AI-900 Azure AI Fundamentals - Set 1 - Part 1

Correct Answer: Artificial intelligence umbrella

Artificial intelligence is the overarching term that encompasses all software imitating human behaviors and capabilities, including machine learning, computer vision, natural language processing, and anomaly detection working in combination.

Option a is incorrect because deep learning is only one subset of techniques used in AI systems and does not encompass all the different capabilities mentioned in the scenario.

Option c is incorrect because machine learning specifically focuses on systems that learn and predict without explicit programming, not the broader range of AI capabilities like natural language processing.

Option d is incorrect because data science refers to the discipline and professionals who build models, not the technology capabilities that enable systems to mimic human behavior.

Correct Answer: Machine learning foundation

Machine learning is the foundation of AI systems that can learn from data and improve predictions over time without explicit programming, exactly matching the requirement described in the scenario.

Option a is incorrect because conversational AI specifically focuses on dialogue and communication with humans, not learning from sales data to make predictions about product recommendations automatically.

Option b is incorrect because computer vision deals with interpreting visual information from images and videos, not analyzing sales data patterns to improve recommendations over time.

Option c is incorrect because natural language understanding processes human language contexts, not historical sales patterns, and would not be the primary component for recommendation improvement.

Correct Answer: Anomaly detection capabilities

Anomaly detection specifically identifies outliers and unusual patterns that deviate from normal behavior, making it the most directly applicable AI element for detecting fraudulent transactions in the scenario.

Option b is incorrect because while machine learning can be used for fraud detection, the specific capability needed here is detecting outliers, which is more precisely described as anomaly detection.

Option c is incorrect because natural language processing deals with understanding and processing human language, not analyzing numerical transaction patterns to identify unusual spending behaviors in financial data.

Option d is incorrect because computer vision involves interpreting visual information from images or videos, not analyzing transaction data patterns to detect unusual spending behaviors in financial systems.

Correct Answer: Natural language processing

Natural language processing enables systems to understand, process, and contextualize human language in text form, making it essential for reading medical reports and understanding symptom descriptions effectively.

Option a is incorrect because anomaly detection focuses on identifying outliers and unusual patterns in data, not understanding and processing the contextual meaning of written medical reports and documentation.

Option b is incorrect because computer vision deals with interpreting visual information like images and videos, not processing textual information from medical reports and doctor notes about symptoms.

Option d is incorrect because while classification is part of the solution, natural language processing is the more specific and central capability needed to understand textual medical information.

Correct Answer: Conversational AI capability

Conversational AI specifically enables systems to hold natural dialogues with humans, maintain context across exchanges, and respond appropriately to follow-up questions in an interactive manner as required.

Option a is incorrect because while natural language understanding is a component, conversational AI is the broader capability that encompasses dialogue management, context maintenance, and interactive response generation.

Option b is incorrect because computer vision focuses on interpreting visual information from images and videos, not engaging in text-based or voice-based interactive conversations with human users.

Option d is incorrect because anomaly detection identifies unusual patterns or outliers in data, not maintaining conversational context or generating appropriate dialogue responses for customer service interactions.

Correct Answer: MNIST handwritten digits

The MNIST database contains images of handwritten digits specifically designed for testing classification and clustering algorithms related to translating handwriting into digital text, making it ideal for this purpose.

Option a is incorrect because the Common Objects in Context dataset focuses on object detection and segmentation in general images, not specifically on handwritten character recognition for translation purposes.

Option c is incorrect because Azure open datasets is a collection of various datasets, not a specific dataset designed for handwriting recognition, and lacks the focused structure needed for this task.

Option d is incorrect because Coco format refers to a JSON file structure for organizing image annotations, not a dataset itself, and is typically used for object detection tasks.

Correct Answer: Common objects context format

The Common Objects in Context dataset provides comprehensive annotations including object segmentation, recognition, and contextual information within images, making it ideal for complex scene understanding and object localization tasks.

Option a is incorrect because the MNIST database only contains simple handwritten digit images without multiple objects or complex scenes, lacking the segmentation information needed for street photography analysis.

Option b is incorrect because Azure labeled data is a general term for data with annotations, not a specific dataset format that provides the structured object and segmentation information required.

Option d is incorrect because ground truth samples refer to properly labeled training data in general, not a specific dataset format that provides object identification and pixel-level segmentation information.

Correct Answer: Data labeling annotation

Data labeling is the process of identifying raw data and adding meaningful, informative labels to provide context so that machine learning models can learn patterns from the examples.

Option a is incorrect because feature engineering involves extracting and transforming data into numerical formats that models can process, not the initial process of adding descriptive tags to raw data.

Option c is incorrect because model training occurs after data has been labeled and prepared, using the labeled data to teach the model patterns, not the labeling process itself.

Option d is incorrect because ground truth validation refers to having a properly labeled dataset as an objective standard, not the actual process of creating those labels through human review.

Correct Answer: Ground truth dataset

Ground truth refers to a properly labeled dataset that serves as the objective standard for training and assessing models, where the accuracy of labels directly impacts model performance.

Option a is incorrect because while it describes labeled data used for training, it does not convey the authoritative, objective standard nature that ground truth specifically represents in model development.

Option b is incorrect because validation data is used to assess model performance during development, but ground truth specifically emphasizes the properly labeled, authoritative nature as the objective standard.

Option d is incorrect because supervised learning input describes the general category of data used in supervised learning, not the specific concept of an authoritative, properly labeled objective standard.

Correct Answer: Supervised learning regression

Supervised learning with labeled historical data is appropriate here, and regression specifically handles predicting continuous numerical values like house prices from features, making it the ideal approach for this scenario.

Option a is incorrect because unsupervised learning works with unlabeled data to find patterns, but predicting specific house prices requires labeled examples showing the relationship between features and prices.

Option c is incorrect because reinforcement learning involves agents learning through interaction with an environment and rewards, not learning from historical labeled examples to predict continuous numerical values.

Option d is incorrect because while deep learning could potentially be used, supervised learning regression more specifically describes the task-driven approach needed to predict precise numerical house price outcomes.

Correct Answer: Unsupervised clustering analysis

Unsupervised learning with clustering is designed to discover natural groupings and patterns in unlabeled data, making it ideal for finding customer segments without predefined categories in the dataset.

Option a is incorrect because supervised classification requires labeled data with known categories, but the scenario explicitly states there are no predefined categories and the data is unlabeled.

Option b is incorrect because reinforcement learning involves agents making decisions in an environment based on rewards, not discovering patterns or groupings in existing behavioral data without predefined categories.

Option d is incorrect because regression predicts continuous numerical values from labeled data, not discovering hidden groupings or natural segments in unlabeled behavioral patterns and purchasing data.

Correct Answer: Reinforcement learning agents

Reinforcement learning involves agents learning through interaction with an environment, making decisions, receiving feedback through rewards or penalties, and improving over time, perfectly matching the self-driving car scenario.

Option a is incorrect because supervised classification requires labeled training data with known correct answers, but the car must learn through trial and error in an environment, not from pre-labeled examples.

Option b is incorrect because unsupervised learning discovers patterns in unlabeled data, but the scenario requires active decision-making and learning from environmental feedback, not pattern discovery in static data.

Option d is incorrect because clustering groups similar data points without labels, but the scenario requires active decision-making based on environmental feedback and rewards, not finding similarities in data.

Correct Answer: Connection weight value

The weight is the adjustable numerical value assigned to connections between neurons that determines the strength of influence one neuron has on another during data propagation through the network.

Option a is incorrect because activation functions are algorithms applied to nodes to transform their output, not the numerical values assigned to connections that determine influence strength between nodes.

Option c is incorrect because thresholds determine when neurons activate, not the strength of connections between neurons, and do not represent the adjustable influence values between connected nodes.

Option d is incorrect because bias terms are additional parameters added to neurons to shift activation, not the connection strength values between neurons that determine how much one node influences another.

Correct Answer: Backpropagation learning

Backpropagation is the process of moving backwards through a forward feed neural network, adjusting connection weights based on how much each contributed to the error, enabling the network to learn.

Option a is incorrect because forward feed propagation describes the process of moving data forward through the network from input to output, not the backward adjustment of weights for learning.

Option b is incorrect because activation functions are algorithms applied to nodes to transform their output, not the process of adjusting weights by moving backwards through layers to improve performance.

Option d is incorrect because the loss function calculates the error between predictions and actual values, but backpropagation is the process that uses this information to adjust weights throughout the network.

Correct Answer: Sparse dimensionality reduction

When layers progressively decrease in node count, this creates sparsity and achieves dimensionality reduction by condensing information into fewer dimensions, making the representation more compact and focused.

Option a is incorrect because dense layers increase the number of nodes, not decrease them, and do not achieve the information condensing effect described in the scenario.

Option c is incorrect because forward feed propagation describes how data moves through the network from input to output, not the architectural pattern of progressively decreasing node counts.

Option d is incorrect because activation functions are algorithms applied to nodes to transform outputs, not the architectural pattern describing how the number of nodes changes between successive layers.

Correct Answer: Graphics processing unit

Graphics processing units have thousands of cores designed for parallel processing, making them ideal for both rendering graphics and training neural networks which require simultaneous operations on multiple data sets.

Option a is incorrect because central processing units typically have only four to sixteen cores, limiting their parallel processing capabilities compared to the thousands of cores needed for efficient neural network training.

Option c is incorrect because quantum computing processors use quantum mechanics principles for computation, not the massively parallel architecture of thousands of traditional cores that neural networks benefit from most.

Option d is incorrect because while application-specific chips can be optimized for certain tasks, the scenario specifically describes the parallel processing architecture characteristic of graphics processing units with thousands of cores.

Correct Answer: CUDA enabled GPU

CUDA is Nvidia's parallel computing platform with deep learning SDK libraries that provide highly optimized implementations for neural network operations like convolution, making it ideal for efficient deep learning framework integration.

Option a is incorrect because standard CPU threading does not provide the specialized neural network libraries and thousands of parallel cores that deep learning frameworks require for efficient training and inference.

Option c is incorrect because while distributed cloud computing can help with scale, it does not specifically provide the optimized neural network operation libraries that CUDA offers for efficient deep learning computations.

Option d is incorrect because multi-core processing is a general concept that does not specifically reference the CUDA platform with its specialized deep learning libraries optimized for neural network routines.

Correct Answer: Feature engineering extraction

Feature engineering extracts and transforms raw data into numerical formats that machine learning models can process, including converting text into numerical vectors that capture important features like sentiment.

Option a is incorrect because model training occurs after data has been transformed into numerical features, not during the process of converting raw text into numerical vectors for the model.

Option c is incorrect because hyperparameter tuning involves optimizing model parameters for better performance, not transforming raw text data into numerical representations that models can understand and process.

Option d is incorrect because inference deployment makes trained models accessible for predictions, occurring after all preprocessing and training, not during the transformation of raw data into numerical features.

Correct Answer: Hyperparameter tuning

Hyperparameter tuning is the process of systematically testing different model parameter combinations to optimize performance, especially important in deep learning where manual parameter adjustment becomes impractical due to complexity.

Option a is incorrect because feature engineering focuses on extracting and transforming data into formats models can process, not on optimizing model configuration parameters like learning rate and batch size.

Option b is incorrect because data preprocessing involves preparing and cleaning data before training, not the process of testing different model parameter combinations to find optimal configuration settings.

Option d is incorrect because model deployment involves making trained models accessible for predictions, not the optimization process of testing different parameter combinations to improve model performance during development.

Correct Answer: Serving or deploying

Serving or deploying a model makes it accessible by hosting it in infrastructure like virtual machines or containers, enabling users to send requests and receive predictions through endpoints.

Option a is incorrect because model training is the process of teaching the model using data, not making an already trained model accessible to users through scalable infrastructure for prediction requests.

Option b is incorrect because feature engineering transforms data into numerical formats for model consumption, not the process of making trained models accessible through web services or scalable infrastructure endpoints.

Option d is incorrect because hyperparameter optimization tests different parameter combinations to improve model performance during development, not the process of hosting models to make them accessible for user predictions.