This domain is 20% of the scored exam. Questions are short and vocabulary-heavy: define a term, match a business problem to an ML technique, match a task to the managed AWS AI service that does it, or put an ML pipeline stage next to the SageMaker feature that supports it. You never write code or tune a model, but you must recognise every term.
Core terms
| Term | Meaning |
|---|---|
| Artificial intelligence (AI) | Any technique that lets a machine do something that normally needs human intelligence. The widest circle |
| Machine learning (ML) | A subset of AI where a model learns patterns from data instead of following hand-written rules |
| Deep learning | A subset of ML that uses neural networks with many layers. Good with images, audio and text, and hungry for data and compute |
| Neural network | Layers of connected nodes (neurons) whose weights are adjusted during training |
| Generative AI | Deep learning models that create new content (text, images, audio, code). Covered in the next domains |
| Algorithm | The procedure that learns from data (linear regression, XGBoost, k-means) |
| Model | The artefact the algorithm produces: learned parameters you can use to make predictions |
| Training | Fitting the model to data. Inference is using the trained model on new data |
| Computer vision | Getting meaning from images and video |
| NLP | Getting meaning from, or producing, human language |
| LLM | A very large language model trained on huge text corpora; predicts the next token |
| Features / label | Input columns / the answer you want to predict |
| Parameters | Values the model learns (weights). Hyperparameters are settings you choose before training (learning rate, tree depth, epochs) |
Nesting to remember: AI ⊃ ML ⊃ deep learning ⊃ generative AI / LLMs.
Fit, bias and variance
- Underfitting: the model is too simple; it does badly on training and test data. High bias. Fix with a more complex model, better features, more training.
- Overfitting: the model memorised the training data; great on training, poor on new data. High variance. Fix with more (and more varied) data, regularisation, a simpler model, early stopping.
- Good fit: balanced. This is the bias-variance tradeoff.
- “Bias” has a second meaning in responsible AI: systematic unfairness toward a group. Read the context to see which one is meant.
Types of data
| Dimension | Types |
|---|---|
| Labels | Labeled (each example has the answer) and unlabeled (no answer) |
| Structure | Structured (rows and columns, database tables), semi-structured (JSON, XML) and unstructured (free text, images, audio, video) |
| Shape | Tabular, time-series (ordered by time: sales per day, sensor readings), image, text, audio |
Most business data is unstructured. Tabular and time-series data suit classic ML; images, text and audio suit deep learning.
Learning styles
| Style | Data | Idea | Examples |
|---|---|---|---|
| Supervised | Labeled | Learn the mapping from inputs to known outputs | Spam detection, house price, churn prediction |
| Unsupervised | Unlabeled | Find hidden structure | Customer segmentation (clustering), anomaly detection, dimensionality reduction |
| Reinforcement | No dataset; an environment | An agent takes actions and gets rewards or penalties, learning a policy that maximises reward | Robotics, game play, dynamic pricing. RLHF uses it to tune LLMs |
| Self-supervised | Unlabeled; labels come from the data itself | Predict a masked word or the next token | How foundation models are pre-trained |
| Semi-supervised | A little labeled, a lot unlabeled | Use both | Labeling is expensive |
Techniques by problem
| Problem | Technique | Output |
|---|---|---|
| Predict a number | Regression | Continuous value (price, temperature, demand) |
| Pick a category | Classification | Class label. Binary (fraud / not fraud) or multi-class |
| Group similar items, no labels | Clustering | Groups |
| Find unusual items | Anomaly detection | Outlier flag |
| Reduce the number of features | Dimensionality reduction (PCA) | Compressed data |
| Suggest items | Recommendation | Ranked items |
| Predict future values over time | Forecasting | Time-series values |
A frequent trap: “predict whether” means classification; “predict how much” means regression; “group customers” with no labels means clustering.
Inferencing types
| Type | When to use | Notes |
|---|---|---|
| Real-time | Low-latency answers for one request at a time (chatbot, fraud check at checkout) | Always-on endpoint; you pay while it runs |
| Batch | Score a large dataset on a schedule; no one is waiting | Cheaper, higher throughput; results land in S3 |
| Asynchronous | Large payloads or long processing, results can come later | SageMaker async endpoints queue requests |
| Serverless | Spiky or infrequent traffic | Scales to zero; tolerates cold starts |
Practical use cases for AI
AI/ML adds value when it assists human decisions, automates repetitive work, scales a task beyond what people can do, or finds patterns in data too large or complex for rules.
AI/ML is the wrong tool when:
- You need a specific, deterministic outcome every time (a tax calculation, a rule-based eligibility check). ML predicts; it doesn’t guarantee.
- A simple rule or query solves the problem.
- There is little or no good data.
- The cost-benefit doesn’t work: building, running and maintaining the model costs more than the value.
- Errors are unacceptable with no human review.
Real-world applications the exam names: computer vision, NLP, speech recognition, recommendation systems, fraud detection, forecasting.
AWS managed AI services
These are pre-trained, API-based services: no ML expertise needed. The exam asks you to match the task to the service.
| Service | What it does |
|---|---|
| Amazon Rekognition | Image and video analysis: objects, faces, face comparison, celebrity, text in images, content moderation |
| Amazon Textract | Extracts text, forms, tables and key-value pairs from scanned documents (OCR and more) |
| Amazon Comprehend | NLP on text: sentiment, entities, key phrases, language, topics, PII detection |
| Amazon Transcribe | Speech to text |
| Amazon Polly | Text to speech |
| Amazon Translate | Language translation |
| Amazon Lex | Conversational chatbots and voice bots (same technology as Alexa); intents, slots, utterances |
| Amazon Kendra | Intelligent enterprise search over documents with natural-language questions |
| Amazon Personalize | Recommendations (and personalised ranking) from your user and item data |
| Time-series forecasting | Amazon Forecast no longer accepts new customers; today use SageMaker (Canvas or built-in algorithms) |
| Amazon Fraud Detector | Fraud scoring for online transactions and account registrations (check its availability status; AWS has been winding it down for new customers) |
| Amazon Augmented AI (A2I) | Human review workflows for ML predictions |
| Amazon Q / Amazon Bedrock | Generative AI assistants and foundation models |
| Amazon SageMaker | The platform for building, training and deploying your own models |
Choosing: if a managed API already does the task, use it (fastest, cheapest, no ML team). Use SageMaker when you need a custom model.
SageMaker in brief
The service is now called Amazon SageMaker AI inside the broader SageMaker unified platform, but exam questions still say “SageMaker”.
| Capability | Use |
|---|---|
| SageMaker Studio | Web IDE for the whole ML workflow |
| SageMaker Canvas | No-code ML for analysts: click to build models and forecasts |
| SageMaker Data Wrangler | Visual data preparation and feature transformation |
| SageMaker Feature Store | Central store for reusing features in training and inference |
| SageMaker Ground Truth | Data labeling with human workers (including RLHF data) |
| SageMaker JumpStart | Hub of pre-trained and open-source models and solution templates you can deploy or fine-tune |
| Training jobs / built-in algorithms / AutoPilot (AutoML) | Train models; AutoML picks algorithms and tunes |
| Automatic model tuning | Hyperparameter tuning |
| Experiments / Model Registry | Track runs; version and approve models |
| Endpoints | Real-time, serverless, async inference; Batch Transform for batch |
| SageMaker Pipelines | CI/CD-style workflow orchestration for ML |
| Model Monitor | Detects data drift, model-quality drift, bias drift in production |
| Clarify | Bias detection and explainability |
| Model Cards | Documentation of a model’s purpose, data and risks |
The ML development lifecycle
| Stage | What happens | AWS help |
|---|---|---|
| Business goal | Define the problem and success metric | |
| Data collection | Gather data from sources | S3, Glue, Kinesis, Data Exchange |
| Exploratory data analysis (EDA) | Look at distributions, missing values, correlations | SageMaker Studio notebooks, Data Wrangler, QuickSight |
| Pre-processing | Clean, fill gaps, remove duplicates, normalise | Data Wrangler, Glue DataBrew, Glue |
| Feature engineering | Create or transform the inputs the model learns from | Data Wrangler, Feature Store |
| Training | Fit the model | SageMaker training jobs |
| Hyperparameter tuning | Search for the best settings | Automatic model tuning |
| Evaluation | Measure on held-out test data | SageMaker, Clarify |
| Deployment | Make the model available | SageMaker endpoints, Bedrock |
| Monitoring | Watch for drift and degradation; retrain | Model Monitor, CloudWatch |
Data is normally split into training, validation (tuning) and test (final check) sets. Never evaluate on training data.
Sources of models
- Open-source pre-trained models (Hugging Face, JumpStart): fast, cheap, can be fine-tuned.
- Models via a managed service (Bedrock, Rekognition): no training, pay per use.
- Custom-trained models (SageMaker): most control, most effort and cost.
Using a model in production
| Method | Meaning |
|---|---|
| Managed API service | Call an AWS-hosted API (Bedrock, Comprehend). AWS runs the infrastructure |
| Self-hosted API | You deploy the model to your own endpoint (SageMaker endpoint, EC2, ECS/EKS) and manage scaling, patching and cost |
MLOps
MLOps applies DevOps ideas to ML. Principles to know:
- Experimentation: track experiments so results are reproducible.
- Repeatable processes: automate pipelines instead of running notebooks by hand.
- Scalable systems: infrastructure grows with data and traffic.
- Managing technical debt: untracked data, features and code decay over time.
- Production readiness: testing, versioning, security, rollback.
- Monitoring and retraining: data drift (inputs change) and concept drift (the relationship changes) degrade accuracy, so monitor and retrain.
Model performance metrics
Classification
A confusion matrix counts true positives (TP), false positives (FP), true negatives (TN) and false negatives (FN).
| Metric | Formula idea | Use when |
|---|---|---|
| Accuracy | Correct / all | Classes are balanced. Misleading on imbalanced data (99% “not fraud” gives 99% accuracy for a useless model) |
| Precision | TP / (TP + FP) | False positives are costly (flagging legitimate customers) |
| Recall (sensitivity) | TP / (TP + FN) | False negatives are costly (missing cancer, missing fraud) |
| F1 score | Harmonic mean of precision and recall | Imbalanced data, balance both |
| AUC-ROC | Area under the ROC curve (true-positive rate against false-positive rate). 1.0 is perfect, 0.5 is random | Compare classifiers across all thresholds |
Regression
MAE, MSE, RMSE (average error size; lower is better) and R² (share of variance explained; higher is better).
Business metrics
A technically good model must still pay off. Track ROI, cost per user, development and running costs, customer feedback and satisfaction, and productivity gains.
Common traps
- A managed AI service already does the job, but the answer offers building a custom model in SageMaker. Pick the managed service.
- “Group similar customers with no labels” is clustering, not classification.
- High accuracy on an imbalanced dataset doesn’t mean a good model.
- Overfitting is a variance problem; underfitting is a bias problem.
- Batch inference for scheduled bulk scoring; real-time endpoint for interactive use.
Exam checklist
- Nest AI, ML, deep learning, generative AI and say what each adds.
- Name supervised, unsupervised, reinforcement (and self-supervised) learning with an example each.
- Match regression, classification, clustering to a problem statement.
- Say when ML is not appropriate.
- Map each managed service (Rekognition, Textract, Comprehend, Transcribe, Polly, Translate, Lex, Kendra, Personalize, A2I) to its task.
- Order the pipeline stages and name the SageMaker feature for each (Data Wrangler, Feature Store, Model Monitor, Clarify, Model Cards).
- Contrast managed API with self-hosted model, and batch with real-time inference.
- Choose precision, recall, F1 or AUC for a scenario; explain overfitting and underfitting.