Study Hub / AWS Certified AI Practitioner (AIF-C01) / Domain 1.0
AWS AI Practitioner (AIF-C01) Domain 1.0: Fundamentals of AI and ML — Explained
Domain 1.0 is the vocabulary of artificial intelligence: what AI, machine learning, and deep learning actually are, how models learn from data, and the lifecycle every ML project follows. No math required — just the concepts the rest of the exam assumes.
AI, ML, and deep learning: the nesting dolls
- Artificial intelligence — the broadest term: any system performing tasks that normally need human intelligence (reasoning, perception, language).
- Machine learning — AI's biggest subset: systems that learn patterns from data instead of following hand-written rules. A spam filter trained on a million emails is ML; a spam filter with hand-written keyword rules is just software.
- Deep learning — ML's subset using neural networks with many layers. Excels at images, speech, and language — at the cost of more data and more compute.
- Generative AI — models that create new content (text, images, code) rather than classifying or predicting. A subset of deep learning, and its own exam domain later.
How models learn: the core vocabulary
- Training — feeding labeled (or unlabeled) data to an algorithm so it adjusts its internal parameters. Training is expensive; inference (using the trained model) is comparatively cheap.
- Supervised learning — learns from labeled examples: "this email IS spam." Used for classification (spam/not spam) and regression (predicting a number, like a house price).
- Unsupervised learning — finds structure in unlabeled data: clustering customers by behavior, anomaly detection. No "right answers" provided up front.
- Reinforcement learning — learns by trial and error against a reward signal: game-playing agents, robotics. The model explores actions and reinforces what scores well.
- Overfitting vs. underfitting — overfitting memorizes the training data (great on training, fails on new data); underfitting never learned the pattern at all. The goal is generalization: performing on data it hasn't seen.
- Training / validation / test split — train on one slice, tune on a second, and evaluate once on a third the model never saw during development. Testing on training data proves nothing.
The ML lifecycle and where AWS fits (conceptually)
- Business problem → data → train → evaluate → deploy → monitor — the lifecycle to know by heart. Most real projects fail at "data," not at "model."
- Data quality beats model cleverness — biased, incomplete, or mislabeled data produces biased, unreliable models. "Garbage in, garbage out" is an exam answer in plain clothes.
- Deployment concepts — batch inference (predictions on a schedule, e.g., nightly churn scores) vs. real-time inference (predictions on demand, e.g., fraud checks at checkout).
- Monitoring — models decay: the world changes and yesterday's patterns stop working (model drift). Production ML needs retraining pipelines, not one-time training.
- Human in the loop — for high-stakes decisions (hiring, lending, medical), a human reviews or approves model outputs. The exam treats this as a responsible-AI requirement, not an optional nicety.
The must-memorize table: learning types
Match the scenario to the learning type.
| Type | Data | Example task |
|---|---|---|
| Supervised — classification | Labeled (categories) | Spam vs. not spam; fraud vs. legitimate |
| Supervised — regression | Labeled (numbers) | Predict house price; forecast demand |
| Unsupervised — clustering | Unlabeled | Group customers by behavior |
| Unsupervised — anomaly detection | Unlabeled (mostly normal) | Flag unusual transactions |
| Reinforcement learning | Reward signal | Game agent; robot control |
Try it: 4 sample questions
Sample questions to check your understanding of this domain. Answers and full reasoning are tucked under each one.
1. A bank wants a model that predicts a numeric loan-risk score from applicant data. Which ML approach fits?
- A. A. Supervised regression — predicting a number from labeled examples
- B. B. Unsupervised clustering — grouping applicants
- C. C. Reinforcement learning — trial and error
- D. D. Generative AI — creating new applicants
Show answer & reasoning
A. Predicting a number from labeled historical data is regression, a supervised task: past applicants labeled with their actual outcomes train the model. Clustering (B) groups without predicting; reinforcement (C) needs a reward loop; generative (D) creates content rather than predicting values.
2. A model scores 99% on its training data but 61% on new customer data. What happened, and what is the goal it missed?
- A. A. Underfitting — it never learned the pattern
- B. B. Overfitting — it memorized training data and failed to generalize
- C. C. Data drift — the customers changed overnight
- D. D. Label leakage — the test data was too easy
Show answer & reasoning
B. Huge training accuracy plus poor new-data accuracy is the textbook signature of overfitting: the model memorized noise in the training set instead of learning the underlying pattern. The missed goal is generalization — performing on unseen data, which is the only performance that matters.
3. A retailer wants to group shoppers by purchasing behavior without any predefined categories. Which approach fits?
- A. A. Supervised classification
- B. B. Unsupervised clustering
- C. C. Reinforcement learning
- D. D. Regression
Show answer & reasoning
B. No predefined categories and no labels means unsupervised learning; finding groups in the data is clustering. Supervised methods (A, D) need labeled examples that don't exist here; reinforcement (C) needs a reward signal, not a pile of purchase histories.
4. After deployment, a demand-forecasting model's accuracy slowly declines over six months although nothing in the code changed. What is the most likely cause?
- A. A. Overfitting during the original training
- B. B. Model drift — real-world patterns shifted away from the training data
- C. C. The inference hardware is too slow
- D. D. The training labels were mislabeled
Show answer & reasoning
B. The world moves on: seasons change, competitors act, customer habits evolve — and a model frozen at training time decays. That's model drift, and the fix is monitoring plus retraining pipelines, not a one-time training event. Slow hardware (C) affects latency, not accuracy.
This was one domain. The simulator drills all 5.
This free preview covered Fundamentals of AI & ML (Domain 1.0, 20%). Not covered here: Applications of Foundation Models (Domain 3.0, 28%); Security, Compliance & Governance for AI Solutions (Domain 5.0, 14%). The applied domains — foundation-model use cases and AI security/governance — stay in the paid guide and practice bank. The ByteBar Practice Exam Simulator drills every domain with 1,000 exam-style questions — $9.99 one-time unlocks all 28 exams. Prefer to read? The full AWS AI Practitioner Study Guide is in the ByteBar shop on Etsy.
Practice Exam Simulator — $9.99 one-time, all 28 exams Full Study Guides on Etsy