AI-102 Azure AI Engineer Associate - Set 1 - Part 1

Correct Answer: Generative AI models

Generative AI specializes in creating new, novel content and combinations, making it ideal for generating personalized treatment plans based on existing medical knowledge and patient data.

Option a is incorrect because expert systems focus on decision-making and analysis using predefined rules rather than creating new, original content like personalized treatment combinations.

Option c is incorrect because traditional NLP chatbots are designed for understanding and responding to queries, not for generating novel treatment plans or creative medical solutions.

Option d is incorrect because image recognition systems analyze and classify visual data but cannot generate new treatment plans or synthesize therapeutic combinations for patients.

Correct Answer: GenAI encompasses both capabilities

Generative AI can both interpret and analyze data while also generating new content, making it suitable for sentiment analysis and creative marketing copy generation simultaneously.

Option a is incorrect because modern GenAI models are designed to handle both analytical tasks and content generation within a single framework, eliminating the need for separate systems.

Option c is incorrect because expert systems are rule-based and focus on specific decision-making tasks, lacking the creative content generation capability required for marketing copy.

Option d is incorrect because while NLP helps with understanding text, it does not inherently include the generative capabilities needed to create original marketing copy from scratch.

Correct Answer: Simulation of human behaviors

The materials explicitly state that AI creates simulation by mimicking behaviors rather than emulation, which would replicate exact biological processes or virtualize the human mind.

Option d is incorrect because emulation involves replicating exact biological processes or virtualizing the mind, which AI does not accomplish; it only simulates behavior patterns.

Option b is incorrect because AI does not virtualize consciousness or replicate the actual biological mechanisms of awareness; it simulates intelligent behavior without replicating underlying biological processes.

Option c is incorrect because AI does not replicate biological processes of the brain; it uses mathematical computations to simulate intelligent behaviors without biological replication.

Correct Answer: Foundational model

Foundational models are pre-trained on vast amounts of diverse data from multiple sources and can later be fine-tuned for specific tasks like medical imaging.

Option a is incorrect because specialized diagnostic algorithms are typically designed for narrow, specific tasks from the outset rather than being broadly pre-trained on diverse data types.

Option b is incorrect because traditional machine learning models usually require task-specific training data rather than being pre-trained on massive diverse datasets before specialization.

Option c is incorrect because reinforcement learning agents learn through trial and error in environments rather than being pre-trained on diverse text, image, and audio datasets.

Correct Answer: Continuous feedback loop mechanism

The decoder's process of feeding generated tokens back as input to predict subsequent words exemplifies the continuous feedback loop that enables sequential text generation in LLMs.

Option a is incorrect because multi-head attention allows focusing on different input parts simultaneously, not the sequential word-by-word generation and feedback process described in the scenario.

Option b is incorrect because tokenization breaks input text into smaller units before processing, rather than describing the iterative generation process where outputs become inputs.

Option d is incorrect because positional encoding preserves word order information in parallel processing, not the sequential feedback mechanism where each generated word influences the next.

Correct Answer: The black box effect

The materials specifically refer to the black box concept, describing how the reasoning process within LLMs hidden layers is highly complex and not fully understood by researchers.

Option a is incorrect because the transformer paradox is not a term used in the materials; the actual term describing unexplained internal reasoning is the black box effect.

Option b is incorrect because deep learning mystery is not the specific terminology used; the materials explicitly reference the black box to describe this incompletely understood reasoning process.

Option d is incorrect because while hidden layers are mentioned as complex, the specific term used in the materials for this phenomenon is the black box effect.

Correct Answer: Parallel processing capability

Transformers revolutionized NLP by processing data in parallel rather than sequentially, directly solving the scaling issues that plagued earlier recurrent neural network models.

Option a is incorrect because tokenization is the process of breaking text into smaller units, not the architectural innovation that enabled simultaneous processing of all words.

Option b is incorrect because while attention is important, the fundamental breakthrough was parallel processing capability; attention mechanisms determine importance but don't enable simultaneous processing alone.

Option d is incorrect because deep neural networks existed before Transformers; the specific innovation was parallel processing architecture, not simply having multiple layers.

Correct Answer: Encoder reads and decoder generates

The encoder creates semantic representations of input text while the decoder generates output sequences, with cross-attention enabling the decoder to map relationships between source and target languages.

Option b is incorrect because tokenization breaks text into units and vocabulary matching assigns IDs, but these preprocessing steps do not handle the semantic cross-language mapping during generation.

Option d is incorrect because positional encoding preserves word order information but does not perform the actual understanding and generation tasks required for cross-language translation.

Option c is incorrect because while multi-head attention is important, it works within the encoder-decoder framework; the component combination specifically performing translation is encoder and decoder together.

Correct Answer: Positional encoding

Positional encoding is specifically designed to preserve word order information in Transformers, compensating for the parallel processing that would otherwise lose sequential relationships.

Option a is incorrect because tokenization ordering assigns IDs to text units but does not embed positional information into the model's processing to maintain sequential understanding.

Option b is incorrect because multi-head attention allows focusing on different parts simultaneously but does not inherently encode which word came first or last in the sequence.

Option d is incorrect because vector embeddings represent semantic relationships between words but do not specifically encode the sequential position of words in the input text.

Correct Answer: Context window exceeded

The context window represents the combined input and output token limit; with nine thousand tokens already used from a ten thousand limit, insufficient capacity remains for processing.

Option a is incorrect because while memory is a factor, the specific technical constraint being enforced is the context window limit, which manages both memory and compute costs.

Option b is incorrect because tokenization algorithm failure would prevent processing any input; the issue is that processing would succeed but exceed the predetermined context window constraint.

Option d is incorrect because embedding space overflow is not a standard term; the actual limiting factor is the context window, which manages total token capacity.

Correct Answer: Tokenization algorithms differ

Different models use different tokenization algorithms like Byte Pair Encoding or WordPiece, which results in varying approaches to breaking text into tokens.

Option a is incorrect because embedding algorithms convert tokens to vectors after tokenization occurs; they do not determine how words are initially split into processing units.

Option b is incorrect because while vocabulary size affects tokenization, the specific reason models split words differently is the distinct tokenization algorithms they employ, not just vocabulary size.

Option d is incorrect because attention mechanism types determine how the model processes relationships between tokens, not how the original text is broken down into tokens.

Correct Answer: Vector embeddings

Vector embeddings represent text as points in high-dimensional space where semantic similarity is captured by distance, enabling the system to find conceptually related content regardless of exact wording.

Option a is incorrect because keyword matching only finds exact word matches and would miss semantically similar phrases like slow response and delayed replies that use different terminology.

Option b is incorrect because tokenization breaks text into processing units but does not create semantic representations that enable finding conceptually similar content with different words.

Option c is incorrect because attention mechanisms determine importance of words within sequences during processing but do not create the persistent similarity representations needed for search.

Correct Answer: Similar features or categories

In vector spaces, distance represents relationship strength; close proximity indicates semantic or contextual similarity, meaning the products likely share similar features, uses, or categories.

Option a is incorrect because vector embeddings capture semantic and contextual relationships from text, not numerical business data like pricing that would not be encoded in descriptions.

Option d is incorrect because inventory levels are operational data not encoded in product description embeddings; vector proximity reflects semantic content similarity from the text itself.

Option c is incorrect because manufacturing location details would only create vector proximity if extensively discussed in descriptions; semantic similarity indicates conceptual relatedness not geographic data.

Correct Answer: Cross-attention between encoder and decoder

Cross-attention specifically enables the decoder to attend to the encoder's output, creating the crucial mapping between source language input and target language output during translation.

Option a is incorrect because self-attention operates within a single sequence, helping words attend to other words in the same sentence, not mapping between different languages.

Option b is incorrect because multi-head attention runs multiple attention processes in parallel but is not the specific mechanism that maps between encoder input and decoder output.

Option d is incorrect because positional attention is not a standard term; the specific mechanism for input-output mapping across languages is cross-attention between encoder and decoder.

Correct Answer: Self-attention within the sequence

Self-attention allows words in a sequence to attend to other words in the same sequence, enabling the model to use context like steep to disambiguate meanings.

Option a is incorrect because cross-attention operates between different sequences like encoder and decoder, not within a single sentence where contextual disambiguation occurs through self-attention.

Option b is incorrect because while multi-head attention runs multiple processes simultaneously, the specific mechanism examining word relationships within one sentence to determine meaning is self-attention.

Option c is incorrect because positional encoding attention is not a standard mechanism; positional encoding preserves word order while self-attention examines contextual relationships for meaning.

Correct Answer: Multi-head attention

Multi-head attention runs multiple attention processes in parallel, allowing the model to capture various dependencies and relationships such as syntax, semantics, and grammar simultaneously.

Option a is incorrect because cross-attention specifically maps between encoder and decoder sequences, not the parallel processing of multiple relationship types within attention mechanisms.

Option b is incorrect because self-attention is one type of attention mechanism, but the specific capability to run multiple attention processes simultaneously is called multi-head attention.

Option d is incorrect because positional attention encoding is not a standard term; the mechanism for capturing multiple dependencies simultaneously is multi-head attention.

Correct Answer: Supervised learning

Supervised learning uses labeled data where the correct answer is provided, allowing the model to learn patterns that predict precise outcomes like classification tasks.

Option a is incorrect because unsupervised learning works with unlabeled data to find patterns and structures, not labeled examples with correct answers for prediction tasks.

Option b is incorrect because reinforcement learning involves learning through trial and error in an environment with rewards, not learning from pre-labeled image datasets.

Option c is incorrect because self-directed learning is not a standard machine learning paradigm; the correct term for learning from labeled data is supervised learning.

Correct Answer: Unsupervised learning

Unsupervised learning works with unlabeled data to discover structures and patterns like clustering, making it ideal for finding natural customer groupings without predefined categories.

Option a is incorrect because supervised classification requires labeled data with known categories, whereas discovering natural groupings without labels is an unsupervised learning task.

Option b is incorrect because reinforcement optimization involves learning through environmental interaction and rewards, not discovering patterns in static unlabeled purchase data.

Option d is incorrect because labeled regression is a supervised learning task requiring known outputs, not the unsupervised pattern discovery needed to find natural customer groupings.

Correct Answer: Reinforcement learning

Reinforcement learning involves learning from environmental interaction through trial and error, maximizing rewards and minimizing penalties to improve behavior over time.

Option a is incorrect because supervised navigation would require pre-labeled examples of correct paths, whereas the robot learns through environmental feedback rather than labeled training data.

Option c is incorrect because unsupervised learning discovers patterns in unlabeled data, not learning through environmental rewards and penalties as described in the scenario.

Option d is incorrect because labeled pathfinding would involve supervised learning with correct paths provided, not the trial-and-error reward-based learning described in the scenario.

Correct Answer: Deep learning network

The materials specifically define deep learning as a neural network with three or more hidden layers, making this five-layer network a deep learning architecture.

Option a is incorrect because while technically accurate, the specific term used in the materials for networks with three or more hidden layers is deep learning network.

Option b is incorrect because advanced neural network is not the precise terminology; the materials use deep learning to specifically describe networks with three or more hidden layers.

Option d is incorrect because complex processing system is too vague; the materials provide the specific term deep learning for neural networks with three or more hidden layers.