As I embarked on preparing for the Azure AI-900 exam, I made several observations and took notes that helped me navigate the process more effectively. Below are some tips I wish I had known before starting the exam preparation, along with my summarized insights (a crash course 🙂) from the Microsoft AI Fundamentals course.

Tips for Exam Preparation

Use Your Personal Computer
Avoid using a restricted laptop, whether it's for work, school, or any other purpose. I learned this lesson the hard way. If you don’t have a personal computer available, consider borrowing one from a friend. I attempted to use my work laptop and ran into network restrictions that caused me to miss my exam—don't make the same mistake!
Commit to the Training Modules
The official training module is an excellent starting point. With consistent effort, you can complete it in about a week by dedicating 2-3 hours each day. I found that sticking to this schedule helped me stay on track and retain key concepts.
Take the Free Practice Tests
Don't underestimate the value of practice tests. They’re free and can be taken multiple times, allowing you to track your progress. Start taking them as soon as you begin studying. Aim for at least a 90% score on the practice tests before attempting the actual exam.
Expect Different Questions on the Exam
While practice tests are useful for understanding the fundamentals, don’t expect the exact same questions on the actual exam. Practice tests are there to help you master the material, not to memorize specific answers 😊.
Select the Right Proficiency Level
When taking the exam, you'll have the option to choose your proficiency level for each section. If you have some experience with AI, starting at a medium proficiency level can be helpful. If needed, you can retake the exam at a lower proficiency level, so don’t stress about choosing the highest level initially.
Take Advantage of Discounts
You can receive up to 50% discount on the exam if you attend one of Microsoft's virtual training days. This is a great opportunity to gain additional insights and save on the exam cost.

Now, shifting gears to my summarized insights from the prep course, I’d like to share some key takeaways that helped me better understand the concepts and solidify my preparation for the exam.

Disclaimer:
These notes serve as a summarized guide based on my personal understanding of the course materials. They are not a comprehensive or official resource for exam preparation. While I’ve made every effort to ensure their accuracy, I highly recommend reviewing the authorized course materials and official resources provided by the course for the most complete and accurate preparation. I cannot be held responsible for any outcomes based solely on the use of these notes.

Machine Learning Overview

Machine learning involves using features (input data) and labels (predicted outcomes) to build models. The process of defining the relationship between features and labels is known as training. Once trained, the model can be used to predict new values in a process called inferencing.

Supervised Learning: The model is trained using labeled data.
- Regression: Predicts numeric values (e.g., predicting house prices).
- Classification: Predicts categories (e.g., spam or not spam).
  - Binary Classification: Two possible categories.
  - Multiclass Classification: More than two categories.
Unsupervised Learning: The model is trained using only features, with no labels.
- Clustering: Groups similar data points together, useful for exploratory analysis and as a foundation for classification models.
Common ML Algorithms:
- Linear Regression: Used for regression tasks; evaluates model fit with metrics like R-squared (R²) and Root Mean Squared Error (RMSE).
- Logistic Regression: Used for binary classification; evaluated using a confusion matrix (True Positives, False Positives, False Negatives, True Negatives).
- Multiclass Classification: Uses strategies like One-vs-Rest or Multinomial algorithms.
- K-means Clustering: A common algorithm for unsupervised learning tasks, grouping data points based on similarity.
Other Key Concepts:
- Multiple Linear Regression: Assumes that the features are independent of each other.
- Deep Learning: Uses artificial neural networks (ANNs) to model complex relationships, inspired by how the human brain learns. The loss function in deep learning evaluates the difference between predicted and actual values.
Azure Services for developing and deploying machine learning models: Azure Machine Learning Studio (Automated ML), Azure ML Designer

Computer Vision

Computer vision involves analyzing and interpreting images and videos to detect objects, faces, and text. It’s a field of AI that relies on deep learning techniques and pixel manipulation to extract meaningful insights from visual data.

Deep Learning forms the foundation for modern image classification solutions, enabling powerful image recognition capabilities.
Computer Vision Techniques:
- Convolutional Filtering: A technique used to process and analyze image pixels, enabling the identification of important features within images.
- Convolutional Neural Networks (CNNs): A type of deep learning model designed specifically for image classification and recognition tasks. CNNs are highly effective at identifying spatial hierarchies in image data.
Key Computer Vision Taks:
- Facial Detection: Identifies and analyzes human faces in images.
- Image Classification: Categorizes the entire image into specific classes based on its contents.
- Object Detection: Identifies the location of objects in an image using bounding boxes, which highlight areas containing specific objects.
- Semantic Segmentation: Classifies individual pixels in an image to recognize the objects they represent, enabling fine-grained analysis.
- Multi-modal Models: Combines computer vision with generative AI to improve the understanding of images and text.
- Optical Character Recognition (OCR): Extracts text from images, such as handwriting, signs, or printed text.
Face Detection: Involves identifying areas in an image that contain human faces. Azure AI Face offers robust facial analysis capabilities, including:
- Detecting attributes such as glasses or headwear.
- Face identification (1-to-many matching) and face verification (1-to-1 matching) for identifying if two faces are the same person.
Specialized Domain Support: Azure AI Vision offers specialized recognition for domains like:
- Celebrities: Recognizing and identifying famous individuals.
- Landmarks: Detecting and identifying landmarks in images.
Tagging: Associates metadata with images to describe their attributes, making it easier to organize and search for images based on their content.
Azure Services for Computer Vision: Azure AI Face, Azure AI Vision, Azure AI Custom Vision

Natural Language Processing (NLP)

NLP is the technology that enables machines to understand, interpret, and generate human language. It encompasses various tasks such as speech recognition, text analytics, translation, and conversational AI.

Steps in Analyzing Text:
- Preprocessing:
  - Text Normalization: Standardizing text to a consistent format.
  - Stop Word Removal: Eliminating common words (e.g., "the", "and") that don't add significant meaning.
  - N-grams: Breaking text into sequences of words to capture context.
  - Stemming: Reducing words to their base form (e.g., "running" → "run").
- Tokenization: Splitting text into smaller units like words or sentences.
- Frequency Analysis:
  - TF-IDF (Term Frequency-Inverse Document Frequency): A statistical measure used to evaluate the importance of a word within a document relative to its frequency in other documents.
Common NLP Tasks:
- Text Analysis: Extracting insights from raw text data.
- Sentiment Analysis: Determining the emotional tone behind text.
- Machine Translation: Automatically translating text between languages.
- Summarization: Condensing large amounts of text into a shorter form while preserving key information.
- Conversational AI: Building bots that can understand and respond to natural language.
Conversational Language Understanding (CLU):
- CLU focuses on predicting the overall intention of an incoming utterance and extracting key entities from it.
- It enables machines to understand and respond effectively to conversational inputs, such as those in chatbots or virtual assistants.
Azure Bot Service:
- Provides the platform for creating and managing conversational bots.
- Integrates with CLU or LUIS (Language Understanding Intelligent Service) to interpret and respond to user inputs in a conversational manner.
Speech Recognition: Converts spoken language (in real-time or batch) into processable text.
Speech Synthesis: Converts written text into spoken language.
Translation in Azure:
- Azure AI Translator: Provides text-to-text translation between different languages.
- Azure AI Speech: Enables both speech-to-text and speech-to-speech translation, utilizing the Neural Machine Translation (NMT) model for accurate, context-aware translation.
The Language Detection feature in Azure AI Language Service identifies the language of a given text and provides:
- ISO 6391 code.
- Language name.
- Confidence score.
NLP tasks in Azure return a confidence score of NaN (Not a Number) for unknown languages.
Azure Services for NLP: Azure AI Language Service, Azure AI Bot Service, Azure AI Speech Service, Azure AI Translator

Document Intelligence & Knowledge Mining

Data Mining involves searching and indexing vast amounts of data to uncover patterns and insights. Knowledge Mining, on the other hand, focuses on making large datasets easily searchable and accessible, transforming raw data into valuable, actionable information.

Azure AI Document Intelligence

Azure AI Document Intelligence leverages both prebuilt and custom models to analyze documents and forms, extracting relevant content such as text, layout, and key-value pairs.
Key Features:
- Document Analysis: Enables detailed extraction from documents.
- Prebuilt Models: Tailored for common document types like invoices, ID documents, receipts, and business cards.
- Custom Models: Allows the creation of models designed to recognize and extract data from specialized or unique document formats.
- Built on OCR: Utilizes Optical Character Recognition (OCR) to recognize text in scanned or image-based documents.

Indexing and Knowledge Base Creation

An Indexer converts source documents into JSON format, which is then processed by a search engine for efficient indexing and retrieval.
To generate Q&A for a knowledge base, sources can include:
- An existing FAQ document.
- A webpage containing structured information.
- Manually entered data (e.g., product details, policies).

Azure AI Search:

Azure AI Search provides a powerful, private enterprise search solution for organizations. It helps build indexes that allow both internal and public-facing content to be searched efficiently.
The tool can be used to build custom indexes for internal knowledge bases or for exposing searchable content on public websites.

Generative AI

Generative AI refers to models that understand and generate natural language, creating content based on the patterns and structures they've learned from large datasets.

GPT Model:

GPT (Generative Pretrained Transformer) is designed to understand and generate natural language, allowing for tasks such as text generation, translation, and summarization.

Transformer Models:

Encoder: Creates a semantic representation of the input vocabulary.
Decoder: Generates new language sequences based on the encoded information.
- GPT models use the decoder part of the transformer architecture to generate text.

Transformers and GPT:

A transformer model like GPT-4 (the model behind ChatGPT and Bing) processes text input (called a prompt) and generates syntactically correct and contextually relevant output (called a completion). The model’s power lies in its ability to produce coherent and meaningful text based on the given input.

Some Key Elements in Generative AI:

Tokenization: This process breaks text into smaller units (tokens), such as words or subwords, to process input data.
Embeddings: These are vector representations of tokens that capture their semantic meaning, enabling the model to understand relationships between words.
Attention: The model places emphasis on important tokens during text generation, allowing it to focus on contextually significant parts of the input.
System messages: help set the context for the model by outlining expectations, constraints, and response styles. These messages guide the AI’s behavior during interactions.
Grounding: refers to the process of linking AI-generated outputs to reliable and verifiable data sources. This ensures factual accuracy, contextual relevance, and real-world applicability, reducing the risk of hallucinations (false or misleading information).

Embeddings in Azure OpenAI:

Azure OpenAI utilizes embeddings to perform tasks like search, classification, and text comparison, helping to identify similarities across text sources.

Language Models:

Large Language Models (LLMs): These models are trained on massive datasets and can generate complex, coherent responses.
Small Language Models (SLMs): These models are more compact but still capable of performing language generation tasks with limited scope.

Azure AI Foundry:

Azure AI Foundry is a comprehensive platform for AI development, providing various tools and studios for building AI solutions.

Azure AI Agent Service:

This service helps create, test, and manage AI agents within the Azure AI Foundry.

Responsible AI Stages:

Identify potential harm: Recognize areas where the AI model might cause harm.
Measure the presence of harm: Evaluate the impact or likelihood of harm.
Mitigate the harm: Take steps to reduce or eliminate the risk.
Operate the model: Ensure the model is running with proper safeguards in place.
Mitigation Layers:
- Model: Addressing issues at the model level, such as biases.
- Safety Systems: Implementing content filters or moderation tools.
- Metaprompt/System Messages & Grounding: Setting up context and relying on verifiable data sources.
- User Experience: Designing the AI interaction to ensure transparency and clarity.

Generative AI Image Features:

Create a New Image: Generate entirely new images from text descriptions.
Create Variations of an Image: Modify or generate different versions of an existing image.
Edit an Image: Make adjustments to an image while retaining its original features.
- Example: DALL-E, a model that can generate and manipulate images from text.

Azure OpenAI Service:

The Azure OpenAI Service provides a cloud platform for deploying, customizing, and hosting large language models, such as GPT, enabling the development of generative AI solutions.

Azure AI-900: My Exam Prep Journey and Essential Tips

Table of contents