Microsoft Azure AI Fundamentals Practice Test (AI-900)

Clustering - This is the correct answer. Clustering is an unsupervised machine learning technique used to group data into segments based on similarities without requiring labeled data. It is ideal for grouping customers into segments based on their behavior.

Regression is used for predicting continuous numerical values based on input features, not for grouping data or segmenting customers.

Classification is a supervised learning technique where the goal is to categorize data into predefined classes. It requires labeled data, which is not available in this case.

Time Series Analysis is used to analyze data that is collected over time (e.g., stock prices, sales data) to identify trends or patterns. It is not focused on segmenting data into groups based on behavior.

Natural Language Processing (NLP) workload - This is the correct answer. Analyzing customer reviews to determine overall satisfaction involves processing and understanding text data, which is exactly what Natural Language Processing (NLP) is designed for. NLP techniques such as sentiment analysis can be applied to assess customer sentiment from reviews and feedback.

Computer Vision workload - Computer vision is used for analyzing visual data, such as images or videos.

Content Moderation workload - Content moderation focuses on filtering inappropriate or harmful content.

Document Intelligence workload - Document intelligence focuses on understanding and extracting structured information from documents, often for tasks like form processing or invoice extraction.

This statement is False.

Azure OpenAI Service's image generation capabilities, such as those offered by the DALL-E model, generate images based on textual descriptions provided by the user, not audio prompts. Users input text prompts that describe the desired image, and the service generates images accordingly. Audio input is not a supported method for image generation in Azure OpenAI Service.

Generative AI models learn from massive text and image corpora that contain historical and social biases. Those biases can surface or even amplify in model outputs. Responsible AI guidance from Microsoft requires developers to continuously identify, measure, and mitigate potential harms by combining automated tools (such as content filters) with human-in-the-loop reviews, red-team testing, and ongoing monitoring. Ignoring these steps risks unfair, stereotyped, or otherwise harmful generations. Therefore, the only correct statement is that bias can occur and must be addressed through sustained human oversight.

Automated Machine Learning enables data scientists to automatically train and tune models by selecting the best algorithms and hyperparameters for a given dataset.

Azure Machine Learning Designer provides a visual interface for building models, it still requires manual selection of algorithms and setup.

Azure Databricks and Azure Synapse Analytics are comprehensive data processing platforms but do not specialize in automated model training and selection.

Codex is the Azure OpenAI model specifically designed for code generation tasks, allowing developers to transform natural language prompts into code in various programming languages.

GPT-3's text-davinci-003 is powerful for natural language understanding and generation but is not optimized for code generation.

DALL·E is used for image generation.

Azure's Computer Vision API is intended for analyzing visual content, not generating code.

They create new content by modeling the structure of the training data - This is the correct answer. Generative AI models create new and original content by learning the underlying structure, patterns, and relationships within the training data. This allows them to generate creative outputs, like story ideas, that are inspired by but not directly copied from the original material.

They store exact copies of data for direct replication - This is characteristic of retrieval-based systems, which retrieve exact content from a database or dataset. This does not apply to generative AI, which creates novel content rather than copying it.

They categorize input data into predefined classes - This is the function of classification models, which categorize data into specific classes based on predefined labels. It does not involve generating new content.

They detect anomalies and irregularities in data - This describes anomaly detection models, which focus on identifying outliers or unusual patterns in data, not generating new and original content.

Key Phrase Extraction is the process of automatically identifying the most significant and relevant phrases within a text. This technique helps in summarizing the main topics or issues discussed, enabling the company to quickly understand customer concerns.

Sentiment analysis identifies the emotional tone behind words, language detection identifies the language of the text, and entity recognition identifies and classifies named entities, such as names of people or organizations, which may not provide the overall themes or issues.

Document intelligence workloads involve extracting structured information from unstructured or semi-structured documents like invoices, receipts, and forms. A service that extracts key information from invoices and receipts is an example of this, as it automates data extraction from documents.

The other options represent different AI workloads: translating spoken language into text is speech recognition (natural language processing), recommending products based on user preferences is personalization, and identifying objects in images is computer vision.

Classification - This is the correct answer. Fraud detection is a classification task because the model needs to classify each transaction as either fraudulent or legitimate, which involves assigning data points to predefined categories or labels.

Regression is used for predicting continuous numerical values for example sales forecasts or prices, not for classifying transactions into categories such as "fraudulent" or "legitimate."

Clustering is an unsupervised learning technique used to group data based on similarities, but it is not suitable for determining whether a transaction is fraudulent or legitimate, which requires labeled data and a classification approach.

Dimensionality Reduction is used to reduce the number of features in the data, typically for improving performance or visualization, but it is not a task in itself for determining fraud or legitimacy.

The Codex model in Azure OpenAI Service is specifically designed for translating natural language input into programming code. It supports multiple programming languages and can interpret user intent to generate accurate code snippets, making it ideal for this scenario.

GPT-3 is capable of understanding and generating human-like text, it is not optimized for code generation tasks.

DALL-E focuses on creating images from textual descriptions.

Embeddings are used for semantic understanding and similarity tasks, not code generation.

Regression is the appropriate technique for predicting continuous numerical values, such as the number of customer service calls. It models the relationship between dependent and independent variables to forecast future values.

Classification is used for predicting categorical outcomes

Clustering groups data points based on similarity without prior labels.

Generative AI - This is the correct answer. A system that creates original artworks and designs based on existing art styles would fall under generative AI. Generative models, like GANs (Generative Adversarial Networks), can learn from existing artworks and create new, unique pieces in similar styles.

Computer Vision - Computer vision is used for analyzing and processing visual data, such as recognizing objects or identifying patterns in images. While it could be used as part of the system (e.g., for analyzing art), it is not the primary workload for creating original art.

Knowledge Mining - Knowledge mining involves extracting insights and patterns from unstructured data like documents and media files. It is not focused on generating new creative content such as artworks.

Predictive Analytics - Predictive analytics uses historical data to make predictions about future events or trends. It is not suitable for creating original artworks or designs based on existing styles.

Deep Learning - This is the correct answer. Deep learning is the most appropriate technique for analyzing unstructured data like images and speech, as it can automatically learn hierarchical representations from raw data. Deep learning models, especially neural networks, excel at processing and extracting meaningful features from complex data without requiring extensive manual feature engineering.

Clustering Algorithms are used for grouping data into clusters based on similarity, but they are not designed to automatically learn hierarchical representations from raw unstructured data like images and speech.

Regression Algorithms are used for predicting continuous numerical outcomes based on input features, but they are not suited for learning hierarchical representations from unstructured data.

Decision Trees are used for classification or regression tasks by splitting data into decision rules based on features. They require predefined features and are less effective than deep learning for automatically learning hierarchical representations from raw, unstructured data like images and speech.

Predicting the content category of an image - This is the correct answer. Image classification involves assigning a label or category to an image based on its content. For example, an image classification solution might categorize an image as "cat," "dog," or "car" based on the objects it contains.

Identifying and locating objects within an image - This describes object detection, not image classification. Object detection involves both identifying objects and locating them within the image (often with bounding boxes), while classification only assigns a category label.

Extracting text from images for text analysis - This is Optical Character Recognition (OCR), which is focused on extracting text from images, not classifying the entire image into categories.

Detecting facial features to analyze emotions - This is facial recognition or emotion detection, which focuses on detecting and analyzing facial features, often for applications like identifying emotions or individuals, but it is not image classification.

To evaluate the model's performance on unseen data, helping to prevent overfitting - This is the correct answer. Dividing a dataset into training and validation subsets allows the model to learn from the training data while being evaluated on a separate validation set. This helps assess the model's ability to generalize to new, unseen data, preventing overfitting where the model memorizes the training data but performs poorly on new data.

To increase the total amount of data available for training - This is incorrect because splitting the data reduces the amount of data available for training. The goal is to use a portion for training and another for validation, not to increase the total amount of data.

To reduce the training time by using smaller datasets - Splitting the data does not specifically aim to reduce training time. It’s more about evaluating model performance and preventing overfitting.

To ensure the model memorizes the training data perfectly - The goal is not for the model to memorize the training data. In fact, memorization (overfitting) is something practitioners want to avoid. The focus is on ensuring the model generalizes well to new, unseen data.

Sentiment Analysis is used to determine the emotional tone behind words, identifying and categorizing opinions expressed in text to understand the customers feelings.

Entity Recognition focuses on identifying and classifying named entities within text, such as people or organizations, which doesn't provide insights into emotions.

Key Phrase Extraction identifies important keywords or phrases but doesn't assess emotional context.

Topic Modeling discovers abstract topics within documents but doesn't directly evaluate emotional sentiment.

In machine learning, the label is the variable that represents the output or the value the model is trying to predict. In this scenario, the model aims to predict the 'Final Selling Price' of products, making it the label.

The other variables such as 'Production Cost', 'Marketing Budget' and 'Competitor Prices' are features that the model uses to make the prediction.

Implementing robust encryption techniques for data at rest and in transit - This is the correct answer. To address concerns about privacy and security, the company should implement strong encryption techniques to protect user data both when it is stored at rest and when it is transmitted in transit. This ensures that sensitive data is secure and reduces the risk of unauthorized access.

Collecting as many data points as possible to improve recommendations - While collecting data may help improve recommendations, it raises privacy concerns if personal data is not properly protected or managed. This approach does not directly address the need for secure handling of user data.

Giving developers access to user data for debugging purposes - Giving developers access to user data can create significant privacy and security risks. It is important to ensure that user data is only accessible to those who have a legitimate need and that proper safeguards are in place.

Storing user data on shared servers to reduce costs - Storing user data on shared servers without adequate security measures can expose the data to higher risks of breaches and unauthorized access. It is essential to store user data securely, even if it means higher costs for private or dedicated infrastructure.

Entity Recognition - This is the correct answer. Entity Recognition (also known as Named Entity Recognition or NER) is a natural language processing technique used to extract structured information such as names of people, places, organizations, dates and other specific entities from unstructured text documents.

Key Phrase Extraction - Key phrase extraction identifies important phrases in text but does not specifically target named entities like people, places, or dates.

Sentiment Analysis - Sentiment analysis is used to evaluate the emotional tone or sentiment (positive, negative, neutral) in text, not to extract specific entities.

Language Translation - Language translation is used to convert text from one language to another, but it does not focus on extracting structured information like names or dates.