WorksheetsMachine Learning and Generative AI — MCQ Worksheet Extraction
Total questions: 111
Worksheet time: 56mins
MCQ — LEC 1: Which statement best describes supervised learning in machine learning?
The algorithm is trained using only unlabeled data and finds patterns without any guidance.
The algorithm tries to generate new data similar to the training data without any specific labels
The algorithm learns by receiving rewards or penalties for its actions in a simulated environment.
The algorithm is given input data paired with the correct output labels and learns to predict new outputs based on this labeled data.
MCQ — LEC 1: ………………… deals with predicting categorical target variables, which represent discrete classes or labels.
Classification
Regression
Clustering
Association
MCQ — LEC 1: Which task is most appropriate for a regression algorithm?
Sorting emails into “spam” or “not spam” folders.
Estimating the selling price of a used car based on its characteristics.
Grouping photos by the people in them without labels.
Identifying faces in crowded images and marking their positions.
MCQ — LEC 1: The primary goal of Unsupervised learning is…………………………
To predict future outcomes based on historical labeled data and maximize prediction accuracy
To discover hidden patterns, similarities, or clusters within data for purposes such as data exploration, visualization, and dimensionality reduction
To classify data into predefined categories using training examples with known correct answers
To optimize decision-making processes through trial-and-error interactions with an environment
MCQ — LEC 1: What is association rule learning?
A supervised learning technique that predicts customer purchasing behavior based on demographic data and past transaction labels
A clustering algorithm that groups similar customers together based on their shopping preferences and spending patterns
A rule-based machine learning technique used to discover interesting relationships between variables in large datasets, particularly known for market basket analysis
A reinforcement learning method that optimizes product recommendations by learning from customer feedback and purchase rewards.
MCQ — LEC 1: What are the main characteristics that distinguish reinforcement learning from other machine learning approaches?
Uses labeled datasets, requires human supervision, and focuses on finding patterns in static data without environmental interaction
Discovers hidden structures in data, performs clustering and dimensionality reduction, and works with unlabeled datasets exclusively
Learns through trial-and-error interactions with an environment, uses reward feedback to improve performance, and excels at autonomous decision-making
Processes sequential text data, generates natural language responses, and relies on attention mechanisms for understanding context
MCQ — LEC 1: What are Neural Networks (NNs) in the context of deep learning?
The building blocks of deep learning that mimic the structure of the brain to identify patterns and make decisions
Statistical models that use linear regression techniques to analyze data relationships and make predictions based on correlation coefficients
Database management systems designed to store and retrieve large amounts of structured data efficiently using indexing algorithms
Programming frameworks that provide pre-written code libraries for implementing machine learning algorithms and data processing tasks
MCQ — LEC 1: What are the two main components of a Generative Adversarial Network (GAN)?
Encoder and decoder networks
Generator and discriminator networks
Input and output layers
Training and testing datasets
MCQ — LEC 1: RNNs are specifically designed to process ……………… data and can maintain ……………… to remember information from previous steps in a sequence.
static image; long-term memory
tabular; permanent storage
sequential; short-term memory
categorical; external memory
MCQ — LEC 1: Which statement best explains why transformers revolutionized NLP compared to previous architectures?
They process words sequentially one at a time, maintaining better memory of previous contexts
They require less computational power by processing only the most relevant words in a sentence
They generate synthetic text by competing with discriminator networks through adversarial training
They use attention mechanisms to consider all words simultaneously and weigh their importance
MCQ — LEC 1: What is the main difference between traditional AI and on-device AI?
Traditional AI uses smaller models while on-device AI requires large cloud servers for processing
Traditional AI relies on cloud servers to process data from edge devices, while on-device AI operates directly on devices using integrated chipsets and smaller LLMs
Traditional AI focuses on mobile applications while on-device AI is designed only for desktop computers
Traditional AI processes data faster while on-device AI provides more accurate results through cloud computing
MCQ — LEC 2: What does Generative AI refer to?
A set of artificial intelligence methodologies that can predict future events
A set of artificial intelligence methodologies that can produce novel content resembling their training data
A set of artificial intelligence methodologies that only analyze data
A set of artificial intelligence methodologies that create physical objects
MCQ — LEC 2: Which scenario best illustrates the concept of a foundation model?
A model trained specifically for medical diagnosis that only works with X-ray images and cannot be used for other medical tasks
A small specialized model designed to perform one specific function like spam email detection with high accuracy
A large pre-trained model like GPT that can be fine-tuned for various tasks such as writing, coding, translation, and summarization
A model that requires training from scratch for each new application without leveraging any previous learning
MCQ — LEC 2: What is the primary purpose of Gen AI Studio?
To allow developers to quickly explore, customize, and deploy GenAI models on Google Cloud with collaborative tools and resources
To provide cloud storage solutions for large datasets used in traditional machine learning applications
To offer computational resources for training AI models from scratch without any pre-built components
To serve as a marketplace for purchasing pre-trained AI models from third-party vendors
MCQ — LEC 2: What is the main issue with LLM consistency?
Always provide the same answer regardless of context
They produce wildly different answers when asked the same question multiple times
They only work with specific types of questions
They require constant internet connection to function
MCQ — LEC 2: Why do LLMs often exhibit bias in their responses?
They inherit biases present in their internet-scraped training data
They are designed to favor certain user groups
They lack sufficient computational power
They are programmed with specific political viewpoints
MCQ — LEC 2: In the context of LLMs, what are hallucinations?
Visual distortions that occur when the model processes image inputs incorrectly
Words or phrases generated by the model that are often nonsensical or grammatically incorrect
Temporary system failures that cause the model to stop responding to user queries
Intentional creative outputs designed to enhance the model's storytelling capabilities
MCQ — LEC 2: What makes LLM interpretability challenging?
They process data too slowly for analysis
They require specialized hardware to operate
It's difficult to understand why they generate particular responses
They only work with text-based inputs
MCQ — LEC 2: The main memory challenge for LLMs is ……………………
They cannot remember any previous conversations
They have unlimited memory capacity
They can only store numerical data
Efficiently remembering important details over long conversations remains difficult
MCQ — LEC 2: What limitation do LLMs have regarding current information?
They can only process historical data from decades ago
They lack information about changes that occurred after their training cutoff
They cannot access any external data sources
They require manual updates every few hours
MCQ — LEC 2: What is the primary purpose of PaLM API in Google's generative AI ecosystem?
To serve as a simple entry point for developers to access Google's LLMs optimized for various use cases
To provide a complex multi-step authentication system for accessing Google's cloud services
To offer data storage solutions for large-scale machine learning datasets
To provide hardware acceleration for training custom neural networks from scratch
MCQ — LEC 2: How does MakerSuite facilitate generative AI application development?
By requiring extensive coding knowledge and manual model training processes
By providing an approachable platform for prototyping, prompt iteration, dataset augmentation, and model customization
By focusing exclusively on image generation and computer vision tasks
By offering only pre-built applications without customization options
MCQ — LEC 3: What is the primary function of sequence-to-sequence (seq2seq) models?
To classify images into categories
To map an input sequence to an output sequence
To detect anomalies in large datasets
To generate static numerical summaries of tabular data
What is one of the main issues associated with processing long sequences in RNNs?
Increased accuracy
Faster training
Reduced memory usage
Slow computations
Which problem in RNNs occurs when gradients become too large during training?
Vanished memory
Slow computations
Exploding gradients
Vanishing gradients
What does the term 'vanished memory' refer to in the context of RNNs?
The ability to store large datasets
The loss of information from earlier time steps during training
The increase in computational speed
The improvement in gradient flow
Which of the following is NOT listed as a problem with RNNs?
Vanishing gradients
Exploding gradients
Perfect memory retention
Slow computations
What is the eventual size of the gradient when vanishing gradients occur?
Infinitely large
Infinitely small
Moderately stable
Equal to one
What happens to the gradient size in each step due to vanishing gradients?
It becomes smaller and smaller
It remains constant
It becomes larger and larger
It fluctuates randomly
What happens to the gradient when it is multiplied by a number larger than one in the context of exploding gradients?
It keeps increasing
It remains constant
It becomes smaller
It becomes zero
What can happen to the parameters of a network when gradients become extremely large?
They become negative
They stabilize automatically
They decrease to zero
They result in NaN values
What is a potential drawback of word-based tokenization?
Short sequences
Out-of-vocabulary (OOV) words
Small vocabulary size
Easy semantic capture
Which tokenization method breaks words into smaller units and is used in GPT?
Character-based tokenization
Byte-Pair Encoding (BPE)
Word-based tokenization
Sentence Piece
Which tokenization method is used in BERT and breaks words into smaller units?
Sentence Piece
Byte-Pair Encoding (BPE)
Character-based tokenization
Word Piece
Why are positional embeddings used in Transformers?
To increase the vocabulary size
To encode the position or order of tokens
To replace attention scores
To process sequences sequentially like RNNs
What does the attention scores matrix represent in Transformers?
The difference between token positions
The total number of tokens
The similarity between query and key vectors of tokens
The weight of the output sequence
What is the primary role of the query vector (Q) for each token?
To store the token's content
To determine the attention it receives
To measure its relevance to other tokens
To mix data based on relevance scores
What is the function of the key vector (K) in relation to other tokens?
To store the actual information of a token
To compare against query vectors to determine attention
To measure relevance to other tokens
To mix data based on relevance scores
What does the value vector (V) represent for each token?
The relevance score to other tokens
The actual information or content used if deemed relevant
The comparison metric for attention allocation
The position of the token in the sequence
What are the "heads" in multi-head attention?
Sequential layers in the Transformer
Output tokens
Single query vectors
Parallel attention mechanisms
What is a key characteristic of multi-head attention in Transformers?
It processes tokens sequentially
It uses multiple parallel attention mechanisms
It relies solely on a single attention score
It does not use query, key, or value vectors
What is a key requirement of a causal model?
It must not be able to see future words
It must predict future words accurately
It must process all words simultaneously
It must ignore past words
What do the key and value vectors produced by the encoder represent?
The order of tokens in the output sequence
The contextual information and what the model has learned about the input
The probabilities of the next token
The similarity between input and output tokens
How does the decoder generate the output sequence?
By processing the entire sequence at once
By ignoring the input sequence
One token at a time
By using only key vectors
What is a potential drawback of using greedy search?
It always leads to optimal results
It can lead to suboptimal results due to its lack of exploration
It increases the speed of decoding
It eliminates the need for SoftMax
What is the primary purpose of fine-tuning a pre-trained Large Language Model (LLM)?
To erase all previous knowledge
To adapt it for a specific task or domain
To increase the size of the dataset
To train it from scratch
What does it mean when a model is described as pre-trained?
It has been trained only on a small dataset
It has been exposed to a massive dataset with general knowledge
It has never been trained before
It is designed for a single task
What is a key characteristic of Self-Supervised Learning (SSL) in domain-specific fine-tuning?
It requires fully labeled datasets
It generates its own labels from the input data
It only works with small datasets
It ignores the surrounding context
How does Self-Supervised Learning help models like LLMs?
By predicting missing words based on context
By limiting the use of unlabeled data
By avoiding the use of training corpora
By focusing only on labeled data
How does supervised fine-tuning improve a model's performance?
By mapping inputs to correct outputs based on explicit feedback
By ignoring question-answer pairs
By limiting the model to general knowledge
What is the core of the Reinforcement Learning from Human Feedback (RLHF) procedure?
Using only supervision and demonstration data
Establishing and training a reward model (RM)
Ignoring human feedback
Fine-tuning without a reward model
How does RLHF improve an LLM's behavior?
By collecting feedback from humans to train a reward model
By limiting the model to unsupervised learning
By avoiding candidate response exploration
By aligning with machine preferences only
What is the main benefit of Low-Ranking Adaptation (LoRA) in fine-tuning LLMs?
It increases the number of parameters
It reduces computational and financial costs using low-rank approximation
It requires full 16-bit finetuning
It avoids specific task adaptation
How does Parameter Efficient Fine Tuning (PEFT) enhance LLM performance?
By fine-tuning all parameters
By adjusting a small set of parameters to reduce costs and combat forgetting
By limiting performance to general tasks
By avoiding multimodal applications
What is the main benefit of Low Ranking Adaptation (LoRA) in fine-tuning LLMs?
It increases the number of parameters
It reduces computational and financial costs using low-rank approximation
It requires full 16-bit finetuning
It avoids specific task adaptation
During fine-tuning of a pre-trained model like DistilBERT for a specific task, what strategy can help reduce catastrophic forgetting?
Fine-tuning with a larger batch size
Using Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA
Increasing the maximum sequence length during tokenization
Training on noisy, un-labelled data
What is a primary cause of model drift in a fine-tuned model?
Insufficient computational resources
Changes in data distribution or environment
Overfitting during initial training
Lack of model complexity
What is a potential consequence of bias amplification during fine-tuning?
Improved model accuracy
Increased computational efficiency
Ethical concerns due to biased predictions
Reduced need for re-fine-tuning
What is the primary function of the reward model in the RLHF process?
To generate initial text outputs for evaluation
To assign a score to the base model's output based on human preferences
To directly update the base model's parameters
To reduce the computational resources needed for fine-tuning
How does the reinforcement learning algorithm (e.g., PPO) utilize the reward model’s scores in RLHF?
It uses the scores to generate new training data
It adjusts the base model to favor high-reward outputs and avoid lowerward ones
It replaces the base model with a new model
It ignores low-reward outputs without updating the model
What is the purpose of human annotators ranking responses in the reward model training process?
To directly update the fine-tuned language model
To provide preference data to train the reward model to predict humanpreferred scores
To generate new prompts for the language model
To reduce bias in the fine-tuned model's outputs
Why is fine-tuning a pre-trained model more efficient than training a model from scratch?
It requires no computational resources
It skips initial training stages and converges faster with less data
It eliminates the need for task-specific adaptation
It avoids the use of pre-trained model knowledge
What is a key benefit of fine-tuning in continuous learning scenarios?
It eliminates the need for any initial training
It allows periodic model updates to adapt to changing data without starting from scratch
It ensures the model never requires additional data
It completely removes biases from the pre-trained model
What is a prompt in the context of NLP and LLMs?
A pre-trained model output
An input provided to the model to generate a response or prediction
A type of fine-tuning technique
A method to evaluate model performance
What does prompt engineering involve?
Only asking simple questions
Designing high-quality prompts, tinkering for the best prompt, and optimizing length and style
Ignoring creativity and logic
Avoiding task-specific adjustments
What is the primary function of an instruction in a prompt?
To provide background context
To tell the model what action to take, e.g., "Summarize the following text"
To specify the output format
To process input data
How does an output indicator enhance a prompt?
By providing additional context
By reducing the prompt length
By guiding the model on the desired response format, e.g., "In the style of Shakespeare"
By limiting the input data
How does context enhance a prompt?
By giving the model a broader scenario or background, e.g., "Considering the economic downturn"
By limiting the input data
By defining the output style
By reducing the prompt length
What defines a zero-shot prompt?
It includes multiple examples
It provides a task description and some text without examples
It relies on specific training data
It requires prior model training on the task
What is a key ability tested by the zero-shot technique?
The model's ability to generalize and produce outputs without prior examples
The model's ability to memorize training data
The model's ability to follow complex instructions
The model's ability to process large datasets
What is the main difference between zero-shot and one-shot prompting?
Zero-shot uses multiple examples, while one-shot uses none
Zero-shot provides no examples, while one-shot uses a single example
One-shot requires specific training data, while zero-shot does not
Zero-shot is more complex than one-shot
How does one-shot prompting help the model perform a task?
By providing multiple demonstrations
By avoiding any examples
By relying on prior training data alone
By showing the model a single example to understand and generate text accordingly
How can the effectiveness of one-shot prompting be improved for difficult tasks?
By reducing the number of examples
By removing the prompt context
By experimenting with increasing demonstrations (e.g., 3-shot, 5-shot, 10-shot)
By limiting the input data
What is the main benefit of chain-of-thought (CoT) prompting?
It simplifies the task by avoiding examples
It limits the model's output to simple responses
It reduces the need for model training
It enables complex reasoning through intermediate steps, improving language understanding
What is the core idea behind Self-Consistency CoT prompting?
It uses a single reasoning path for all tasks
It calls the LM multiple times on the same prompt and selects the most consistent answer using Majority Voting
It avoids intermediate reasoning steps
It relies solely on standard prompting techniques
How does Self-Consistency CoT prompt mimic human reasoning?
By providing diverse perspectives and encouraging critical evaluation of its own reasoning
By limiting the model to a single example
By avoiding consultation of multiple resources
By focusing only on arithmetic tasks
What does RAG primarily enable a generative AI model to do?
Generate images from text descriptions
Access and use external information that was not part of its original training data/parameters
Train itself in real time on user inputs
Compress its model size while keeping performance
Which of the following is NOT listed as one of the core RAG components on this slide?
Retriever
Ranker
Generator
Planner
In a standard RAG pipeline, what is the role of the Retriever component?
To write the final answer in natural language
To fetch relevant passages or documents from a large corpus of external knowledge
To re-rank the retrieved documents by relevance after the initial retrieval
To store and update the external data corpus
What is the primary purpose of a Vector Database in a RAG system?
To store vector embeddings and enable similarity search for efficient retrieval
To compress the base language model to reduce memory footprint
To generate final answers after retrieval
To train the Retriever component on-the-fly
Before data can be stored and searched in a Vector DB for RAG, what mandatory preprocessing step must be?
Data Chunking – dividing the data into manageable “chunks” or segments
Translation into English
Compression using gzip
Removal of all stop-words
Which of the following is NOT a real-world example of a Vector Database?
Chroma
Pinecone
Weaviate
Redis
What is the main job of the RAG Ranker component?
To fetch the initial set of documents from the vector database
To refine the retrieved documents by assigning relevance and importance, assigning scores and prioritizing the most relevant ones
To generate the final answer in natural language
To split the original documents into chunks
According to RAG, which component is responsible for efficiently pulling the initial relevant information from a huge corpus?
The Ranker
The Retriever
The Generator
The Augmentor
What best describes Naive/Standard RAG?
It uses multiple rounds of retrieval and reasoning
It is the simplest form: chunk documents → embed → retrieve top chunks on query → directly send to LLM for answer
It always rewrites the user query before retrieval
It compresses all retrieved documents into a single sentence
What is the key difference between Naive RAG and Advanced RAG?
Advanced RAG removes the vector database entirely
Advanced RAG only works with images, not text
Advanced RAG uses smaller language models
Advanced RAG adds extra processing steps both before and after the initial retrieval
What is the defining characteristic of Modular RAG?
It adds a loop that corrects factual mistakes after generation
It breaks the traditional RAG pipeline into independent, swappable, re-orderable modules (e.g., separate search, re-ranking, summarization, generation modules)
It only works with a single large vector database
It removes the retrieval step completely
Corrective RAG is specifically designed for scenarios where:
Speed is more important than accuracy
The knowledge base is very small
A factual error in the answer could have serious negative consequences
Users only ask creative questions
What is the key new capability that Agentic RAG introduces compared to all previous RAG variants?
It uses a larger vector database
It removes the generator LLM completely
It only works with images
The system gains autonomy: the model actively decides if/when it needs more information and iteratively refines its own queries
Fusion RAG is distinguished from standard RAG because it:
Uses only one document at a time
Pulls relevant information from multiple heterogeneous data sources and blends/synthesizes them into a single cohesive answer
Relies solely on the LLM’s internal knowledge without any retrieval
Runs everything in parallel on a single GPU
What is currently one of the biggest real-world benefits of Self-RAG?
It is the fastest RAG method
It drives hallucination rates in RAG systems close to zero while keeping high accuracy
It requires no vector database
It works offline without any external data
According to Anthropic, both Workflows and Agents belong to the same broad category. What is that category?
Large Language Models
Agentic Systems
Tool-augmented Models
Reasoning Engines
What is the key characteristic of a Workflow according to Anthropic?
The model dynamically decides the next step at runtime.
The model can change its goals during execution.
It operates without any tools or external functions
The sequence of steps is predefined by the designer.
In an Anthropic-style Agent, who or what decides what to do next during task execution?
A hardcoded script written by the developer.
The user in real time.
The model itself dynamically
A separate supervisor model
Which of the following statements is true about the process flexibility of Agents compared to Workflows (per Anthropic’s distinction)?
Workflows can adapt and replan, while Agents follow a fixed path
Both Workflows and Agents have completely fixed processes
Agents have a fixed process, while Workflows can react and adapt
Agents can react, adapt, and plan as they go, whereas Workflows follow a fixed script
What is Prompt Chaining according to Anthropic?
Asking the same LLM multiple times in parallel for better answers
Breaking a large task into a sequence of smaller, well-defined sub-tasks
Letting the model freely decide the order of operations
Combining multiple unrelated prompts into one giant prompt
In Anthropic’s Prompt Chaining pattern, how are the individual sub-tasks typically executed?
All sub-tasks run simultaneously on different models
Each sub-task is handled by a separate LLM call (or sequence of calls), with results passed forward
The entire chain is executed inside a single LLM call
Sub-tasks are performed only by custom Python code without any LLM
In the diagram shown, what is the purpose of the “Gate” component placed after LLM Call 1?
To add new tools to the next LLM call
To merge outputs from multiple parallel LLM calls
To perform a quality check and decide whether to continue the chain (Pass) or stop early (Exit/Fail)
To rewrite the original user query
Which of the following is NOT listed as a benefit of Prompt Chaining?
Precision
Control
Modularity
Autonomous decision-making
Which statement best reflects Anthropic’s view on Prompt Chaining as presented in this slid?
It is a type of fully agentic system where the model freely decides every next action
It puts the entire process on “guardrails” with every step defined in advance
It is inferior to agentic approaches in every situation
It removes the need for any LLM calls after the first one initial prompt
What is the primary role of the LLM Router in Anthropic’s Routing design pattern?
To generate the final answer itself
To act as a “traffic controller” that decides which specialized (“expert”) model should handle the incoming task
To combine outputs from all specialist models into one response
To translate the user query into a different language before processing
Which of the following is NOT listed as a benefit of using the Routing pattern?
Specialization
Efficiency
Modularity
Full autonomy (the router can change the user’s original request)
What is the core idea of the Parallelization design pattern according to Anthropic?
Running the same task on multiple identical LLMs and voting on the best answer
Splitting a task into parts and having multiple specialist LLMs work on different parts simultaneously, then combining results
Letting one LLM decide everything step-by-step without any parallel work
Sending the entire task to the fastest available LLM
According to Anthropic, Parallelization is classified as:
A fully agentic system
A workflow pattern (not an agent), because the splitting and aggregation logic is coded by humans
Exactly the same as Prompt Chaining
In orchestrator-work pattern, which component is responsible for deciding how to break down the original task?
Human-written code
The user in real time
The Orchestrator LLM
The Synthesizer LLM
What is the key advantage highlighted for using LLMs as both Orchestrator and Synthesizer?
It is cheaper because fewer API calls are needed
It guarantees perfect parallelism every time
It completely eliminates the need for any worker LLMs
The process becomes more dynamic and can better adjust to the specifics of each individual input thanks to the LLM’s reasoning abilities
What is the main purpose of the Evaluator-Optimizer (Validation Agent) pattern?
To parallelize work across many LLMs
To iteratively improve an answer by having one LLM generate solutions and another LLM critically check and reject/improve them
To route the query to the single best specialist LLM
To translate the user query into multiple languages simultaneously
In Evaluator-Optimizer (Validation Agent) pattern, what happens when the Evaluator LLM rejects a proposed solution?
The process stops immediately and returns an error
The rejected solution (plus feedback) is sent back to the Generator LLM for another attempt
The Evaluator automatically fixes the answer itself
The input is forwarded to a completely different model
What is the biggest quality-related benefit of the Evaluator-Optimizer pattern?
It is faster than a single LLM call
Solutions get progressively better each round as the Generator corrects its own mistakes using feedback
It guarantees the answer is 100% factually correct
It reduces token usage to the absolute minimum
The Evaluator-Optimizer pattern delivers “more consistent, high-quality outputs” even when:
Using very small models
The Generator LLM is “messy” or inconsistent on its own
There is no internet access
The user provides very short prompts
In the context of Anthropic’s overall workflow-vs-agent distinction, the Evaluator-Optimizer loop is still categorized as:
A workflow (not an agent), because the loop structure, roles, and exit condition are all predefined by the human designer
A true agent, because it can loop until satisfied
The only pattern that is neither workflow nor agent
The first real agentic pattern in the deck
