Microsoft AI-900 Azure AI Fundamentals 1-10

In machine learning, the step of checking the state of the data and shaping it into an appropriate form before training a model is extremely important. This step is called data preparation.
Data preparation includes checking and imputing missing values, detecting outliers, understanding the data distribution, normalization and standardization, and encoding categorical variables.

Option B, model evaluation, is the step of checking performance after training; option C, model deployment, is releasing the model to production; and option D, hyperparameter tuning, is optimizing the settings of the learning method.
Train machine learning models with Azure Machine Learning

In Azure Machine Learning designer, to divide a single dataset into “training” and “test” sets, you use the Split Data module.
This module lets you split the data based on a specified ratio (for example, 70% / 30%) or on conditions, establishing the prerequisites for properly performing training and evaluation.

Option B, Select Columns, chooses which features to keep or drop; option C, Join Data, merges multiple datasets; and option D, Normalize Data, adjusts the numeric scale.
Split Data component

Finding a specific target within an image (in this case, “cars”) and indicating the location of each with a bounding box (rectangular frame) corresponds to object detection in computer vision. A characteristic of object detection is that it can identify not only “what is shown” but also “where it is shown” at the same time.

By contrast, image classification only assigns a single label to the whole image and does not handle location information.
OCR is a technology for recognizing text within an image, and sentiment analysis is a process that estimates emotions mainly from text or speech.
Object detection

In Azure AI Language service, the feature that determines the emotional tone of text, such as “positive,” “negative,” or “neutral” is sentiment analysis.
Sentiment analysis returns a sentiment score or classification result for the whole document or per sentence, and is widely used for tasks such as analyzing customer reviews and evaluating social media posts.

By contrast, key phrase extraction pulls out important terms, named entity recognition identifies items such as people and place names, and language detection determines the language of the text.
What is sentiment analysis and opinion mining?

The feature that enables a chatbot to respond to purposeless everyday greetings and small talk such as “Hello” or “How are you?” is called chit-chat.
Rather than resolving user questions or handling business tasks, chit-chat aims to increase friendliness and encourage continued use through natural conversation.

Multi-turn conversation refers to the mechanism for maintaining a dialogue across multiple turns, and is one of the underlying technologies for realizing chit-chat.
Active learning is a method for improving learning efficiency, and knowledge mining is a technology for extracting information from documents.

Therefore, the most appropriate term for the feature that handles greetings and small talk itself is “chit-chat.”
Use chit-chat in your project

To identify a specific kind of animal (for example, a rare species of bird), when you want to train a model using images you have collected yourself, Azure AI Custom Vision is the best choice. Custom Vision is a service that lets you label images you prepare and easily build custom image classification or object detection models.

Option A, Azure AI Vision, mainly provides prebuilt models for general-purpose image analysis (such as tagging and caption generation) and cannot be trained for a specific target.
Option C, Azure AI Face, is dedicated to face recognition, and option D, Azure AI Document Intelligence, is for document analysis.
Azure AI Custom Vision

Machine learning that predicts continuous numeric values, such as future “sales amount” or “temperature,” is the regression method.
In regression, the model learns the relationship between the input features and the numeric value (label) to be predicted, and outputs the result as a real number.

By contrast, classification is a method that predicts discrete categories such as “approve / reject” or “yes / no.”
Clustering groups data that has no correct labels based on similarity, and is not used directly to predict future values.
How to select algorithms for Azure Machine Learning

The idea that “AI systems should empower and engage everyone, including people with disabilities” refers to the principle of inclusiveness.
Inclusiveness emphasizes designing and providing AI that everyone can use and participate in, regardless of differences in age, gender, culture, or physical ability.
It also includes ensuring accessibility (support for assistive technologies, easy-to-use UIs, and so on).

Fairness is mainly about avoiding unfair bias, transparency about the understandability of the mechanism, and accountability about clarifying where responsibility lies. The one that best matches the intent of this question is inclusiveness.
What is responsible AI?

Among the Azure OpenAI models, the one used to generate high-quality images from text is DALL·E.
DALL·E is an image generation model that can create images reflecting the composition and style based on a prompt described in natural language (for example, “a robot walking on a beach at sunset”).

GPT-4 is a large language model mainly for text understanding and generation, Whisper is for speech recognition (speech to text), and Codex is a model specialized in code generation and completion.

Therefore, the model best suited for generating images from text is DALL·E.
How to use Azure OpenAI image generation models

Anomaly detection is a machine learning approach aimed at finding behavior or values that deviate from normal patterns.
Typical uses include detecting fraudulent transactions, predicting equipment failures, and monitoring sensor values for abnormalities, and it analyzes time-series data to identify changes that differ from normal conditions.

Option A describes natural language processing, and options B and D describe image classification, which are different from the essence of anomaly detection.

Therefore, C, which identifies abnormal changes based on how the data evolves, is the correct machine learning approach that can be used for anomaly detection.
AI Anomaly Detector