Skill (02): (QRP) Describe fundamental principles of ML on Azure

[MohamedRadwan-DevOps]

Skill (02): (QRP) Describe fundamental principles of ML on Azure (15–20%)

Document Type: QRP (Quick Revision Pack)
Scope: This document provides a high signal, minimal noise revision pack optimized for last-mile review. It focuses on crisp bullet points, micro examples, and checklists that cover the most important concepts and exam-relevant angles.

Identify common machine learning techniques

Common ML techniques/What ML is (exam level)

• Machine learning = models learn patterns from data to make predictions/decisions without being explicitly programmed with rules.
• Big split:
  • Supervised learning: learn from labeled examples (regression, classification).
  • Unsupervised learning: discover structure in unlabeled data (clustering).

Common ML techniques (you should memorize)/Regression vs classification vs clustering (comparison)

• Regression → predict a number (continuous value).
  • Keywords: “predict price”, “estimate demand”, “forecast value”, “how much/how many”.
• Classification → predict a category/label (class).
  • Keywords: “spam/not spam”, “fraud/not fraud”, “churn yes/no”, “which category”.
• Clustering → group similar items with no labels (discover segments).
  • Keywords: “segment customers”, “group similar”, “discover clusters”, “no labels available”.

Common ML techniques/Quick check

• Trap: If the output is a number but you see “classes,” it’s probably classification; if the output is a label but you see “score/probability,” it’s still classification (probability of a class).

Common ML techniques/Micro example

• “We want to segment customers into 5 groups based on buying behavior; no target column.” → Clustering.

Source:
Study guide for Exam AI-900: Microsoft Azure AI Fundamentals (Microsoft Learn)
Introduction to machine learning concepts (Microsoft Learn)
Machine learning tasks in ML.NET (Microsoft Learn)
Train and evaluate clustering models (Microsoft Learn)
Machine Learning Algorithm Cheat Sheet (Azure ML designer) (Microsoft Learn)

Identify regression machine learning scenarios

Regression/What it is

• Regression predicts a continuous numeric value (one number per row).
• Typical targets: price, revenue, temperature, time-to-failure, demand.

Regression/What to spot in questions

• Output is a number (not a class): “£”, “$”, “units”, “minutes”, “kWh”, “probability as a number” only if it’s explicitly the predicted value.
• Often includes “forecast” wording (note: forecasting is still regression-style output).

Regression/Quick check

• Trap: “0/1 output” can look numeric, but it’s usually binary classification (churn yes/no). Look for “class/label/category” language.

Regression/Micro example

• “Predict house price from size, location, bedrooms.” → regression.

Source:
Train and evaluate regression models (Microsoft Learn)
Introduction to machine learning concepts (Microsoft Learn)
Machine learning tasks in ML.NET (Microsoft Learn)
Machine Learning Algorithm Cheat Sheet (Azure ML designer) (Microsoft Learn)

Identify classification machine learning scenarios

Classification/What it is

• Classification predicts a category (class/label). Can be:
  • Binary (2 classes) or multiclass (>2).

Classification/What to spot in questions

• Keywords: “approve/deny”, “fraud/not”, “spam/not”, “diagnosis type”, “which category”, “route ticket to team”.
• Often output includes probabilities per class (“confidence”), but the task is still classification.

Classification/Quick check

• Trap: “predict a risk score” can be either:
  • classification (risk level: low/med/high), or
  • regression (risk score: 0–100).
    Use the stated target type (label vs number).

Classification/Micro example

• “Detect whether a transaction is fraudulent.” → binary classification.

Source:
Train and evaluate classification models (Microsoft Learn)
Introduction to machine learning concepts (Microsoft Learn)
Machine learning tasks in ML.NET (Microsoft Learn)
Machine Learning Algorithm Cheat Sheet (Azure ML designer) (Microsoft Learn)

Identify clustering machine learning scenarios

Clustering/What it is

• Clustering groups items by similarity when you don’t have labels (unsupervised).

Clustering/What to spot in questions

• Keywords: “segment”, “group similar”, “discover patterns”, “no target column”, “unknown categories”.
• Output: cluster IDs like “Cluster 1/2/3” (not meaningful labels unless you interpret them).

Clustering/Quick check

• Trap: If the question says “we already know the categories” → that’s classification, not clustering.

Clustering/Micro example

• “Group stores into similar performance profiles (sales mix, footfall, region) without predefined labels.” → clustering.

Source:
Train and evaluate clustering models (Microsoft Learn)
Introduction to machine learning concepts (Microsoft Learn)
Machine learning tasks in ML.NET (Microsoft Learn)
Machine Learning Algorithm Cheat Sheet (Azure ML designer) (Microsoft Learn)

Identify features of deep learning techniques

Deep learning/What it is

• Deep learning is ML using neural networks with multiple layers to learn complex patterns (often excels on images, audio, and text).

Deep learning/What to spot in questions

• Keywords: “neural network”, “CNN”, “transfer learning”, “GPU”, “large dataset”, “images/text/audio at scale”.
• Typical strengths: representation learning (less manual feature engineering than classic ML for unstructured data).

Deep learning/Quick check

• Trap: Deep learning isn’t automatically “better” — it often needs more data/compute and careful training; classic ML can win on small tabular datasets.

Deep learning/Micro example

• “Classify product images using a pretrained CNN and fine-tune on 2k labeled images.” → deep learning + transfer learning.

Source:
Deep Learning vs. Machine Learning (Azure ML) (Microsoft Learn)
Train and evaluate deep learning models (Microsoft Learn)
Introduction to machine learning concepts (Microsoft Learn)
Understand compute targets (GPU/CPU options) (Microsoft Learn)

Identify features of the Transformer architecture

Transformer/What it is

• Transformer architecture uses self-attention to model relationships between tokens, enabling strong NLP performance and powering many LLMs.

Transformer/What to spot in questions

• Keywords: “attention”, “self-attention”, “encoder/decoder”, “context”, “LLM”, “foundation model”.
• High-level properties:
  • Handles long-range dependencies via attention (vs only sequential recurrence).
  • Efficient parallel processing during training compared with purely sequential architectures.

Transformer/Quick check

• Trap: “Transformer = LLM” is too narrow. LLMs commonly use transformers, but transformers are a broader architecture used in many NLP tasks.

Transformer/Micro example

• “Summarize and answer questions over text using a foundation model in Azure ML.” → transformer-based model usage pattern.

Source:
Explore foundation models in the model catalog (Transformer + LLMs) (Microsoft Learn)
Deep Learning vs. Machine Learning (transformers + attention) (Microsoft Learn)
Study guide for Exam AI-900 (Microsoft Learn)
Train and evaluate deep learning models (Microsoft Learn)

Describe core machine learning concepts

Core ML concepts/What to remember

• ML flow (exam level): data → features/labels → train → validate → deploy → monitor.
• Two common failure modes:
  • Underfitting: model too simple (poor on train + validation).
  • Overfitting: model memorizes (good on train, worse on validation).

Core ML concepts/Quick check

• Trap: “high training accuracy” is meaningless without validation/test performance.

Core ML concepts/Micro example

• “Model performs great in dev but degrades in production due to changing customer behavior.” → conceptually: drift/monitoring needed.

Source:
Introduction to machine learning concepts (Microsoft Learn)
Build & train models (Azure ML) (Microsoft Learn)
MLOps model management with Azure Machine Learning (Microsoft Learn)
Train and evaluate a model (ML.NET) — split data concepts (Microsoft Learn)

Identify features and labels in a dataset for machine learning

Features vs labels/Definitions

• Features = input variables used to make a prediction (columns you feed into the model).
• Label = the target/output you want to predict (the “answer” column in supervised learning).

Features vs labels/What to spot in questions

• If the question says “predict X”, then X is the label (target).
• Everything used to predict X are features (unless explicitly excluded).

Features vs labels/Quick check

• Trap: Including the label as an input feature (“data leakage”) inflates results and fails in production.

Features vs labels/Micro example

• Predict churn (label) from tenure, plan type, complaints count (features).

Source:
Introduction to machine learning concepts (Microsoft Learn)
Train and evaluate a model (ML.NET) — data prep/splitting (Microsoft Learn)
Machine learning tasks in ML.NET (Microsoft Learn)
Machine Learning Algorithm Cheat Sheet (Azure ML designer) (Microsoft Learn)

Describe how training and validation datasets are used in machine learning

Training vs validation/Definitions (exam level)

• Training dataset: used to fit/learn model parameters.
• Validation dataset: used during development to tune/compare models and estimate generalization (helps detect overfitting).

Training vs validation/What to spot in questions

• “holdout set”, “evaluate on unseen data”, “tune hyperparameters” → validation usage.

Training vs validation/Quick check

• Trap: Repeatedly tuning using the same validation set can “leak” knowledge; cross-validation is a common mitigation conceptually.

Training vs validation/Micro example

• “Use 80% of labeled data to train, 20% to validate and select the best model.”

Source:
Introduction to machine learning concepts (Microsoft Learn)
Configure training/validation/test splits in AutoML (Microsoft Learn)
Train and evaluate a model (ML.NET) — split data (Microsoft Learn)
Build & train models (Azure ML) (Microsoft Learn)

Describe Azure Machine Learning capabilities

Azure ML/What it is (exam level)

• Azure Machine Learning provides a managed platform to prepare data, train models, track runs, register/version models, and deploy to endpoints.

Azure ML/What to spot in questions

• Keywords: “workspace”, “experiments/runs”, “model registry”, “endpoints”, “MLOps”.

Azure ML/Quick check

• Trap: Confusing “Azure Machine Learning” (platform) with “Azure AI services” (prebuilt APIs). AI-900 tests both.

Azure ML/Micro example

• “Train a model weekly, register best model, deploy to an endpoint, monitor performance.”

Source:
Build & train models (Azure ML) (Microsoft Learn)
MLOps model management with Azure Machine Learning (Microsoft Learn)
Register and work with models (Azure ML) (Microsoft Learn)
Endpoints for inference (Azure ML) (Microsoft Learn)
Study guide for Exam AI-900 (Microsoft Learn)

Describe capabilities of automated machine learning

AutoML/What it does

• AutoML trains multiple models and tunes them automatically to optimize a metric you choose (you define data, target/label, task type, and constraints).
• Common supported task types include classification, regression, forecasting (and other areas depending on setup).

AutoML/What to spot in questions

• Keywords: “no-code/low-code”, “automatically select algorithm”, “hyperparameter tuning”, “best model”.

AutoML/Quick check

• Trap: AutoML doesn’t replace validation—your job is still to choose correct label/metric and evaluate results.

AutoML/Micro example

• “Upload tabular dataset, set label ‘Churn’, run AutoML to find best classification model.”

Source:
What is automated ML? (Azure ML) (Microsoft Learn)
Set up Automated ML for tabular data in the studio (Microsoft Learn)
Evaluate AutoML experiment results (Microsoft Learn)
Tutorial: Train a classification model with no-code AutoML (Microsoft Learn)
Study guide for Exam AI-900 (Microsoft Learn)

Describe data & compute services for data science and machine learning

Data (Azure ML)/What to remember

• Azure ML uses data assets (versioned references to data) and datastores (connections to storage) to access data consistently.

Compute (Azure ML)/What to remember

• Core compute options include:
  • Compute instance (interactive dev)
  • Compute cluster (scalable training)
  • Serverless compute (managed job execution where supported)

Data & compute/Quick check

• Trap: Picking “compute instance” for large-scale training when you need an autoscaling cluster (or serverless).

Data & compute/Micro example

• “Data in Blob Storage via datastore; train on GPU compute cluster; deploy later to managed endpoint.”

Source:
Data concepts in Azure Machine Learning (data assets) (Microsoft Learn)
Create data assets (how-to) (Microsoft Learn)
Use datastores (connect storage) (Microsoft Learn)
Understand compute targets (Microsoft Learn)
Create compute clusters (Microsoft Learn)

Describe model management and deployment capabilities in Azure Machine Learning

Model management/What to remember

• Model management includes registering/versioning models, tracking metadata, and promoting models across environments (dev → test → prod).

Deployment/What to remember

• Deployment patterns:
  • Online endpoints for real-time inference.
  • Batch endpoints for asynchronous, large-scale scoring.
• MLOps capabilities include repeatable pipelines, packaging, and deployment workflows.

Model management & deployment/Quick check

• Trap: confusing “register model” with “deploy model” (registration = catalog/version; deployment = serving/scoring).

Model management & deployment/Micro example

• “Register model v3 → deploy to managed online endpoint → shift traffic gradually to new deployment.”

Source:
Register and work with models (Azure ML) (Microsoft Learn)
Machine Learning registries (promote models across environments) (Microsoft Learn)
MLOps model management with Azure Machine Learning (Microsoft Learn)
Online endpoints for real-time inference (Microsoft Learn)
Batch endpoints (Microsoft Learn)