Back to Subject

Unit 1: Foundations of Generative AI - Practice Quiz

CSG202 — Generative Ai Fundamentals 60 Questions
0 Correct 0 Wrong 60 Left
0/60

1 What is the primary purpose of generative AI?

Define generative AI Easy
A. To physically assemble hardware components
B. To delete outdated files from a database
C. To only classify existing data into categories
D. To create new content such as text, images, or audio

2 Which of the following best describes generative AI?

Define generative AI Easy
A. A system that only follows fixed if-else rules
B. A tool that only stores data in the cloud
C. A type of AI that produces original outputs from learned patterns
D. A program that manages computer memory allocation

3 Generative AI is a subset of which broader field?

Define generative AI Easy
A. Network security
B. Web development
C. Database administration
D. Machine learning

4 Which of the following is an example of a generative AI output?

Define generative AI Easy
A. A calculation of the sum of two numbers
B. A record deleted from a spreadsheet
C. An AI-written poem about the ocean
D. A sorted list of email addresses

5 What do generative AI models learn from during training?

Explain how generative AI works Easy
A. A single fixed rulebook
B. Only user passwords
C. The computer's clock speed
D. Large amounts of data

6 In language models, a prompt refers to what?

Explain how generative AI works Easy
A. The input text given to the model
B. The final saved model file
C. The electricity powering the server
D. The hardware running the model

7 What is a token in the context of language models?

Explain how generative AI works Easy
A. A small unit of text such as a word or part of a word
B. A physical coin used to pay for AI
C. A backup copy of the model
D. A type of computer virus

8 How does a text-generating model typically produce its output?

Explain how generative AI works Easy
A. By searching only the local file system
B. By predicting the next token based on previous ones
C. By copying answers directly from a fixed list
D. By randomly selecting characters with no learning

9 What is the process of adjusting a pre-trained model on a smaller, specific dataset called?

Explain how generative AI works Easy
A. Fine-tuning
B. Compiling
C. Encrypting
D. Formatting

10 Which component in modern generative models helps focus on relevant parts of the input?

Explain how generative AI works Easy
A. The cooling fan
B. The power supply
C. The keyboard driver
D. The attention mechanism

11 What does the acronym GAN stand for?

Describe generative AI model types Easy
A. Grouped Attention Node
B. Global Access Number
C. Generative Adversarial Network
D. General Analog Network

12 A GAN consists of which two main components?

Describe generative AI model types Easy
A. A compiler and a linker
B. A router and a switch
C. A generator and a discriminator
D. A server and a database

13 Which architecture is the foundation of most modern large language models?

Describe generative AI model types Easy
A. Spreadsheet
B. Transformer
C. Firewall
D. Router

14 What does LLM stand for?

Describe generative AI model types Easy
A. Linear Learning Method
B. Local Log Manager
C. Long Latency Machine
D. Large Language Model

15 Which model type is commonly used to generate high-quality images from text prompts?

Describe generative AI model types Easy
A. Diffusion models
B. Sorting algorithms
C. Spreadsheet macros
D. Firewall rules

16 What does VAE stand for in generative modeling?

Describe generative AI model types Easy
A. Variational Autoencoder
B. Virtual Access Engine
C. Verified Audio Encoder
D. Variable Array Editor

17 Which of the following is a common application of generative AI?

Describe generative AI applications Easy
A. Cleaning a computer keyboard
B. Measuring network cable length
C. Cooling the CPU during heavy load
D. Chatbots that answer questions

18 Generative AI tools that create pictures from a text description perform what task?

Describe generative AI applications Easy
A. Text-to-image generation
B. File encryption
C. Image compression only
D. Text-to-speech deletion

19 Which task is an example of generative AI assisting software developers?

Describe generative AI applications Easy
A. Printing physical documents
B. Managing office electricity
C. Replacing the computer monitor
D. Suggesting and completing code

20 How can generative AI help with long documents?

Describe generative AI applications Easy
A. By formatting the hard drive
B. By physically shredding them
C. By summarizing them into shorter text
D. By adjusting screen brightness

21 A team builds a system that produces brand-new product descriptions rather than labeling existing ones as spam or not-spam. Which characteristic best classifies this system as generative AI?

Define generative AI Medium
A. It only retrieves and copies stored text from a database
B. It creates novel content by modeling the distribution of training data
C. It reduces the dimensionality of input features for visualization
D. It assigns each input to one of a fixed set of predefined categories

22 Which of the following tasks is the clearest example of generative rather than discriminative modeling?

Define generative AI Medium
A. Predicting whether an email is spam or not
B. Synthesizing a realistic human face that does not exist
C. Detecting the language of a given sentence
D. Ranking search results by relevance score

23 A discriminative model learns while a generative model typically learns or . What practical capability does learning give a generative model?

Define generative AI Medium
A. It removes the need for any training data at all
B. It can only draw boundaries between existing classes
C. It guarantees perfect accuracy on classification tasks
D. It can sample and produce new data instances resembling the training set

24 In a transformer-based language model, the self-attention mechanism primarily allows the model to do what?

Explain how generative AI works Medium
A. Convert images into fixed-length numerical labels
B. Weigh the relevance of different tokens in a sequence to each other
C. Store the entire training corpus inside its parameters verbatim
D. Guarantee that outputs never repeat any training text

25 During text generation, increasing the temperature parameter of a language model generally has what effect on the output?

Explain how generative AI works Medium
A. It permanently retrains the model on new data
B. It forces the model to always pick the single most likely token
C. It increases randomness and diversity in the generated tokens
D. It reduces the size of the model's vocabulary

26 A large language model predicts text by estimating the probability of the next token given previous tokens. This objective is best described as:

Explain how generative AI works Medium
A. Autoregressive next-token prediction
B. Reinforcement of pixel-level rewards
C. Supervised image segmentation
D. Unsupervised clustering of documents

27 Why are embeddings important in how generative models process text?

Explain how generative AI works Medium
A. They label each sentence with a spam or not-spam tag
B. They compress the model file so it uses less disk space
C. They convert vectors back into raw pixels for display
D. They map words into dense vectors that capture semantic relationships

28 In the training of a generative model, what does a loss function typically measure?

Explain how generative AI works Medium
A. The physical temperature of the GPU during training
B. The number of parameters stored in the network
C. The difference between the model's predictions and the target data
D. The total number of tokens in the vocabulary

29 A Generative Adversarial Network (GAN) trains two competing networks. What are their roles?

Describe generative AI model types Medium
A. Two identical encoders compress the same image in parallel
B. A generator creates samples while a discriminator judges their authenticity
C. A translator converts text while a classifier labels it
D. A generator stores data while a decoder deletes duplicates

30 Modern image generators like Stable Diffusion are based on diffusion models. How do they generate images?

Describe generative AI model types Medium
A. By iteratively denoising random noise into a coherent image
B. By retrieving the closest matching image from a fixed gallery
C. By compressing an image into a single scalar value
D. By classifying an image into one of several object categories

31 A Variational Autoencoder (VAE) differs from a plain autoencoder mainly because it:

Describe generative AI model types Medium
A. Only reconstructs inputs and can never generate new samples
B. Uses two adversarial networks competing against each other
C. Learns a probabilistic latent space that can be sampled to generate new data
D. Requires labeled data for every training example

32 The transformer architecture became dominant for large language models primarily because it:

Describe generative AI model types Medium
A. Eliminates the need for GPUs during training
B. Requires far less training data than any other model type
C. Processes sequence tokens in parallel using attention instead of recurrence
D. Can only handle fixed images of a single resolution

33 Which model type is most naturally suited for generating a coherent paragraph of text one word at a time?

Describe generative AI model types Medium
A. An autoregressive transformer language model
B. A k-means clustering algorithm
C. A convolutional neural network for image classification
D. A simple linear regression model

34 A researcher wants a model that maps an input distribution to a latent space and back using an encoder-decoder pair with a regularized latent code. Which model best fits this description?

Describe generative AI model types Medium
A. Decision Tree Classifier
B. Support Vector Machine (SVM)
C. Variational Autoencoder (VAE)
D. Random Forest Regressor

35 A company uses generative AI to automatically draft first-pass replies to customer support emails. This application is best categorized as:

Describe generative AI applications Medium
A. Anomaly detection in network traffic
B. Text generation for automated content creation
C. Time-series forecasting of stock prices
D. Image classification for object detection

36 A studio uses a generative model to turn a text prompt like "a red fox in a snowy forest" into an original image. This is an example of:

Describe generative AI applications Medium
A. Text-to-image generation
B. Sentiment analysis
C. Data compression
D. Optical character recognition

37 In software development, tools like code assistants that suggest complete functions from a comment demonstrate which generative AI application?

Describe generative AI applications Medium
A. Counting the lines of code in a repository
B. Code generation from natural-language descriptions
C. Encrypting source files for security
D. Compiling machine code into assembly

38 A drug-discovery lab uses a generative model to propose novel molecular structures with desired properties. This best illustrates generative AI used for:

Describe generative AI applications Medium
A. Payroll processing automation
B. Synthetic data and candidate generation in scientific research
C. Database index optimization
D. Real-time video streaming compression

39 Which scenario is the least appropriate use of a generative AI model on its own?

Describe generative AI applications Medium
A. Producing legally binding financial figures that must be exactly correct
B. Generating a sample dialogue for a chatbot
C. Drafting a creative marketing tagline
D. Creating a rough concept illustration

40 A music platform generates original background tracks from a short melodic prompt. This is an example of generative AI applied to:

Describe generative AI applications Medium
A. Warehouse inventory counting
B. Spam email filtering
C. Audio and music synthesis
D. Fingerprint identification

41 A researcher claims their model is 'generative' because it outputs class labels with calibrated probability distributions . Why is this claim technically incorrect from a probabilistic modeling standpoint?

Define generative AI Hard
A. It models the conditional rather than the joint or data distribution , making it discriminative
B. It cannot output labels, since generative models only produce continuous values
C. It lacks a softmax layer, which is mandatory for all generative architectures
D. It uses probabilities instead of hard decision boundaries, which only clustering models do

42 Which characteristic most precisely distinguishes a generative model from a discriminative one, even when both are neural networks trained on the same dataset?

Define generative AI Hard
A. The generative model always has more parameters than the discriminative one
B. The generative model can synthesize novel samples resembling the training distribution
C. The generative model must use unsupervised gradient descent exclusively
D. The generative model requires labeled data while the discriminative one does not

43 A team argues that a lookup table returning memorized training sentences is 'generative AI' because it produces text. What is the strongest counterargument?

Define generative AI Hard
A. It has no loss function, which is the sole requirement for being generative
B. It is deterministic, and all generative models are strictly stochastic
C. It uses text output, and generative AI must only output images or audio
D. It fails to generalize to a learned distribution and only reproduces stored samples without novel synthesis

44 In an autoregressive language model, the joint probability of a sequence is factorized as which of the following?

Explain how generative AI works Hard
A.
B.
C.
D.

45 Lowering the sampling temperature in a language model's softmax produces what effect on generated output?

Explain how generative AI works Hard
A. Output probability mass spreads evenly, increasing diversity
B. Output becomes maximally random, sampling all tokens uniformly
C. Output halts because the softmax becomes undefined at low temperature
D. Output becomes near-deterministic, collapsing toward the highest-probability token (greedy behavior)

46 A diffusion model is trained to reverse a gradual noising process. During generation, what does the model fundamentally learn to predict at each denoising step?

Explain how generative AI works Hard
A. The class label of the eventual output before any denoising
B. The exact final image in a single forward pass without iteration
C. The discriminator's gradient to fool an adversarial critic
D. The noise (or score) to remove, progressively transforming random noise into a data sample

47 In the self-attention mechanism, attention weights are computed as . Why is the scaling factor included?

Explain how generative AI works Hard
A. To prevent large dot products from pushing softmax into regions with vanishingly small gradients
B. To guarantee that attention weights sum to exactly
C. To normalize the values matrix to unit length before multiplication
D. To reduce the parameter count of the query and key projections

48 A variational autoencoder (VAE) optimizes the Evidence Lower Bound (ELBO). Which term in the ELBO acts as a regularizer forcing the latent posterior toward the prior?

Explain how generative AI works Hard
A. The KL divergence
B. The adversarial loss between encoder and decoder
C. The cross-entropy between input and output labels
D. The reconstruction likelihood

49 Why can teacher forcing during training cause 'exposure bias' at inference time in autoregressive generators?

Explain how generative AI works Hard
A. Training conditions on ground-truth tokens, but inference conditions on the model's own possibly-erroneous outputs, compounding errors
B. Teacher forcing disables backpropagation, so the model never learns dependencies
C. Inference uses ground-truth tokens while training uses generated ones, reversing the bias
D. Training uses larger batches than inference, changing the loss landscape

50 GAN training is often unstable due to 'mode collapse.' What does mode collapse specifically describe?

Describe generative AI model types Hard
A. The latent space becomes larger than the data space
B. The discriminator's loss diverges to infinity, halting training
C. The generator perfectly matches every mode, ending training prematurely
D. The generator produces limited, repetitive outputs covering only a few modes of the true data distribution

51 Which comparison between VAEs and GANs is accurate regarding sample quality and likelihood?

Describe generative AI model types Hard
A. VAEs use adversarial critics while GANs use a KL-regularized encoder
B. Both compute exact log-likelihoods and produce identically sharp outputs
C. GANs provide exact likelihoods while VAEs cannot compute any probability
D. VAEs offer tractable likelihood estimates but often blurrier samples, while GANs yield sharper samples without explicit likelihoods

52 Why are diffusion models generally more stable to train than GANs despite being more computationally expensive at inference?

Describe generative AI model types Hard
A. They require no neural network and use closed-form solutions
B. They train a single step, avoiding all iterative computation
C. They discard the data distribution entirely and sample uniformly
D. They optimize a well-defined denoising regression objective rather than a min-max adversarial game

53 An autoregressive image model like PixelCNN differs from a GAN primarily in that it:

Describe generative AI model types Hard
A. Cannot generate images and only classifies them
B. Models the explicit joint distribution as a product of pixel-wise conditionals, enabling exact likelihood
C. Uses a discriminator network to score full images at once
D. Generates all pixels simultaneously in a single parallel pass

54 Normalizing flows achieve exact likelihood computation through which key architectural constraint?

Describe generative AI model types Hard
A. Invertible transformations with tractable Jacobian determinants relating data and latent densities
B. Discrete latent codebooks quantized via nearest neighbor
C. Adversarial competition between two subnetworks
D. Iterative noise injection followed by learned denoising

55 A transformer-based LLM and an RNN-based language model both generate text autoregressively. What is the transformer's core architectural advantage during training?

Describe generative AI model types Hard
A. Parallel processing of all sequence positions via attention instead of sequential recurrence
B. It eliminates the need for a probability distribution over tokens
C. It uses fewer parameters by sharing a single weight across all layers
D. It requires no positional information because order is irrelevant

56 In a retrieval-augmented generation (RAG) system, why does grounding an LLM with retrieved documents primarily reduce hallucination?

Describe generative AI applications Hard
A. It increases the model's parameter count during inference
B. It retrains the model weights on each query in real time
C. It replaces the softmax with a deterministic lookup, removing all randomness
D. It conditions generation on relevant external evidence, anchoring outputs to verifiable source content

57 A company deploys a generative model for synthetic tabular data to train downstream classifiers. What is the key risk that undermines this application if unaddressed?

Describe generative AI applications Hard
A. The classifier will refuse to train on any non-real data
B. The synthetic data may memorize and leak sensitive records or fail to preserve real statistical dependencies
C. Synthetic data always improves classifier accuracy regardless of quality
D. Generative models cannot produce tabular data, only images

58 For real-time interactive applications (e.g., live chat), why might a diffusion-based text generator be a poor choice compared to an autoregressive transformer?

Describe generative AI applications Hard
A. Diffusion requires many iterative denoising steps, increasing latency, while transformers generate tokens more directly
B. Diffusion produces only single-word outputs by design
C. Diffusion models cannot represent language at all
D. Autoregressive models require retraining for each new message

59 When using a text-to-image model for product design, which limitation best explains why fine-grained spatial instructions (e.g., 'exactly three buttons in a row') often fail?

Describe generative AI applications Hard
A. Text prompts are ignored entirely during image synthesis
B. The model captures statistical associations rather than precise compositional or counting constraints
C. The model has no access to any training images
D. The model can only output grayscale images

60 A developer uses an LLM for code generation and observes plausible-looking but nonexistent library functions. What underlying property of the model best explains this?

Describe generative AI applications Hard
A. It only outputs code it has verified by executing it internally
B. It cannot produce code and only produces prose
C. It generates the most statistically likely tokens, which can produce syntactically plausible yet factually invalid API calls
D. It intentionally sabotages code to protect proprietary libraries
Next MCQ
Unit 2