Exam AI-900: Microsoft Azure AI Fundamentals

 Describe Artificial Intelligence workloads and considerations (15–20%)

 Describe fundamental principles of machine learning on Azure (15–20%)

 Describe features of computer vision workloads on Azure (15–20%)

 Describe features of Natural Language Processing (NLP) workloads on Azure (15–20%)

 Describe features of generative AI workloads on Azure (20–25%)

1
Describe Artificial Intelligence workloads and considerations
(15–20%)
Identify Features of Common AI Workloads

AI workloads refer to the different types of tasks or problems AI can help solve. The most common AI workloads
include:

Workload Type Description Example

Machine Learning Using data to train models that make Predicting customer churn
predictions
Anomaly Detection Finding unusual patterns in data Fraud detection in transactions
Computer Vision Analyzing visual content like images or Detecting faces in a photo
videos
Natural Language Processing Understanding and generating human Translating languages
(NLP) language
Conversational AI Building bots or voice assistants Chatbots for customer service
Generative AI Creating new content from data Generating text, images, or code

AI Workload Type Features Examples

Common AI Workloads Data processing, model training & Predictive analytics, recommendation
deployment, automation, decision- systems
making, integration
Computer Vision Workloads Image classification, object detection, Security systems, autonomous
semantic segmentation, facial vehicles, image search
recognition, OCR
NLP Workloads Text classification, named entity Sentiment analysis, chatbots,
recognition (NER), language translation, translation apps
question answering, summarization
Document Processing Data extraction, classification, table Invoice processing, contract analysis,
Workloads extraction, validation, compliance email categorization
Generative AI Workloads Content creation, personalization, data AI writing tools, image synthesis
augmentation, creative applications models, music composition

AI Workload Type Key Features Examples

Computer Vision • Image classification • Autonomous vehicles detecting pedestrians and
• Object detection and tracking traffic signs
• Image segmentation • Medical imaging for disease diagnosis
• Facial recognition • Quality control systems in manufacturing
• Pose estimation • Surveillance systems with person/vehicle tracking
• Optical character recognition • Retail analytics tracking customer movement
• Scene understanding • Visual search in e-commerce
Natural Language • Text classification • Customer service chatbots
Processing • Named entity recognition • Email spam detection
• Sentiment analysis • Social media sentiment monitoring
• Language translation • Automatic transcription services
• Text summarization • Language translation apps
• Question answering • Voice assistants (Siri, Alexa)
• Speech recognition • Content recommendation systems
• Topic modeling
Document • Document classification • Invoice processing automation
Processing • Information extraction • Resume parsing for HR
• Form processing • Contract analysis in legal departments
• Table extraction • KYC verification in banking
• Document parsing • Medical record analysis
• Handwriting recognition • Tax document processing
• Document summarization • Insurance claim processing
Generative AI • Text generation • Content creation assistants
• Image generation • DALL-E/Midjourney image creation
• Audio synthesis • Text-to-speech systems
• Video generation • Music composition tools

2
 • Code generation • Code completion tools
• Multimodal capabilities • Conversational AI systems
• Fine-tuning/prompt engineering • Design tools for creative professionals
Each of these workload types requires different infrastructure considerations, specialized model architectures, and
varied approaches to data processing and optimization.

AI Workload Key Features Examples

Category
1. Computer
Vision -Processes and analyzes visual data (images -Object Detection: Autonomous vehicles
and videos). identifying pedestrians and other cars.
-Focuses on tasks like object detection, image
classification, segmentation, facial recognition, -Image Classification: Identifying the
and pose estimation. breed of a dog in a photo.

-Often involves convolutional neural networks -Facial Recognition: Unlocking a

(CNNs). smartphone or security systems
identifying individuals.

-Medical Imaging Analysis: Detecting

anomalies in X-rays or MRI scans.
2. Natural
Language -Deals with understanding, interpreting, and -Sentiment Analysis: Determining the
Processing (NLP) generating human language. positive, negative, or neutral sentiment
of customer reviews.
-Involves tasks like text classification, sentiment
analysis, machine translation, question -Machine Translation: Translating text
answering, and named entity recognition. between different languages.

-Commonly uses recurrent neural networks -Chatbots: Answering customer queries

(RNNs), transformers, and large language in a conversational manner.
models (LLMs).
-Text Summarization: Condensing long
articles into shorter summaries.
3. Document
Processing -Automates the extraction, understanding, and -Invoice Processing: Automatically
organization of information from various extracting key information like vendor
document formats (PDFs, scanned documents, name, invoice number, and amount due.
etc.).
-Legal Document Review: Identifying key
-Combines elements of computer vision (OCR) clauses and entities in contracts.
and NLP.
-Automated Data Entry: Populating
-Focuses on tasks like information extraction, databases from scanned forms.
document classification, data entry automation,
and compliance checks. -Compliance Checking: Ensuring
documents adhere to specific
regulations.
4. Generative AI
-Focuses on creating new, original content that -Text Generation: Writing articles, poems,
resembles the data it was trained on. scripts, or code.

-Can generate text, images, audio, video, code, -Image Generation: Creating realistic or
and more. artistic images from text descriptions.

-Often relies on generative adversarial networks -Music Composition: Generating new

(GANs), variational autoencoders (VAEs), and musical pieces in various styles.
transformer models.
-Synthetic Data Generation: Creating
artificial data for training other AI
models.

3
Identify Computer Vision Workloads
Computer Vision enables computers to understand and interpret visual information. Some common workloads
include:

Computer Vision Task Description Example

Image Classification Classify images into categories Is this a dog or a cat?
Object Detection Identify and locate objects in Detecting vehicles in traffic
images camera feeds
Semantic Segmentation Assign a label to each pixel in an Highlighting roads vs. pedestrians
image
Face Detection & Detect and recognize human Unlocking phones with facial
Recognition faces recognition
Image Analysis Extracting metadata like color, Auto-tagging photos
size, etc.

Feature Description Examples

Image Classification Identifying objects or scenes in images Sorting photos into categories
Object Detection Locating multiple objects in Autonomous vehicles detecting
images/videos obstacles
Semantic Segmentation Classifying each pixel of an image Medical imaging for tumor
identification
Facial Recognition Recognizing and verifying faces Security systems
Optical Character Recognition Extracting text from images or scanned Digitizing printed invoices
(OCR) documents

Feature Description Examples

Image Classification Categorizing images into predefined • Product categorization in e-commerce
classes or labels • Plant/animal species identification
• Medical image classification (X-rays,
MRIs)
• Content moderation systems
Object Detection & Identifying and locating objects within • Autonomous vehicles detecting road
Tracking images and tracking their movement users
across frames • Manufacturing defect detection
• Wildlife monitoring systems
• Sports analytics tracking players
Image Pixel-level classification dividing images • Medical tumor segmentation
Segmentation into segments representing different • Satellite imagery analysis
objects • Augmented reality applications
• Background removal tools
Facial Recognition Identifying or verifying a person's identity • Security access control systems
using facial features • Phone unlocking features
• Automated photo tagging
• Law enforcement applications
Pose Estimation Detecting human figure positioning and • Fitness applications tracking form
joint locations • Animation/motion capture
• Ergonomic monitoring
• Rehabilitation therapy systems
Optical Character Converting handwritten or printed text into • Digitizing documents/books
Recognition (OCR) machine-encoded text • License plate recognition
• Handwriting recognition
• Street sign interpretation
Scene Comprehending the entire visual scene, • Robot navigation
Understanding including spatial relationships • Mixed reality applications
• Smart retail analytics
• Urban planning from imagery
Anomaly Detection Identifying unusual patterns or outliers in • Industrial equipment fault detection
visual data • Security surveillance
• Medical scan anomaly identification
• Quality assurance in manufacturing

4
Feature Description Examples Underlying Key Challenges
Techniques/Models
Input Data Images (static, single Still photographs, Convolutional Neural Handling
frames), videos surveillance footage, Networks (CNNs), variations in
(sequences of frames), medical scans, satellite Recurrent Neural lighting, pose,
real-time video streams. imagery, output from Networks (RNNs) for occlusion,
cameras on robots or video, Transformers viewpoint, scale,
drones. and image quality.
Core Tasks Object detection Autonomous driving CNN architectures Real-time
(locating and classifying (detecting pedestrians, (e.g., ResNet, YOLO, processing
objects), image traffic signs), medical Faster R-CNN, Mask requirements,
classification (assigning diagnosis (identifying R-CNN), Vision need for large
a label to an entire tumors), security Transformers (ViT) labeled datasets,
image), image surveillance robustness to
segmentation (pixel- (identifying intruders), adversarial
level classification), robotic manipulation attacks.
facial recognition, pose (object grasping),
estimation. augmented reality.
Output Bounding boxes around Coordinates and labels Probability scores, Ensuring the
detected objects, class of detected cars, a coordinate values, accuracy and
labels for images or label indicating "cat," a pixel assignments. reliability of the
objects, pixel-wise highlighted region of a output in critical
segmentation masks, tumor in a scan, the applications.
facial identities, key name of a recognized
points representing person, joint angles of
body or object pose. a human body.
Key Data augmentation Using techniques like Optimization for Ethical
Consideratio (creating variations of rotation, scaling, and speed and efficiency, considerations
ns existing data), transfer cropping to increase model interpretability related to privacy
learning (leveraging the size and diversity in some applications. and bias in facial
pre-trained models), of training data; fine- recognition.
computational tuning models trained
resources (GPU on large image
acceleration), real-time datasets.
processing needs.

5
Identify Natural Language Processing (NLP) Workloads
NLP enables machines to understand, interpret, and generate human language.

NLP Workload Description Example

Text Classification Categorize text into topics Spam vs. not spam
Sentiment Analysis Detect emotion in text Positive/negative customer reviews
Language Translation Translate text between languages English to Spanish
Named Entity Recognition Identify entities like people, dates, "John flew to London" → John (person),
(NER) and places London (location)
Text Summarization Generate short summaries of longer Summarize a news article
text
Speech Recognition Convert spoken language to text Voice-to-text on mobile apps

Feature Description Examples

Text Classification Categorizing text into predefined labels Email spam detection
Named Entity Recognition Identifying entities like names and locations Extracting key information from
(NER) in text articles
Language Translation Translating text between languages Multilingual communication
Question Answering Enabling conversational systems or direct Chatbots answering FAQs
responses
Text Summarization Condensing long pieces of text Creating concise summaries of
reports

Natural Language Processing Workloads

Feature Description Examples

Text Classification Categorizing text documents into • Email spam filtering
predefined categories • News article categorization
• Support ticket routing
• Content moderation
Named Entity Identifying and classifying named entities • Information extraction from legal
Recognition (NER) in text into predefined categories documents
• Automated resume parsing
• Research paper analysis
• News monitoring systems
Sentiment Analysis Determining the emotional tone or opinion • Brand reputation monitoring
expressed in text • Customer feedback analysis
• Social media monitoring
• Product review analysis
Machine Translation Converting text from one language to • Real-time translation services
another • Localization of content
• Multilingual support systems
• International document processing
Text Summarization Creating concise versions of longer texts • News article summarization
while preserving key information • Research paper condensation
• Meeting minutes generation
• Legal document summarization
Question Answering Understanding questions in natural • Customer support systems
language and providing accurate answers • Educational tools
• Search engines
• Virtual assistants
Speech Recognition Converting spoken language into text • Transcription services
• Voice assistants
• Call center automation
• Accessibility tools
Topic Modeling Discovering abstract topics in a collection • Content recommendation systems
of documents • Trend analysis in social media
• Document organization
• Research synthesis
Coreference Determining when different mentions refer • Document understanding systems
Resolution to the same entity • Chatbots with memory
• Relationship extraction

6
 • Advanced information retrieval

Feature Description Examples Underlying Key Challenges

Techniques/Models
Input Data Textual data in various Customer reviews, news Recurrent Neural Ambiguity of
forms: documents, articles, tweets, support Networks (RNNs), language
articles, social media tickets, voice commands, LSTMs, GRUs, (polysemy,
posts, emails, speech programming code. Transformers (e.g., homonymy),
transcripts, code. BERT, GPT), Word handling context,
Embeddings understanding
nuances and
sarcasm.
Core Tasks Text classification Spam detection, Language models, Dealing with
(categorizing text), identifying sequence-to- different
sentiment analysis positive/negative sequence models, languages and
(determining emotional feedback, translating a attention cultural contexts,
tone), machine website, answering factual mechanisms. the vastness and
translation, question questions based on a variability of
answering, named entity document, extracting human language.
recognition (identifying names of people and
key entities), text organizations, creating
summarization. short summaries of long
texts.
Output Category labels, "Spam" or "Not Spam," a Probability Evaluating the
sentiment scores, score indicating positive distributions over quality and
translated text, answers sentiment, the translated categories, numerical fluency of
to questions, lists of version of a sentence, a scores, generated generated text,
identified entities, direct answer to a text sequences. ensuring the
concise summaries. question, a list of names accuracy of
and their types, a shorter extracted
version of a document. information.
Key Tokenization (splitting Using libraries like NLTK or Computational cost Ethical
Consideratio text into units), spaCy for tokenization, of training large considerations
ns embedding techniques employing word language models, the related to bias in
(representing words as embeddings like Word2Vec need for large and language models
vectors), handling or GloVe, utilizing diverse text datasets. and the
sequence data, transformer architectures generation of
attention mechanisms to capture context. harmful content.
for long-range
dependencies.

7
Identify Document Processing Workloads
Document processing uses AI to extract data from structured or unstructured documents.

Task Description Example

Form Recognizer Extracts key-value pairs, tables, and text Reading data from invoices or
from forms receipts
OCR (Optical Character Converts printed or handwritten text to Scanning paper documents
Recognition) digital text
Document Classification Categorizes documents into types Classifying resumes vs. invoices
Entity Extraction Identifies and pulls out relevant Extracting names and dates from
information contracts

Feature Description Examples

Data Extraction Pulling relevant data from structured Extracting invoice totals
documents

Classification Organizing documents into categories Sorting emails into folders

Table Extraction Extracting and structuring tabular data from Pulling rows of data from contracts
documents

Validation & Ensuring documents meet specified criteria Checking regulatory compliance in
Compliance filings

Feature Description Examples

Document Categorizing documents based on their • Sorting incoming mail/emails
Classification content or type • Routing documents to departments
• Organizing digital archives
• Legal document categorization
Information Extraction Identifying and extracting specific data • Invoice data extraction
points from documents • Resume information extraction
• Contract clause identification
• Medical record data extraction
Form Processing Extracting structured data from form • Tax form processing
documents • Insurance applications
• Survey processing
• Registration form handling
Table Extraction Identifying and extracting tabular data from • Financial statement analysis
documents • Product catalog processing
• Scientific data extraction
• Inventory document processing
Document Parsing Breaking down documents into structured • PDF to structured data conversion
components • Legal document analysis
• Technical document processing
• Academic paper parsing
Handwriting Converting handwritten text into digital text • Check processing
Recognition • Historical document digitization
• Form field recognition
• Signature verification
Document Creating concise summaries of document • Legal case summarization
Summarization content • Report condensation
• Research paper abstracts
• News briefing generation
Layout Analysis Understanding the spatial arrangement of • Complex form processing
document elements • Magazine/newspaper digitization
• Document reformatting
• Accessibility document conversion
Document Verifying document legitimacy or detecting • Passport/ID verification
Authentication tampering • Certificate authentication
• Fraud detection
• Secure document processing

8
Feature Description Examples Underlying Key Challenges
Techniques/Models
Input Data Documents in various Invoices, contracts, legal Optical Character Handling variations
formats: PDFs, scanned documents, financial Recognition (OCR), in document
documents (images), reports, medical records, Computer Vision for layout, image
Word documents, application forms. layout analysis, Natural quality in scanned
spreadsheets. Often Language Processing documents, and
involves a combination for text understanding. the presence of
of text and visual tables and forms.
elements (tables,
forms).
Core Tasks Information extraction Extracting amounts and CNNs for visual feature Integrating OCR
(identifying and dates from invoices, extraction, RNNs and with NLP
extracting specific data classifying documents as Transformers for effectively,
fields), document "invoice" or "receipt," sequence processing, maintaining
classification automatically filling specialized models for accuracy despite
(categorizing fields in a database from table and form document
documents), data entry a form, extracting data understanding. variations,
automation, table from tables in a report, handling complex
extraction, form identifying required document
understanding, signatures in a legal structures.
compliance checking. document.
Output Structured data (key- A JSON object containing Structured data Ensuring data
value pairs, tables), "Invoice Number": "INV- formats (JSON, CSV), accuracy and
document categories, 123", "Total Amount": database entries, completeness,
automated data "$100.00"; a label "Legal annotations. dealing with noise
entries, extracted Contract"; data inserted and errors
tables, filled forms, into database fields; a introduced by OCR.
flags indicating structured
compliance or non- representation of a
compliance. table; a completed
digital form.
Key OCR accuracy, layout Using Tesseract or Robustness to different Automating
Consideratio analysis to understand Google Cloud Vision API document formats and complex workflows
ns document structure, for OCR, employing quality, maintaining involving multiple
handling both textual computer vision models data privacy and document types.
and visual information, to identify tables and security.
rule-based systems sections, using regular
combined with machine expressions and NLP
learning models. models for information
extraction.

9
Identify Features of Generative AI Workloads
Generative AI refers to AI that can create new content (text, code, images, audio, video).

Generative AI Type Description Example

Text Generation Generate responses, stories, or ChatGPT writing blog posts
content
Image Generation Create images from descriptions AI generating art from prompts
Code Generation Auto-generate code snippets GitHub Copilot writing Python
code
Music Generation Compose music tracks AI composing background
music
Video/Animation Generate short clips or Creating animated avatars or
Generation animations scenes

Feature Description Examples

Content Creation Generating new text, images, audio, Writing essays, synthesizing art
or videos
Personalization Tailoring content to specific needs Generating customer-specific
emails
Data Creating synthetic data for training Expanding datasets for model
Augmentation models training
Creative Producing creative outputs Generating music or designing
Applications products

Feature Description Examples

Text Generation Creating human-like text based on • Content writing assistants
prompts or context • Marketing copy generation
• Chatbots and virtual assistants
• Creative writing tools
Image Generation Creating new images from textual • Product visualization
descriptions or other images • Concept art creation
• Design ideation tools
• Visual content for marketing
Audio Synthesis Generating realistic speech, music, or • Text-to-speech systems
sound effects • Voice cloning
• Music composition
• Sound effect generation
Video Generation Creating video content from prompts or • Animation generation
existing media • Training video creation
• Special effects
• Virtual presenter systems
Code Generation Automatically generating programming • Software development assistance
code from descriptions • Automated testing
• Low-code development platforms
• API integration tools
3D Model Generation Creating three-dimensional models from • Product design
descriptions or images • Game asset creation
• Architectural visualization
• Virtual reality content
Multimodal Generation Creating content that spans multiple • Interactive storytelling
formats (text, image, audio) • Mixed media content creation
• Augmented reality experiences
• Enhanced educational materials
Style Transfer Applying the style of one piece of content • Artistic photo filters
to another • Voice style transformation
• Text style adaptation
• Design theme application
Fine-tuning & Prompt Adapting foundational models for specific • Domain-specific assistants
Engineering tasks through training or prompt design • Brand voice adaptation
• Specialized creative tools
• Custom chatbot personalities

10
Feature Description Examples Underlying Key Challenges
Techniques/Models
Input Data Large datasets of the A large corpus of text for Generative Adversarial Ensuring the
type of content to be language models, a Networks (GANs), quality, coherence,
generated (text, dataset of images for Variational and relevance of
images, audio, etc.). image generation, audio Autoencoders (VAEs), the generated
Can also take prompts recordings for music Transformer models (for content.
or conditions to guide generation, source code text, image, and audio
the generation process. for code generation. Text generation), Diffusion
prompts like "a cat Models.
wearing a hat."
Core Tasks Text generation, image Writing articles, creating Generator and Controlling the
generation, audio realistic images from Discriminator networks generation process
synthesis, video descriptions, generating (in GANs), encoder and to produce desired
generation, code music in a specific style, decoder networks (in outputs, avoiding
generation, synthetic creating short video clips, VAEs), attention mode collapse (in
data generation, 3D generating Python mechanisms (in GANs), ensuring
model generation. functions, creating Transformers). diversity in
artificial data for training generated content.
other AI models,
generating 3D models of
objects.
Output New, original content A generated news article, Generated sequences, Evaluating the
that resembles the a photorealistic image of pixel arrays, audio "quality" and
training data: text a landscape, a newly waveforms, code originality of
passages, images, composed musical piece, strings, synthetic data generated content,
audio samples, video a short animation, a points, 3D mesh data. addressing ethical
clips, code snippets, Python function for concerns related to
synthetic datasets, 3D sorting a list, a table of misuse (e.g.,
models. artificially generated deepfakes).
customer data, a 3D
model of a chair.
Key Model architecture Using large language Computational Interpretability of
Consideratio selection (GAN, VAE, models with specific resources required for generative models
ns Transformer, Diffusion), prompting strategies, training and inference, and understanding
training data quality training GANs with careful the risk of generating how they create
and quantity, loss hyperparameter tuning, biased or harmful new content.
functions that using diffusion models for content.
encourage realistic and high-quality image
diverse generation, generation.
techniques for
controlling the
generation process
(e.g., conditioning).

11
Identify guiding principles for responsible AI
Principle Description / Considerations
Fairness AI systems should treat all people fairly. Consider avoiding biases in data and algorithms that
can lead to discrimination based on gender, race, age, etc.
Reliability & AI should perform reliably and safely in all expected conditions. Include testing, validation, and
Safety fail-safes to handle unexpected behavior or inputs.
Privacy & AI systems must ensure data privacy and be secure against unauthorized access or misuse.
Security Implement data protection policies, encryption, and compliance with regulations.
Inclusiveness AI systems should be designed to empower everyone and be usable by people with diverse
backgrounds and abilities (e.g., accessibility features).
Transparency Users should understand how and why an AI system makes decisions. Use explainable AI
methods and provide documentation or model interpretability.
Accountability Organizations and developers must be accountable for the AI systems they build. Assign clear
responsibility and ensure mechanisms for auditing and redress.

Principle/Consideration Description Examples/Implications

Guiding Principles for Ethical AI development that ensures Creating AI systems that serve humanity
Responsible AI fairness, reliability, safety, privacy, while minimizing risks and biases.
inclusiveness, transparency, and
accountability.
Fairness Ensuring that AI solutions do not exhibit Removing biases in training data to avoid
bias and work equally for all users. discriminatory outcomes, such as
unequal credit approvals.
Reliability and Safety Developing AI systems that operate Testing models rigorously to ensure
consistently and protect users from harm. consistent performance and prevent
erroneous predictions.
Privacy and Security Safeguarding user data and preventing Implementing robust encryption and
unauthorized access or misuse. adhering to data privacy regulations like
GDPR.
Inclusiveness Creating AI solutions that are accessible Designing interfaces that accommodate
and beneficial to everyone, including users with disabilities or language
marginalized groups. differences.
Transparency Clearly documenting AI processes, Publishing model details and explaining
decisions, and limitations for users and how predictions are made to build trust
stakeholders. and understanding.
Accountability Holding developers and organizations Monitoring systems for unintended
responsible for AI’s impacts and consequences and establishing
outcomes. remediation processes for errors.

Principle Description Key Considerations

Fairness AI systems should treat all people • Identify and mitigate bias in training data
equitably and avoid creating or • Regular fairness assessments across different
reinforcing bias demographic groups
• Balanced representation in datasets
• Fairness-aware model development
• Continuous monitoring for emerging biases
Reliability & AI systems should perform reliably, • Robust testing across various scenarios
Safety safely, and consistently under normal • Fail-safe mechanisms and graceful degradation
and unexpected conditions • Ongoing performance monitoring
• Safety boundaries and constraints
• System resilience to adversarial attacks
Privacy & AI systems should respect user • Data minimization practices
Security privacy and maintain data security • Strong encryption and security protocols
• Privacy-preserving techniques (differential privacy,
federated learning)
• Clear data collection and usage policies
• Compliance with relevant regulations (GDPR, CCPA,
etc.)
Inclusiveness AI systems should empower • Accessible design for people with disabilities
everyone and engage people of all • Support for multiple languages and cultures
abilities • Consideration of diverse user needs

12
 • Testing with diverse user groups
• Avoiding exclusionary design patterns
Transparency AI systems should be understandable • Explainable AI methods
and explainable in appropriate • Clear documentation of model capabilities and
context limitations
• Understandable user interfaces
• Disclosure of AI involvement in interactions
• Accessible explanations of decision processes
Accountability Organizations should be accountable • Clear governance structures
for their AI systems and their impacts • Human oversight of critical decisions
• Audit trails for significant decisions
• Mechanisms for redress when systems cause harm
• Regular ethical impact assessments

Guiding Principle Description Key Considerations/Practices

Guiding Principles Broad, overarching values that -Fairness: Ensuring AI systems do not perpetuate or amplify biases.
for Responsible AI should inform the design, -Reliability & Safety: AI systems should function as intended and be
development, and deployment of safe under foreseeable conditions.
-Privacy & Security: Protecting data used and processed by AI
AI systems to ensure positive
systems.
societal impact and minimize -Inclusiveness: Designing AI systems that benefit and are accessible
harm. These principles often to all.
overlap and reinforce each other. -Transparency: Providing insights into how AI systems work and make
decisions.
-Accountability: Establishing responsibility for the outcomes of AI
systems.
Fairness Ensuring that AI systems treat all -Bias Detection: Identifying and mitigating biases in training data and
individuals and groups equitably algorithms.
and avoid creating or reinforcing -Fairness Metrics: Defining and measuring fairness using appropriate
metrics (e.g., demographic parity, equalized odds).
unfair biases.
-Algorithmic Auditing: Regularly evaluating AI systems for fairness
across different demographic groups.
-Diverse Teams: Involving individuals with diverse backgrounds in the
development process.
Reliability and Ensuring that AI systems perform -Robustness Testing: Evaluating AI systems against a wide range of
Safety consistently, accurately, and inputs, including noisy or adversarial data.
safely under various conditions, -Error Handling: Designing systems to gracefully handle errors and
failures.
including edge cases and
-Monitoring and Validation: Continuously monitoring performance
unexpected inputs. and validating the system's behavior over time.
-Explainable AI (XAI): Understanding the reasons behind AI outputs to
identify potential safety issues.
Privacy and Protecting the privacy of -Data Minimization: Collecting and retaining only the necessary data.
Security individuals and ensuring the -Anonymization and Pseudonymization: Techniques to de-identify
security of the data used and data.
-Secure Data Storage and Transmission: Implementing robust
processed by AI systems against
security measures to protect data.
unauthorized access, use, or -Compliance with Regulations: Adhering to relevant privacy laws and
disclosure. standards (e.g., GDPR, CCPA).
Inclusiveness Designing AI systems that are -Accessibility: Ensuring AI interfaces and outputs are usable by
accessible and beneficial to a people with disabilities.
wide range of individuals, -Representation in Data: Using diverse and representative training
data to avoid bias against certain groups.
including those from diverse
-User-Centered Design: Involving diverse users in the design and
backgrounds, abilities, and testing process.
demographics. -Language and Cultural Sensitivity: Considering linguistic and cultural
variations.
Transparency Providing clear and -Explainable AI (XAI) Techniques: Using methods to provide insights
understandable information about into the reasoning behind AI outputs (e.g., feature importance,
how AI systems work, make saliency maps).
-Model Documentation: Clearly documenting the AI system's
decisions, and their limitations.
architecture, training data, and limitations.
-User Communication: Communicating clearly with users about how
the AI system operates and its potential impact.
Accountability Establishing clear roles, -Clear Lines of Responsibility: Defining who is accountable at
responsibilities, and processes for different stages of the AI lifecycle.
the development, deployment, -Auditability: Enabling the tracking and review of AI system decisions
and actions.
and consequences of AI systems.
-Feedback Mechanisms: Providing channels for users to report issues
Ensuring that there are and concerns.
mechanisms in place to address -Remediation Processes: Establishing procedures for addressing
any negative impacts or negative impacts and making necessary corrections.
unintended outcomes.

13
Describe fundamental principles of machine learning on Azure
(15-20%)
Identify common machine learning techniques

 Identify regression machine learning scenarios

 Identify classification machine learning scenarios

 Identify clustering machine learning scenarios

 Identify features of deep learning techniques

 Identify features of the Transformer architecture

Describe core machine learning concepts

 Identify features and labels in a dataset for machine learning

 Describe how training and validation datasets are used in machine learning

Describe Azure Machine Learning capabilities

 Describe capabilities of automated machine learning

 Describe data and compute services for data science and machine learning

 Describe model management and deployment capabilities in Azure Machine Learning

14
Identify common machine learning techniques
Category Description Examples
Supervised Learning - Learns from labeled data (input-output - Regression: House price prediction
pairs). - Classification: Spam detection,
- Objective: Map input to known outputs. Image recognition
- Two types: Regression & Classification.
Unsupervised Learning - Learns from unlabeled data. - Clustering: Customer segmentation
- Objective: Find hidden - Dimensionality Reduction: PCA for
patterns/structures. visualization
- Two types: Clustering & Dimensionality
Reduction.
Reinforcement Learning - Learns by interacting with an - Game AI (e.g., AlphaGo)
environment. - Robotics control
- Objective: Maximize reward through - Autonomous driving
actions.
- Uses agents, states, actions, rewards.
Regression ML Scenarios - Predict continuous numerical values. - Predicting temperature
- Forecasting stock prices
Classification ML - Predict discrete categories/labels. - Fraud detection
Scenarios - Sentiment analysis
Clustering ML Scenarios - Group similar data points without labels. - Customer grouping
- Market segmentation
Deep Learning Features - Deep neural networks (CNNs, RNNs, - Image classification
etc.). - Speech recognition
- Handles unstructured data (images, text, - Natural Language Processing
audio).
- Requires large datasets and high
computational power.
Transformer Architecture - Self-attention mechanism. - BERT
Features - Processes entire input sequences in - GPT (like me!)
parallel. - Translation systems
- Used for NLP tasks.

Category Description Azure Tools/Services

Regression Predicting continuous values, such as house Azure Machine Learning Studio (supports
Scenarios prices, sales forecasts, or stock trends. regression models)
Classification Categorizing data into predefined classes, e.g., Azure ML Studio, Azure Cognitive
Scenarios spam email detection, customer segmentation, Services (e.g., Custom Vision, Form
or fraud detection. Recognizer)
Clustering Grouping data into clusters based on similarity, Azure ML Studio, Azure Synapse
Scenarios e.g., customer segmentation, document topic Analytics (for clustering analysis)
modeling.
Features of Deep Extracting hierarchical features from data, Azure Cognitive Services, Azure Machine
Learning highly suitable for image processing, speech Learning (supports deep learning
Techniques recognition, and more. frameworks like TensorFlow, PyTorch)
Features of Efficient handling of sequential data with Azure OpenAI Service, Azure Cognitive
Transformer attention mechanisms, revolutionizing tasks like Search (leverages Transformer models)
Architecture natural language processing and AI generation.

Machine Learning Scenarios

Scenario Description Examples Azure Implementation

Regression Predicts continuous • House price prediction • Azure Machine Learning regression
numerical values based • Stock price forecasting algorithms
on input features • Temperature prediction • Azure Automated ML for regression
• Sales forecasting • Azure Databricks regression models
• Customer lifetime value • Azure Synapse Analytics regression
estimation pipelines
Classificati Categorizes inputs into • Email spam detection • Azure Machine Learning classification
on discrete classes or • Sentiment analysis algorithms
labels • Disease diagnosis • Azure Cognitive Services Text
• Credit approval Analytics
• Image recognition • Azure Custom Vision Service

15
 • Azure Automated ML for classification
Clustering Groups similar data • Customer segmentation • Azure Machine Learning clustering
points together without • Anomaly detection algorithms
prior labels • Document categorization • Azure Databricks K-means clustering
• Network traffic analysis • Azure Synapse Analytics clustering
• Image segmentation • Azure Cognitive Search semantic
clustering
Deep Learning Techniques

Technique Key Features Applications Azure Implementation

Deep • Multiple layers of neural • Computer vision • Azure Machine Learning deep
Learning networks • Natural language learning frameworks
• Automatic feature extraction processing • Azure Cognitive Services
• Handles unstructured data • Speech recognition • Azure Neural Network Intelligence
well • Anomaly detection (NNI)
• Requires significant training • Time series • Azure GPU-accelerated VMs
data forecasting • Azure FPGA for model inference
• Computationally intensive
Transformer Architecture

Feature Description Benefits Azure Implementation

Attention Self-attention allows • Captures long-range • Azure OpenAI Service
Mechanism model to weigh dependencies • Azure Cognitive Services
importance of different • Processes sequences in parallel Text Analytics
• Handles variable-length inputs • Azure Bot Service
input elements
Parallel Processes entire • Faster training • Azure ML distributed
Processing sequences • Better handling of long-range training
simultaneously instead dependencies • Azure GPU clusters
of sequentially
Encoder- Encoder processes input, • Effective for sequence-to- • Azure OpenAI Service
Decoder decoder generates sequence tasks • Azure Machine Learning
Structure output • Modular architecture allows custom models
pretraining
Positional Embeds position • Preserves sequence order in • Azure Cognitive Services
Encoding information into tokens parallel processing Language Understanding
• Enables contextual
understanding
Pre-training & Models are pre-trained • Transfer learning • Azure OpenAI Service fine-
Fine-tuning on large datasets then • Reduces data requirements for tuning
fine-tuned for specific specific tasks • Azure ML custom model
• Improves performance on deployment
tasks
specialized domains

Category Scenario/Feature Azure Relevance

Regression Predicting house prices based on Azure Machine Learning provides tools for building and
features like size, location, and deploying regression models, such as linear regression,
number of rooms. decision trees, and neural networks.
Forecasting sales revenue for Azure Time Series Insights and Azure Machine Learning
the next quarter. can be used for time series forecasting using regression
techniques.
Estimating the remaining useful Azure IoT Hub and Azure Machine Learning can ingest
life of industrial equipment. sensor data and build regression models to predict
equipment failure.
Predicting stock prices based on Azure Data Factory can be used to ingest financial data,
historical data and market and Azure Machine Learning to build forecasting models.
indicators.
Classification Identifying whether an email is Azure Cognitive Services offers pre-trained models for text
spam or not spam. analytics, including spam detection. Azure Machine
Learning can also be used.
Diagnosing a disease based on Azure Health Data Services can integrate with Azure
patient symptoms and medical Machine Learning for building diagnostic classification
test results. models.
Categorizing news articles into Azure Cognitive Services provides text classification
different topics (e.g., sports, capabilities. Azure Machine Learning allows for custom
politics, technology). model development.
Detecting fraudulent credit card Azure Machine Learning offers tools and techniques for

16
 transactions. building fraud detection classification models.
Clustering Grouping customers based on Azure Synapse Analytics and Azure Machine Learning can
their purchasing behavior. be used to perform customer segmentation using
clustering algorithms.
Segmenting images into While not strictly clustering in the traditional sense, Azure
different regions or objects. Computer Vision can identify and segment objects in
images. Azure Machine Learning supports clustering on
image features.
Identifying patterns in network Azure Network Watcher and Azure Sentinel can leverage
traffic for anomaly detection. machine learning, including clustering, for security
analysis.
Grouping similar documents Azure Cognitive Search can perform semantic search and
together for topic discovery. clustering of documents. Azure Machine Learning can also
be used for custom topic modeling.
Deep Learning Automatic feature extraction Azure Machine Learning supports various deep learning
Features from raw data (e.g., images, frameworks like TensorFlow, PyTorch, and ONNX, enabling
text, audio). automatic feature learning.
Ability to learn complex, Azure GPUs and optimized compute instances in Azure
hierarchical representations of Machine Learning accelerate the training of deep neural
data. networks.
Scalability to handle large Azure Machine Learning provides distributed training
datasets and complex models. capabilities across multiple GPUs and nodes.
End-to-end learning, directly Azure Machine Learning simplifies the deployment of
mapping inputs to outputs. trained deep learning models for inference.
Transformer Self-attention mechanism, Azure Machine Learning supports Transformer-based
Architecture allowing the model to weigh the models for natural language processing tasks.
Features importance of different parts of
the input sequence.
Parallel processing of the input Azure's infrastructure enables efficient training and
sequence. inference of Transformer models.
Positional encoding to Azure Cognitive Services for Language often utilizes
understand the order of Transformer architectures under the hood.
elements in the sequence.
Encoder-decoder structure for Azure Machine Learning facilitates the fine-tuning and
sequence-to-sequence tasks. deployment of Transformer models for tasks like
translation and text generation.
Contextual understanding of Azure OpenAI Service provides access to powerful
input data. Transformer-based language models.

17
Describe core machine learning concepts
Identify features and labels in a dataset for machine learning
In machine learning, a dataset is usually structured like a table with rows and columns. Each row is a data point (or
sample), and each column is a variable (or attribute).
 Features are the input variables — the data we use to make predictions.
 Label (also called the target) is the output we want to predict.
Example:
Ag Salar Owns Will Buy
e y House Product
25 40,00 No No
0
45 85,00 Yes Yes
0
35 60,00 No Yes
0
 Features: Age, Salary, Owns House
 Label: Will Buy Product (this is what we want the model to predict)

Features and Labels in a Dataset

Concep Definition Example
t
Featur Input variables that the model uses to make House size, number of bedrooms,
es predictions. location
Labels Output variables or targets that the model House price
predicts.

Features and Labels in a Dataset

In machine learning:
 Features are the input variables or attributes of the dataset used to predict outcomes. For example, if
you're predicting house prices, features might include the size of the house, number of bedrooms, or its
location.
 Labels are the output variables or target values you’re trying to predict. In the same example, the label
could be the actual price of the house.
Think of features as the “questions” and labels as the “answers” you want your model to learn.

In machine learning, datasets are composed of two main components:

Features:
 These are the input variables or attributes used to make predictions
 Also called independent variables, predictors, or X variables
 Features represent the characteristics or properties of each data point
 Examples: age, income, temperature, pixel values in an image, words in text
Labels:
 These are the output variables that the model aims to predict
 Also called dependent variables, targets, or Y variables
 Labels represent what you're trying to predict or classify
 Examples: customer churn (yes/no), house price, disease diagnosis, image category
The relationship between features and labels is what machine learning models attempt to learn. For instance, in a
dataset predicting house prices, features might include square footage, number of bedrooms, and neighborhood,
while the label would be the house price.

18
Describe how training and validation datasets are used in machine learning

When training a machine learning model, the dataset is usually split into two (or sometimes three) parts:
✅ Training Dataset
 This is the part of the data used to "teach" the model.
 The model learns patterns and relationships between features and labels.
✅ Validation Dataset
 Used to evaluate the model’s performance during training.
 Helps to tune the model and avoid overfitting (when a model performs well on training data but poorly on
new data).
 It's not shown to the model during training, only used to check how well it's generalizing.
(Optional) Test Dataset
 Used after training and validation to give a final unbiased evaluation of the model.

Training and Validation Datasets

In machine learning, we split datasets into two (or three) parts:

 Training Dataset: This is used to train the machine learning model, teaching it how to identify patterns and
relationships between features and labels. It’s essentially where the model learns.

 Validation Dataset: This is used during the training phase to check how well the model is performing. It
helps in fine-tuning parameters and avoiding overfitting (when the model performs well on training data
but poorly on new data).

In practice, we also use a test dataset, which is separate from the training and validation datasets, to evaluate the
model's performance on unseen data.

Dataset Type Purpose How It's Used

Training Dataset Teaches the model by finding patterns between features Used to build the model.
and labels.
Validation Dataset Evaluates model performance during training and helps Used for fine-tuning while
tune parameters to prevent overfitting. training.
Test Dataset Assesses model performance on unseen data to ensure Used after training is
reliability. complete for final evaluation.

Visualization Idea
Here’s a simple conceptual diagram to illustrate how data is split and used in machine learning:
Dataset Split:
1. Training Dataset → The model learns patterns.
2. Validation Dataset → The model is checked and refined.
3. Test Dataset → The model is evaluated on unseen data.

Here's a concise table that captures the key information about features, labels, and datasets in machine learning:

Concept Definition Examples Purpose

Features Input variables or attributes Age, income, temperature, Provide information for the model
used for prediction pixel values, text words to learn patterns
Labels Output variables the model Customer churn (yes/no), The target variable the model is
aims to predict house price, disease trained to predict
diagnosis
Training 70-80% of data used to train Model learns patterns by
Dataset the model adjusting parameters based on
this data
Validation 10-15% of data used during Tune hyperparameters, prevent
Dataset training overfitting, select best model
configuration
Test Dataset 10-15% of data used after Evaluate final model
training performance on completely
unseen data

19
Describe Azure Machine Learning capabilities

Capabilities of Automated Machine Learning (AutoML)

Automated machine learning streamlines the process of developing machine learning models
by automating key steps like data preprocessing, feature engineering, model selection,
hyperparameter tuning, and deployment. Key capabilities include:
 Ease of Use: AutoML makes it accessible to non-experts by simplifying complex
workflows.
 Efficiency: Accelerates the experimentation phase by testing multiple models and
configurations automatically.
 Transparency: Provides insights into the models, including metrics and interpretability,
for informed decision-making.
 Customization: Allows users to adjust certain aspects like algorithms, configurations,
and validation processes while automating others.

Feature Description
Algorithm Selection Automatically tests and selects the best algorithm for the task.
Feature Engineering Creates and transforms input features to boost model performance.
Hyperparameter Tuning Optimizes algorithm settings (like learning rate, depth, etc.).
Model Evaluation Compares models using metrics (accuracy, precision, recall, etc.).
Task Types Supports classification, regression, and time-series forecasting.
Ease of Use No need for deep data science knowledge.
Efficiency Faster and more efficient than manual model training.

Capability Description
Algorithm selection Automatically tests multiple algorithms to find the best for your
data
Feature engineering Creates and selects the most relevant features from raw data
Hyperparameter tuning Optimizes model parameters to improve performance
Cross-validation Ensures models generalize well to unseen data
Model evaluation Compares metrics across different models
Model explanation Provides interpretability features to understand model decisions
Time series forecasting Specialized capabilities for time-dependent data
NLP tasks Support for text classification and other natural language
processing tasks

Capability/Service Description
Automated Data Automatically handles missing values, encodes categorical features, scales
Preprocessing & Feature numerical features, performs text featurization, and selects/engineers
Engineering relevant features.
Algorithm Selection Automatically tries a range of suitable machine learning algorithms for the
given task.
Hyperparameter Tuning Automatically searches for optimal hyperparameter settings for selected
algorithms.
Model Training & Trains multiple models in parallel and automatically evaluates their
Evaluation performance using appropriate metrics.
Best Model Selection Identifies and recommends the best-performing model based on the chosen
metric.
Explainability Provides insights into feature importance and model interpretability.
Integration with Azure ML Seamlessly integrates with Azure Machine Learning workspace and other
Azure services.
Support for Various ML Supports classification, regression, time series forecasting, computer vision,
Tasks and natural language processing tasks.
Scalability & Efficiency Leverages Azure's compute infrastructure for parallel processing, reducing
training time.
MLOps Integration Facilitates the operationalization of AutoML models through registration,

20
deployment, and monitoring options.

21
Data and Compute Services for Data Science and Machine Learning

Azure provides robust services for managing data and compute resources:
1. Data Services:
o Azure Blob Storage: Stores large-scale, unstructured data for training models.
o Azure Data Lake: Handles big data analytics and provides a scalable, secure platform.
o Azure SQL Database: Manages structured data for ML training and predictions.
2. Compute Services:
o Azure Machine Learning Compute Clusters: Scalable clusters for distributed training and inference
tasks.
o Azure Kubernetes Service (AKS): Manages containerized workloads for training and deploying ML
models.
o Azure Virtual Machines: Offers flexible computing environments tailored to your workload.
Service Type Azure Service Description
Data Storage Azure Blob Storage Stores large datasets like CSVs, images, etc.
Azure Data Lake Optimized for analytics on big data.
Azure SQL Database Stores structured, queryable data.
Compute Azure ML Compute Instance Pre-configured VM for development and testing.
Azure ML Compute Cluster Auto-scalable compute power for training.
Azure Kubernetes Service Used for scalable deployment and hosting models.
(AKS)
Inference Clusters Specialized for running deployed models.

Category Service Description

Data Services Azure Storage Blob, File, and Data Lake storage for various data types
Azure Databricks Apache Spark-based analytics platform
Azure Synapse Analytics Data integration, warehousing, and big data analytics
Azure Data Factory Data integration service for ETL/ELT workflows
Azure SQL Database Managed relational database service
Cosmos DB Globally distributed multi-model database for NoSQL
needs
Compute Azure ML Compute Managed compute targets for ML workflows
Services
Azure Kubernetes For containerized model deployment
Service
Azure Container Serverless container hosting
Instances
GPU and CPU VMs Specialized virtual machines for ML workloads
Low-priority VMs Cost-effective option for non-critical workloads
Compute Instances Fully managed workstations for data scientists

Data Services for ML & Data Science

Azure Blob Storage Scalable and cost-effective object storage for unstructured data.
Azure Data Lake Storage (ADLS) Highly scalable data lake optimized for big data analytics.
Gen2
Azure SQL Database Fully managed relational database for structured data.
Azure Synapse Analytics Limitless analytics service for data integration, warehousing, and big
data analytics.
Azure Data Factory Cloud-based ETL and data integration service for orchestrating data
pipelines.
Azure Databricks Fast, easy-to-use Apache Spark-based analytics service.
Azure Cosmos DB Globally distributed, multi-model database for scalable applications.
Azure Purview Unified data governance service for understanding and managing data.
Azure Machine Learning Data Assets Versioned and managed references to data sources.
Azure Feature Store (Preview) Centralized repository for managing and reusing machine learning
features.
Compute Services for ML & Data Science
Azure Machine Learning Compute Managed, scalable CPU and GPU compute clusters optimized for ML
workloads.

22
 Azure Virtual Machines (VMs) Customizable infrastructure as a service, including GPU-enabled
options.
Azure Databricks Compute Managed Spark clusters with CPU and GPU options for data processing
and ML.
Azure Container Instances (ACI) Fast and simple way to run Docker containers without managing
infrastructure.
Azure Kubernetes Service (AKS) Fully managed Kubernetes service for deploying and scaling
containerized applications.
Azure Functions Serverless compute service for on-demand code execution.
Azure Data Science Virtual Machines Pre-configured VMs with popular data science and ML tools.
(DSVM)
Model Management and Deployment in Azure Machine Learning
Model Management and Deployment in Azure Machine Learning
Azure Machine Learning provides comprehensive tools for managing and deploying models:
 Model Management:
o Version Control: Tracks different versions of models for comparison and auditing.
o Registry: Stores and organizes models in a centralized repository.
o Monitoring: Observes model performance and detects drifts over time.
 Deployment:
o Endpoint Creation: Deploys models as REST APIs for integration with applications.
o Scalability: Ensures models can handle varying loads through scaling mechanisms.
o Deployment Options: Offers real-time (online) and batch (offline) inference options.
o Azure Kubernetes Service: Supports deployment on containers for efficient scaling and
orchestration.

Capability Description
Model Registry Central store to manage and version models.
Versioning Tracks different versions of a model.
Metadata Tracking Records training data, parameters, and metrics.
Real-time Deployment REST API deployment using AKS or managed endpoints.
Batch Deployment Run predictions on large data sets on a schedule.
MLOps Integration Enables CI/CD pipelines for training and deployment.
Monitoring Tracks performance and detects data drift.

Category Capability Description

Model Management Model registry Central repository to store, version, and track
models
Model versioning Maintain multiple versions of models with metadata
Model lineage Track data, code, and parameters used to create
models
Model tagging Apply custom tags for organization and search
Model approval Governance processes for model promotion
workflows
Deployment Batch inference Process data in batches for non-real-time scenarios
Capabilities
Real-time endpoints Deploy as REST API endpoints for low-latency
inference
Azure Kubernetes Scalable production deployments with high
Service availability
Container Instances Quick, simple deployments for testing or low-traffic
scenarios
Edge deployment Deploy to IoT Edge devices for offline inferencing
CI/CD integration MLOps pipelines with GitHub Actions or Azure
DevOps
A/B testing Compare model performance in production
Monitoring Track model performance, data drift, and endpoint
metrics

Model Management in Azure ML

Model Registry Centralized repository to store and version trained models.

23
 Model Versioning Automatic tracking of different versions of a registered model.
Metadata Tracking Ability to add tags and properties to models for organization and information.
Lineage Tracking Tracks the origin of models, including experiments and data.
Model Profiling (Preview) Provides insights into the input data expected by a model.
Model Deployment in Azure ML
Managed Online Endpoints (ACI) Fully managed real-time inference for testing and low-scale deployments using
Azure Container Instances.
Managed Online Endpoints (AKS) Fully managed real-time inference for high-scale, production deployments using
Azure Kubernetes Service.
Managed Batch Endpoints Enables batch inference on large volumes of data.
Deployment to Azure Compute (VMs, Flexible options to deploy models to various Azure compute resources.
Functions, App Service, IoT Edge)
Packaging and Containerization Simplifies the creation of Docker containers for consistent deployments.
Environment Management Allows defining and managing software dependencies for model execution.
Integration with MLOps Pipelines Enables automated deployment through Azure Machine Learning Pipelines and
Azure DevOps.
Model Monitoring Tools to monitor model performance, detect data drift, and identify issues in
deployed models.
Traffic Management & Blue/Green Allows controlled rollout of new model versions for online endpoints.
Deployments

Describe features of generative AI workloads on Azure (20–25%)

Features of Generative AI Models

Feature Description
Content Creation Generates text, images, audio, video, or code.
Pretrained & Fine-tunable Trained on large datasets; can be customized for specific tasks.
Prompt-driven Uses natural language input (prompts) to generate content.
Multimodal Capabilities Can work across different data types like text-to-image.
Self-learning/ Predicts next element in a sequence (e.g., next word or pixel).
Autoregressive

Feature Description
Content Generation Generates diverse outputs such as text, images, music, or videos based on
input prompts.
Context Understanding Analyzes input context to produce coherent and relevant responses.
Fine-Tuning Capability Customizes models for specific use cases or domains through additional
training.
Creativity & Variability Produces unique and varied outputs, often blending imagination with real-
world data.
Multi-Modal Integration Combines data types (e.g., text and images) for richer output capabilities.
Few-Shot & Zero-Shot Responds effectively with limited examples or none at all during input
Learning prompts.

Feature/Scenario/ Description
Consideration
Learning Data Distributions Models learn the underlying probability distribution of the training data.
Generating Novel Data Creates new, synthetic data instances that resemble the training data but
were not explicitly present.
Sampling from Latent Many models learn a lower-dimensional representation and generate data by
Spaces sampling from it.
Diverse Output Capabilities Can generate various data types, including images, text, audio, video, 3D
models, and synthetic data.
Conditioning on Input Can generate outputs based on specific prompts or conditions.
Scalability and Complexity Modern models can be very large with billions of parameters, enabling high
realism but requiring significant compute.
Continuous Improvement Performance can improve with more and better training data.
Versatile Architectures Utilizes various neural network architectures like GANs, VAEs, Transformers,
Diffusion Models, and Autoregressive Models.

24
Common Scenarios for Generative AI

These models are widely used across industries for a variety of purposes:
Examples:
1. Text Generation – Chatbots, content writing, summarization (e.g., Copilot, ChatGPT).
2. Image Generation – Design, marketing, art creation (e.g., DALL·E, Midjourney).
3. Code Generation – Assisting developers with code suggestions or generation (e.g., GitHub Copilot).
4. Translation and Language Tasks – Translating documents, answering questions, generating language-
specific content.
5. Personalization – Recommender systems, custom marketing messages.
6. Data Augmentation – Creating synthetic data for model training or testing.

Scenario Application
Content Writing articles, designing graphics, composing music, or generating
Creation animations.
Customer Answering queries, creating automated chat responses, or resolving issues
Support via bots.
Product Design Assisting in prototyping and brainstorming creative solutions for new
products.
Language Translating text across multiple languages while maintaining contextual
Translation accuracy.
Education & Developing interactive learning materials, summarizing topics, or providing
Training explanations.
Healthcare Supporting diagnostics, generating medical reports, or simplifying patient
communication.

Content Creation Writing articles, generating artwork, composing music, creating animated
(Text, Image, Audio, Video) clips.
Data Augmentation and Creating artificial datasets for training other AI models or for privacy-
Synthesis preserving data sharing.
Drug Discovery and Material Generating novel molecular structures and designing new materials.
Science
Product Design and Generating new product designs and creating 3D models.
Prototyping
Fashion and Entertainment Generating new fashion designs and creating virtual environments/characters.
Personalization Generating personalized content recommendations and customized user
experiences.
Education and Research Generating explanations and examples, assisting in scientific discovery.
Code Generation Assisting developers by generating code snippets or entire functions.

25
26
Responsible AI Considerations for Generative AI

Generative AI introduces risks that need to be addressed to ensure ethical use.

Key considerations:
1. Fairness – Avoiding bias in outputs that reflect stereotypes or discrimination.
2. Transparency – Clearly communicating that content was AI-generated.
3. Privacy – Ensuring personal or sensitive data is not leaked in generated outputs.
4. Security – Preventing misuse (e.g., deepfakes, phishing content).
5. Accountability – Keeping humans in the loop and responsible for decisions.
6. Content Safety – Filtering harmful, offensive, or misleading content.
7. Intellectual Property (IP) – Avoiding copyright issues with generated or training content.

Consideration Description
Fairness Ensuring unbiased outputs across diverse users and contexts.
Privacy Safeguarding user data and maintaining confidentiality during input and
generation.
Reliability & Guaranteeing outputs are accurate, safe, and free from harmful content.
Safety
Transparency Providing clear explanations of model behavior and limitations.
Accountability Defining responsibility for misuse or harmful outcomes from generated
outputs.
Inclusiveness Designing generative AI to benefit users with varying needs, preferences,
and accessibility.

Bias and Fairness Generative models can inherit and amplify biases, leading to unfair outputs. Requires
diverse data and bias mitigation techniques.
Transparency and Understanding the generation process is challenging; efforts are needed to improve
Explainability interpretability.
Accountability Determining responsibility for generated content is complex; clear guidelines are
needed.
Privacy and Security Risk of leaking sensitive data from training sets and potential for malicious use (e.g.,
deepfakes). Requires privacy-preserving techniques and security measures.
Safety and Ensuring models do not generate harmful, unsafe, or unreliable content. Requires
Robustness safeguards and robustness against attacks.
Human Oversight Maintaining human involvement in development and deployment to align with values
and Control and prevent unintended consequences.
Societal Impact Considering broader societal implications, including impact on jobs and information
spread.
Intellectual Property Complex issues surrounding the ownership and rights of AI-generated content.
Misinformation and Potential for generating realistic fake content for malicious purposes; requires detection
Deepfakes and mitigation strategies.

27
Identify generative AI services and capabilities in Microsoft Azure

Feature/ Azure AI Foundry Azure OpenAI Service Azure AI Foundry Model

Capability Catalog
Purpose Enterprise-grade AI platform Access OpenAI's powerful Central hub for discovering
for model building and language models and deploying AI models
deployment
Key Focus Generative AI applications, Generative AI, image Model discovery,
collaboration, scalability creation, content comparison, and
summarization customization
Model Combines generative AI with Offers models like GPT-4, Includes models from
Integration enterprise workflows GPT-3.5, DALL-E OpenAI, Hugging Face,
NVIDIA, etc.
Customization End-to-end AI project Fine-tuning OpenAI models Fine-tuning models and
Options customization for specific tasks serverless API deployment
Deployment Test and deploy enterprise- Via REST APIs, SDKs, and Manage compute or use
grade models Azure services serverless APIs
Responsible AI Embedded in the Includes content filtering Provides detailed model
Practices development lifecycle and safeguards cards and usage guidelines
Scalability Scales from proof-of-concept Supports scalable AI Suitable for enterprise-scale
to production implementations deployments
Search/ N/A N/A Task-centric search and
Experiment Tools experimentation in
playground

Feature Azure AI Foundry Azure OpenAI Service Azure AI Foundry Model

Catalog
Primary End-to-end platform for developing, Provides access to Central hub for discovering,
Purpose deploying, and managing AI advanced language and evaluating, and deploying AI
applications, especially generative multimodal AI models from models from various
AI. OpenAI. providers.
Core Unified development hub, project Access to models via APIs, Model discovery, search,
Functionality organization, GenAIOps tools, fine-tuning, content performance benchmarks,
responsible AI safeguards, app filtering, enterprise security deployment options
management, model exploration, features. (managed compute,
deployment. serverless API).
Model Scope Integrates various models, Primarily focuses on Offers a broad selection of
including those from Azure OpenAI, OpenAI's proprietary models from Microsoft,
open-source, and other providers. models (GPT, DALL-E, etc.). OpenAI, open-source
(Hugging Face), and other
partners.
Developmen Application development lifecycle, Utilizing models for specific Selecting and deploying the
t Focus prompt engineering, RAG, agent AI tasks within applications. appropriate model for a
workflows. given task.
Deployment Cloud and edge deployment Integrating models into Managed compute (VMs) and
Focus (Container Apps, AKS), managed applications via APIs. serverless API (pay-per-
endpoints. token) deployments.
Responsible Built-in configurable evaluations, Automated content filtering Integrates with Azure AI
AI safety filters, and security controls. based on a content policy. Content Safety for deployed
models.
Integration Deeply integrated with Azure Integrates well with other Integrates with Azure AI
OpenAI, Azure AI Services, Azure Azure services. Foundry for deployment and
ML, Azure AI Search, etc. management.
User Azure AI Studio interface. Azure portal for resource Azure AI Studio interface.
Interface management, Azure AI
Studio for model
exploration/fine-tuning.
Key Benefit Streamlines and simplifies the Provides access to state-of- Simplifies the process of
entire generative AI application the-art AI models with finding, evaluating, and
lifecycle. enterprise-grade security deploying the best AI model
and scalability. for your needs.

28
Analogy The AI factory where you build and The source of the powerful The marketplace or catalog
run your AI products. AI "engines" you can use. where you shop for AI
"engines."

29
Describe features of computer vision workloads on Azure (15–
20%)

30