Azure AI 900 Study Guide

Introduction to Artificial Intelligence

Core Concepts of AI
 Artificial Intelligence (AI): Machines mimicking human behavior.
 Machine Learning (ML): Machines improving at tasks without explicit
programming.
 Deep Learning: Utilizes artificial neural networks inspired by the human
brain for complex problem-solving.
 Data Scientist: Possesses skills in math, statistics, predictive modeling, and
machine learning for future predictions.
Data Concepts
 Dataset: Logical grouping of closely related data units (e.g., MNIST, COCO).
 Data Labeling: Identifying raw data and adding informative labels for
machine learning model context.
 Supervised Learning (SL): Labeled data for training.
 Unsupervised Learning (UL): Data without labels, requiring the model to
label itself.
 Reinforcement Learning (RL): Environment-driven learning without
predefined data.
Neural Networks and Deep Learning
 Neural Networks (NN): Nodes organized into layers for ML model training.

Internal
Azure AI 900 Study Guide

 Deep Neural Network (DNN): NN with 3+ hidden layers, considered deep

learning.
 Backpropagation (BP): Adjusts weights in NN to improve outcomes.
 Loss Function: Compares ground truth to predictions to determine error
rate.
 Activation Functions: Algorithms applied to hidden layer nodes (e.g., ReLu).
Machine Learning Pipeline
Pipeline Stages
 Pre Processing: Data preparation and feature engineering before model
training.
 Data Cleaning: Correcting errors in datasets to improve results.
 Feature Engineering: Transforming data for ML model ingestion.
 Sampling: Balancing datasets across labels.
 Post Processing: Translating model output into human-readable format.
Training and Inference
 Training: Process of training the ML model.
 Serving: Deploying the model for use at an endpoint.
 Inference: Invoking the model for predictions.
 Real-time Endpoint: Optimized for quick single-item results.
 Batch-transform Endpoint: Optimized for larger batch predictions.
Evaluation and Performance Metrics
Metrics for Evaluation
 Classification Metrics: Accuracy, F1 Score, Precision, Recall.
 Regression Metrics: MSE, RMSE, MAE.
 Regression models are trained to predict numeric label values based on
training data that includes both features and known labels. The process for
training a regression model (or indeed, any supervised machine learning
model) involves multiple iterations in which you use an appropriate algorithm
(usually with some parameterized settings) to train a model, evaluate the
model's predictive performance, and refine the model by repeating the
training process with different algorithms and parameters until you achieve
an acceptable level of predictive accuracy.
Tools and Visualization

Internal
Azure AI 900 Study Guide

 Jupyter Notebook: Web-based tool for combining code, text, and

visualizations.
 Confusion Matrix: Visualizes model predictions against ground truth labels.
Cognitive Services and Responsible AI
Cognitive Services
 Anomaly Detector: Identifies potential problems early.
 Language Services: Includes Language Understanding, QnA Maker, Text
Analytics, Translator.
 Speech Services: Covers Speech to Text, Text to Speech, Speech
Translation, Speaker Recognition.
 Vision Services: Encompasses Computer Vision, Custom Vision, Face
Detection.
Responsible AI Principles
 Fairness: AI systems should treat all people fairly.
 Reliability and Safety: AI systems should perform reliably and safely.
 Privacy and Security: AI systems should be secure and respect privacy.
Microsoft AI Principles
Fairness
 AI systems should treat all people fairly.
 Example: Ensuring unbiased decision-making in loan approvals.
Reliability and Safety
 AI systems should perform reliably and safely.
 Case Study: Implementing fail-safes in autonomous vehicles to prevent
accidents.
Privacy and Security
 AI systems should be secure and respect privacy.
 Reference: GDPR regulations for protecting user data in AI applications.
Inclusiveness
 AI systems should empower everyone and engage people.
 Example: Developing AI tools for accessibility in education.
Transparency

Internal
Azure AI 900 Study Guide

 AI systems should be understandable.

 Historical Context: The need for transparency in algorithmic decision-making.
Accountability
 People should be accountable for AI systems.
 Example: Establishing clear responsibility for AI-generated content.
Common Machine Learning Workloads
Anomaly Detection
 Process of finding outliers within a dataset.
 Example: Identifying fraudulent transactions in financial data.
Computer Vision
 Using ML NN to gain high-level understanding from digital images or video.
 Case Study: Facial recognition technology for security purposes.
Natural Language Processing (NLP)
 ML that can understand the context of a corpus.
 Example: Sentiment analysis of customer reviews.
Conversational AI
 Technology that can participate in conversations with humans.
 Application: Chatbots for customer service interactions.
Azure Machine Learning Service
Authoring Tools
 Notebooks: Jupyter Notebooks for writing Python code.
 AutoML: Automated process for building and training ML models.
 Designer: Visual drag-and-drop tool for constructing ML pipelines.
Assets
 Datasets: Uploaded data used for training, can be versioned.
 Experiments: Logical groupings of runs.
 Models: Registry containing trained models for deployment.
Manage
 Compute Instances: Development workstations for data scientists.
 Environments: Reproducible Python environments for ML experiments.

Internal
Azure AI 900 Study Guide

 Datastores: Securely connect to storage services on Azure.

Text Analytics
Sentiment Analysis
 Identifying sentiments like negative, positive, mixed, or neutral.
 Example: Analyzing social media sentiment towards a product launch.
Named Entity Recognition (NER)
 Detects words and phrases in unstructured text associated with semantic
types.
 Application: Extracting entities like names, dates, and locations from text.
Language Understanding (LUIS)
 No-code ML service for building natural language into apps.
 Components: Intentions, entities, and utterances for training ML models.
Generative AI and Transformer Models
Generative AI
 Focuses on creating new and realistic content or data.
 Example: Generating art, music, or text using AI algorithms.
Large Language Model (LLM)
 Processes and generates human-like text based on massive data.
 Use Case: Chatbots providing natural language responses.
Transformer Model
 Specialized in understanding and generating language.
 Components: Encoder reads input text, decoder generates new text.
Tokenization and Embeddings in Transformer Models
Tokenization Process
 Tokenization in transformer models involves breaking down sentences into
'tokens' to help computers understand language.
 Tokens can represent words or parts of words, aiding in capturing the
meaning of the text.
 Example: Sentence 'I heard a dog bark loudly at a cat' can be tokenized into
individual words or subwords.
 This process is crucial for natural language processing tasks in AI models.

Internal
Azure AI 900 Study Guide

Embeddings in Transformer Models

 Each token in a transformer model is assigned a unique numeric code known
as an embedding.
 Embeddings act as a secret code capturing the semantic meaning of the
corresponding word.
 They help the model understand and process the text data effectively.
 Example: In GPT-4, embeddings play a vital role in representing words during
text generation.
Positional Encoding and Attention in Transformer Models
Positional Encoding
 Technique to maintain the word order in transformer models like GPT to
preserve sentence meaning.
 Prevents the model from losing the sequential information during processing.
 Example: Ensures 'I love pizza' is interpreted differently from 'Pizza loves I'.
 Crucial for accurate natural language understanding in AI systems.
Attention Mechanism
 Allows models to focus on specific parts of input data, enhancing context
comprehension.
 Similar to human attention, it helps prioritize relevant information for better
predictions.
 Improves the model's ability to understand relationships between words in a
sentence.
 Example: Attention mechanism in GPT models aids in generating coherent
text outputs.
Azure OpenAI Service and Model Types
Azure OpenAI Service
 Cloud platform for deploying and managing advanced language models from
OpenAI.
 Integrates OpenAI's latest model developments with Azure's secure and
scalable infrastructure.
 Enables the deployment of various AI models for different applications and
tasks.
 Example: Azure OpenAI Service combines GPT-4 and GPT-3.5 models for text
and code generation.

Internal
Azure AI 900 Study Guide

Model Types in Azure OpenAI

 GPT-4 Models: Latest in GPT series, capable of generating text and code from
natural language prompts.
 GPT-3.5 Models: Similar to GPT-4, with a focus on text and code generation.
 Embedding Models: Convert text into numerical sequences for analysis and
comparison.
 DALL-E Models: Create images from textual descriptions, still in testing phase.
 Example: GPT-3.5-turbo model is optimized for conversational AI applications.
Copilots and Prompt Engineering
Copilots in AI Development
 Tools that assist users in common tasks using generative AI models.
 Integrated into applications to provide real-time support and content
generation.
 Enable developers to create custom copilots tailored to specific business
needs.
 Example: Microsoft Copilot aids in document creation, code writing, and
strategic planning.
Prompt Engineering Techniques
 Process of refining prompts to improve AI responses and user interactions.
 Involves defining system messages to set context and constraints for model
responses.
 Writing precise and explicit prompts enhances the quality of AI-generated
outputs.
 Example: Zero-shot learning allows AI models to perform tasks without prior
training.
Responsible AI Guidelines by Microsoft
Stages of Responsible AI Implementation
 Identify: Recognize potential harms associated with AI solutions.
 Measure: Evaluate the occurrence of harms in AI outputs.
 Mitigate: Take steps to reduce harms and communicate risks to users.
 Operate: Follow a comprehensive plan for deploying and managing AI
models.

Internal
Azure AI 900 Study Guide

 Example: Microsoft's guidance emphasizes ethical and responsible AI

practices.

Internal