Practice questions
The full practice set, 72 questions — each with a full worked answer: derivations, diagrams, and examples.
Generative AI mechanism for medical image synthesis
Explain the fundamental mechanism of Generative AI. Discuss its characteristics and suitability for generating synthetic medical images to augment a small training dataset. Explain the different types of generative models and their advantages.
Four generative model types with real-world applications
Discuss any four major types of Generative AI models and provide one real-world application for each.
Generative AI use cases across six industries
Explain the major use cases of Generative AI in healthcare, education, entertainment, marketing, software development, and content creation. Give a suitable example for each use case and highlight two major challenges associated with its adoption.
Joint vs conditional probability with applications
Explain the fundamental difference between joint probability and conditional probability with a suitable example. Mention two applications of each type of probabilistic model.
Generative vs discriminative models comparison table
Compare Generative and Discriminative Models based on learning objective, probability distribution, training process, decision boundary, data requirements, and applications. Give suitable algorithm examples.
Mathematical foundations of generative vs discriminative learning
Explain the mathematical foundations of Generative and Discriminative learning. Discuss how a generative model learns the joint distribution P(x,y), while a discriminative model learns the conditional distribution P(y|x) directly. Illustrate the difference using a suitable classification example.
Bayes' theorem derivation and full generative-discriminative comparison
Explain Generative and Discriminative learning paradigms in detail. Derive the relationship between P(x,y), P(x|y), P(y), and P(y|x) using Bayes' theorem. Compare their mathematical formulation, model training, decision boundaries, computational complexity, advantages, limitations, and real-world applications.
Naive Bayes vs logistic regression on binary classification
Consider a binary classification problem involving two classes. Explain how a Generative Model and a Discriminative Model approach the same problem differently. Compare their assumptions, training procedures, decision-making processes, advantages, and limitations using Naïve Bayes and Logistic Regression.
Designing a multi-class image classifier: generative vs discriminative
A dataset contains images belonging to multiple classes. Design an AI-based classification system using both Generative and Discriminative approaches, such as Naïve Bayes or Gaussian Mixture Models and Logistic Regression, SVM, or Neural Networks. Compare probability estimation, decision boundaries, handling of unseen data, training requirements, and expected performance. Justify your choice for a practical classification system.
Autoencoder vs variational autoencoder differences
Differentiate between a standard Autoencoder and a Variational Autoencoder with respect to architecture, latent representation, training objective, and generated output. Why is a VAE more suitable for generative tasks?
VAE components and ELBO loss derivation
Explain the roles of the encoder, latent space, reparameterization trick, and decoder in a VAE. Derive and explain the two major components of its loss function and discuss why both reconstruction loss and KL-divergence are required.
KL-divergence term and latent space regularization in VAE
Explain the KL-divergence term in the VAE loss function. Why is it required, and how does it help regularize the latent space?
Reparameterization trick for VAE backpropagation
Explain the reparameterization trick used in Variational Autoencoders. Why is it required for backpropagation through the stochastic latent variable? Derive the mathematical expression for sampling the latent vector and explain its significance in VAE training.
VAE architecture, ELBO derivation, and training losses
Explain the architecture and working principle of a VAE in detail. Derive the Evidence Lower Bound (ELBO) and explain how reconstruction loss and KL-divergence loss contribute to model training. Illustrate the complete process with a suitable diagram.
Critical analysis of VAE generation, ELBO, and limitations
Critically analyze VAE-based generation. Explain how a VAE learns a latent probability distribution, derive the ELBO, discuss the role of KL-divergence and the assumptions made by the model, and analyze at least three limitations compared with other generative approaches.
Designing a VAE pipeline for synthetic image generation
Design a complete VAE-based solution for generating realistic synthetic images. Explain the encoder, latent-space representation, reparameterization trick, decoder, reconstruction process, and complete loss function. Describe how to evaluate whether generated images are useful for data augmentation.
Deriving the ELBO: posterior, prior, likelihood and KL
Derive the Evidence Lower Bound (ELBO) used for training a VAE. Explain the approximate posterior q(z|x), the prior distribution p(z), reconstruction likelihood, and KL-divergence, and show how maximizing ELBO supports generative modelling.
VAE-based augmentation pipeline for a small image dataset
Design a Generative AI solution in which a VAE is used to augment a small image dataset. Explain model training, latent sampling, synthetic-image generation, quality filtering, and how a downstream classifier can be used to evaluate the usefulness of the generated samples.
GAN minimax game: Generator vs Discriminator
Discuss the working principle of Generative Adversarial Networks. Illustrate the minimax game between the Generator and Discriminator with a suitable diagram and explain the objective function.
Adversarial learning: how G and D interact in a GAN
Explain the basic concept of adversarial learning in GANs. Describe how a Generator transforms latent noise into synthetic samples and how a Discriminator evaluates real and generated samples. Illustrate the overall workflow with a suitable diagram.
The adversarial training loop and how G improves over time
Describe the adversarial training process in GANs. Explain how the Generator gradually learns to generate realistic samples through feedback from the Discriminator.
GAN architecture: forward/backward passes and the role of z
Explain the architecture of a GAN and how the Generator and Discriminator interact during training. Illustrate the forward and backward passes and discuss the significance of the latent-space input to the Generator.
Vanilla GAN architecture and information flow
Discuss the architecture of a vanilla GAN in detail. Explain the Generator and Discriminator networks, their information flow, and the objective function associated with the architecture.
Minimax optimization and alternating parameter updates in GANs
Analyze the minimax optimization process used in GAN training. Write the original GAN value function and explain how the Generator and Discriminator update their parameters alternately during adversarial training.
GAN Generator and Discriminator architecture
Explain the architectural design of the Generator and Discriminator in a GAN. Illustrate their input and output layers, activation functions, and information flow during one training iteration.
Standard GAN loss functions for G and D
Explain the standard loss functions used in GANs, including Generator loss and Discriminator loss. Discuss how these loss functions contribute to the adversarial training process.
Comparing GAN loss variants and gradient quality
Compare the standard Discriminator loss, saturating Generator loss, and non-saturating Generator loss in GANs. Explain how these objectives influence gradient quality and training stability.
Three major GAN training challenges
Discuss any three major training challenges in GANs. Explain mode collapse, vanishing gradients, and training instability, and discuss how loss-function modifications can improve training.
Diagnosing mode collapse in a digit GAN
A GAN trained on handwritten digits generates only a small subset of digit styles while the Discriminator becomes highly confident. Identify the training problem, explain its causes, and recommend practical techniques to improve sample diversity.
GAN objective comparison for synthetic MRI generation
A healthcare AI company is training a GAN to generate synthetic MRI scans but observes mode collapse, unstable loss curves, artifacts, and insufficient fine details. Diagnose the problems and compare standard GAN, Least-Squares GAN, and Wasserstein-based objectives for improving stability and medical-image quality.
DCGAN architecture vs vanilla GAN
Explain the architecture of a Deep Convolutional GAN (DCGAN) and distinguish it from a fully connected vanilla GAN. Discuss convolution and transposed-convolution layers, normalization, activation functions, and architectural guidelines for stable image generation.
Conditional GAN architecture and conditioning effect
Explain the architecture and working principle of a Conditional GAN (cGAN). How does conditioning information influence generation compared with a conventional GAN?
Class-conditioned digit generation with cGAN
Explain how class labels can be incorporated into both the Generator and Discriminator of a cGAN. Illustrate the process using class-controlled handwritten-digit generation.
Conventional GAN vs Conditional GAN comparison
Compare conventional GANs and Conditional GANs in terms of architecture, input representation, training objective, and control over generated samples. Explain with a suitable application scenario.
cGAN objective and conditioning-injection methods
Formulate the conditional adversarial objective of a cGAN. Explain different methods for injecting conditioning information into the Generator and Discriminator, and discuss two real-world applications where controlled generation is important.
Complete cGAN architecture and training process
Describe the complete architecture and training process of a Conditional GAN in detail. Explain how conditioning information is provided to both networks, formulate the conditional adversarial objective, and discuss the advantages and limitations of controlled generation.
StyleGAN's core idea vs traditional GANs
Explain the key architectural idea behind StyleGAN and describe how it differs from the traditional GAN architecture.
Mapping network, W space, and style modulation
What is style-based generation in StyleGAN? Explain the mapping network, intermediate latent space, and style modulation. Discuss how coarse, middle-level, and fine image attributes can be controlled.
StyleGAN improvements: style control, quality, diversity
Explain the major improvements introduced by StyleGAN over traditional GAN architectures. Discuss style-based control, progressive control of image features, image quality, and diversity.
From traditional GANs to StyleGAN: architecture trace
Trace the major architectural improvements from traditional GANs to StyleGAN. Discuss the mapping network, style injection, stochastic variation, and multi-scale control of visual features, and explain how they improve image quality and diversity.
Traditional GANs vs modern variants: stability and control
Compare traditional GANs with modern GAN variants such as StyleGAN. Explain how architectural improvements address training instability, image quality, and control over generated content.
Virtual fashion studio with StyleGAN attribute control
A fashion company wants to build a virtual studio that can independently control facial appearance, hairstyle, pose, clothing texture, background, and fine visual details. Design a StyleGAN-based approach and explain style levels, style mixing, and attribute control.
Synthetic faces: StyleGAN latent manipulation vs traditional GANs
A research organization wants to generate realistic synthetic human faces and modify characteristics such as age, hairstyle, facial structure, pose, and image details. Compare traditional GANs with StyleGAN and explain how latent-space manipulation supports controlled generation.
GAN image-to-image translation for product photos
An online shopping platform wants to generate different product images by changing the background, style, or appearance of existing products. Explain how GAN-based image-to-image translation can be used, with a suitable example and architecture.
Colorizing photos with a GAN-based pix2pix approach
A company wants to convert black-and-white photographs into realistic colored photographs. Explain how a GAN-based image-to-image translation approach can solve this task. Describe the Generator, Discriminator, training process, and expected transformation with a suitable diagram.
Diffusion models: forward noising and reverse denoising
Explain the basic idea of a diffusion model. Describe the forward diffusion and reverse denoising processes used for generating data.
Gaussian noise's role in DDPM training and sampling
What is the role of Gaussian noise in a Denoising Diffusion Probabilistic Model (DDPM)? Explain how a model learns to generate clean samples from noisy data.
DDPM forward diffusion: a worked 3-step example
In a DDPM, the forward diffusion process is x_t = sqrt(alpha_bar_t) * x_0 + sqrt(1 - alpha_bar_t) * epsilon. The question's original source gave specific numeric values for beta_1, beta_2, beta_3 (a 3-step toy noise schedule) that were lost in copy-paste — they did NOT survive into this text. (a) Calculate alpha_t and alpha_bar_t at each step. (b) Calculate the noisy sample x_t. (c) Find the separate contributions of the original sample and noise. (d) Explain how these contributions change as the timestep increases.
DDPM noise schedule: alpha-bar and signal-to-noise decay
A DDPM uses a noise schedule beta_1, beta_2, beta_3 over 3 timesteps. The question's original source gave specific numeric values that were lost in copy-paste — they did NOT survive into this text. Given a toy schedule, calculate alpha_t, the cumulative products alpha_bar_t, and sqrt(alpha_bar_t)/sqrt(1-alpha_bar_t). Analyze how the signal and noise contributions change across the diffusion steps.
Diffusion vs GAN for synthetic medical image generation
A medical AI research team has a limited dataset and wants to generate synthetic medical images using DDPM or Stable Diffusion. Explain the forward and reverse diffusion processes, noise prediction during training, and compare diffusion-based generation with GAN-based generation. Evaluate the benefits, challenges, and privacy risks.
Stable Diffusion's text-to-image pipeline end to end
An e-commerce company wants to generate product images from natural-language descriptions using Stable Diffusion. Explain the complete text-to-image pipeline, including the text encoder, latent representation, noise addition, U-Net-based denoising, conditioning mechanism, and decoder. Discuss advantages and limitations.
GAN design for synthetic diseased-leaf images
A smart agriculture company wants to generate realistic images of diseased plant leaves because collecting diseased-leaf images is difficult and expensive. Design a GAN-based solution and explain the Generator, Discriminator, adversarial training, objective function, and synthetic-image augmentation workflow.
GAN design for synthetic network-flow traffic data
A cybersecurity firm wants to generate realistic synthetic network-flow records containing numerical and categorical attributes. Design a GAN suitable for tabular or sequential traffic data. Explain data representation, Generator output design, Discriminator structure, training, evaluation metrics, and possible architectural improvements.
GAN architecture for synthetic satellite imagery
A research group wants to generate synthetic satellite images for environmental monitoring. Explain the architecture of a suitable GAN with a labelled diagram showing the Generator, Discriminator, real-data path, and fake-data path. Discuss training dynamics and the value function.
Full Transformer architecture: encoder, decoder, attention
Explain the architecture of the Transformer model. Describe the encoder, decoder, self-attention, feed-forward networks, residual connections, and layer normalization.
Positional encoding: derivation and role in Transformers
Analyze the role of positional encoding in Transformer-based Generative AI models. Derive the sinusoidal positional encoding expressions and explain how position information is combined with token embeddings.
Worked numeric example of single-head self-attention
Using a small numerical example, demonstrate single-head scaled dot-product self-attention. Compute the Query, Key, and Value matrices, attention scores, scaling, Softmax probabilities, and the final weighted output.
Multi-head self-attention: mechanics and motivation
Analyze Multi-Head Self-Attention in Transformers. Explain how different heads independently project Query, Key, and Value representations, compute attention, concatenate their outputs, and apply the final projection. Discuss why multiple heads are useful.
Cross-attention: decoder using encoder context
Analyze how cross-attention allows a Transformer decoder to use encoder information while generating an output sequence. Explain the Query, Key, and Value roles with a suitable machine-translation example.
Contextual embeddings: same word, different sentences
Pass three semantically different sentences through a Transformer Encoder and analyze their contextual representations. Explain why the same word can obtain different embeddings when it appears in different contexts.
Transformer decoder architecture for text generation
Design and explain a Transformer Decoder for autoregressive text generation. Discuss causal masked self-attention, optional cross-attention, residual connections, layer normalization, feed-forward networks, output projection, and next-token prediction.
Text classification using a Transformer encoder
Develop a text-classification application using a Transformer Encoder. Explain tokenization, contextual encoding, pooling or CLS representation, classification head, training objective, and prediction of the input class.
Text generation with a pretrained Transformer
Implement text generation using a pretrained Transformer model for a given prompt. Show tokenization, model inference, decoding, and the final generated text.
Zero-shot, one-shot, and few-shot prompting compared
Differentiate zero-shot, one-shot, and few-shot prompting. Write an example of all three for the same task and discuss when each prompting strategy is most suitable.
Comparing instruction, role, and reasoning prompting styles
Compare instruction prompting, role prompting, one/few-shot prompting, and structured reasoning prompting. Explain their purpose, advantages, limitations, and suitable Generative AI applications.
RAG pipeline: from documents to grounded answers
Explain the complete step-by-step workflow of a Retrieval-Augmented Generation (RAG) system, covering document ingestion, chunking, embedding generation, indexing, retrieval, prompt augmentation, generation, and source grounding.
RAG system design for a university portal
Design an end-to-end RAG system for a university knowledge portal that answers questions from syllabus documents, academic regulations, examination guidelines, and notices. Explain ingestion, chunking, embeddings, vector storage, retrieval, prompt construction, answer generation, and citation of sources.
Comparing greedy, temperature, and top-k/top-p decoding
Develop a Transformer-based text-generation experiment using three different prompts. Generate outputs using greedy decoding, temperature sampling, and top-k/top-p sampling, and compare their diversity and coherence.
Variational Autoencoder on MNIST, step by step
Implement a simple VAE using Python and TensorFlow or PyTorch on the MNIST dataset. Include the encoder, latent-space sampling using the reparameterization trick, decoder, reconstruction loss, KL-divergence loss, and training procedure. Display original and reconstructed images.
VAE image generation: sampling new MNIST digits
Implement a VAE for image generation using the MNIST dataset. Estimate the latent mean and variance, apply the reparameterization trick, decode latent vectors, calculate reconstruction and KL-divergence losses, and visualize newly generated samples.
Basic GAN for generating handwritten digits
Implement a basic GAN using Python and TensorFlow or PyTorch on the MNIST dataset. Include data preprocessing, Generator and Discriminator construction, adversarial training, loss functions, and visualization of synthetic handwritten digits at different training stages.
Text summarization with a pretrained Transformer
Implement text summarization using a pretrained Transformer model. Display the input paragraph and generated summary, and explain how the model identifies and produces the most important information.