only",
        "WaveNet (autoregressive model)"
      ],
      "correctAnswerIndex": 3,
      "explanation": "WaveNet uses autoregressive convolutions suitable for generating continuous sequences like audio."
    },
    {
      "id": 88,
      "questionText": "Which model can generate new images while maintaining semantic content from a reference image?",
      "options": [
        "PixelCNN",
        "Conditional GAN (e.g., Pix2Pix or CycleGAN)",
        "VAE without conditioning",
        "Vanilla Autoencoder"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Conditional GANs can generate images conditioned on a reference, preserving structure while changing style."
    },
    {
      "id": 89,
      "questionText": "You need to evaluate generated text quality. Which metric is suitable?",
      "options": [
        "BLEU or ROUGE score",
        "MSE",
        "KL divergence only",
        "FID"
      ],
      "correctAnswerIndex": 0,
      "explanation": "BLEU and ROUGE compare generated text against reference text for content quality and fluency."
    },
    {
      "id": 90,
      "questionText": "Which approach can generate realistic images from random noise efficiently?",
      "options": [
        "VAE without adversarial loss",
        "PixelCNN only",
        "GAN with convolutional generator",
        "Autoregressive RNN"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Convolutional GANs transform random latent vectors into high-quality images efficiently."
    },
    {
      "id": 91,
      "questionText": "Which challenge is common in conditional generative models?",
      "options": [
        "Reconstruction error is zero",
        "Generator ignoring conditioning labels (mode collapse)",
        "Gradient explosion in decoder",
        "Overfitting latent space only"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Conditional models may fail to produce diverse outputs for all conditions, leading to mode collapse."
    },
    {
      "id": 92,
      "questionText": "Which technique allows GANs to handle high-resolution images more effectively?",
      "options": [
        "VAE reconstruction only",
        "Progressive growing of generator and discriminator",
        "Reducing latent space dimension to 1",
        "Removing convolutional layers"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Progressively increasing image resolution during training stabilizes GANs and enables high-resolution synthesis."
    },
    {
      "id": 93,
      "questionText": "You want to interpolate between two generated faces smoothly. Which model property is critical?",
      "options": [
        "Regularized latent space (e.g., in VAE)",
        "Autoregressive pixel modeling",
        "Mode collapse prevention only",
        "Large discriminator"
      ],
      "correctAnswerIndex": 0,
      "explanation": "A smooth latent space ensures that interpolating between points generates meaningful intermediate outputs."
    },
    {
      "id": 94,
      "questionText": "Which generative model is suitable for generating tabular data with mixed categorical and continuous features?",
      "options": [
        "WaveNet",
        "CTGAN or GMM-based models",
        "PixelCNN",
        "VAE for images only"
      ],
      "correctAnswerIndex": 1,
      "explanation": "CTGANs can handle tabular data and model mixed feature types effectively."
    },
    {
      "id": 95,
      "questionText": "Which technique improves GAN training stability and reduces oscillations?",
      "options": [
        "Using fully connected layers only",
        "Reducing latent dimension to 1",
        "Using Wasserstein loss with gradient penalty",
        "Only reconstruction loss"
      ],
      "correctAnswerIndex": 2,
      "explanation": "WGAN-GP provides smoother gradients, improving convergence and stability in training GANs."
    },
    {
      "id": 96,
      "questionText": "Which generative model allows controlled attribute manipulation (e.g., changing hair color in images)?",
      "options": [
        "Normalizing Flows only",
        "PixelCNN",
        "Conditional GANs or StyleGAN",
        "Vanilla Autoencoder"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Conditional GANs and StyleGAN allow latent space manipulations to change attributes while keeping other content fixed."
    },
    {
      "id": 97,
      "questionText": "You observe that GAN training oscillates and fails to converge. Which step is recommended?",
      "options": [
        "Use gradient penalty, spectral normalization, or learning rate tuning",
        "Remove the generator entirely",
        "Use only MSE loss",
        "Increase latent dimension to 10,000"
      ],
      "correctAnswerIndex": 0,
      "explanation": "Techniques like gradient penalty and spectral normalization stabilize GAN training and reduce oscillations."
    },
    {
      "id": 98,
      "questionText": "Which generative model is most appropriate for music generation?",
      "options": [
        "DCGAN only",
        "Standard VAE without temporal modeling",
        "RNN-based or Transformer-based models",
        "PixelCNN"
      ],
      "correctAnswerIndex": 2,
      "explanation": "Sequential models like RNNs or Transformers can capture temporal dependencies for music synthesis."
    },
    {
      "id": 99,
      "questionText": "You need a generative model that can produce multiple diverse outputs for a single input. Which approach is suitable?",
      "options": [
        "PixelCNN",
        "Conditional VAE or multimodal GAN",
        "Vanilla GAN only",
        "Standard Autoencoder"
      ],
      "correctAnswerIndex": 1,
      "explanation": "Conditional VAEs or multimodal GANs allow sampling diverse outputs for the same input condition."
    },
    {
      "id": 100,
      "questionText": "Which approach allows a VAE to generate sharper images without sacrificing latent space structure?",
      "options": [
        "Use only MSE loss",
        "Remove KL divergence",
        "Combine with adversarial loss (VAE-GAN)",
        "Reduce latent dimension to 1"
      ],
      "correctAnswerIndex": 2,
      "explanation": "VAE-GAN leverages adversarial loss to improve image sharpness while retaining smooth latent representations."
    }
  ]
}