監督式、非監督式與強化學習

The Supervised, Unsupervised, and Reinforcement Learning taxonomy is the first vocabulary AIF-C01 candidates must master in Domain 1. Task statement 1.1 "Explain basic AI concepts and terminologies" routinely puts a business scenario in front of you and asks which flavor of supervised, unsupervised, and reinforcement learning applies. Pick wrong and a good chunk of the 20% Domain 1 weight evaporates.

This study guide walks every dimension of supervised, unsupervised, and reinforcement learning that appears on AIF-C01, plus the self-supervised and semi-supervised variants that power modern foundation models. You will learn the label-availability heuristic that decides the family, the AWS services that implement each paradigm (SageMaker built-in algorithms, SageMaker JumpStart, Amazon DeepRacer, Amazon Bedrock fine-tuning), and the subtle supervised, unsupervised, and reinforcement learning distinctions the AIF-C01 exam uses as trap doors — including the notorious Domain Adaptation vs Transfer Learning pair flagged in community pain-point reports.

What Are Supervised, Unsupervised, and Reinforcement Learning?

Supervised, unsupervised, and reinforcement learning are the three canonical machine-learning paradigms, distinguished by the kind of training signal the model sees. Supervised learning uses labeled input-output pairs. Unsupervised learning uses inputs only and hunts for structure. Reinforcement learning uses a reward signal earned by interacting with an environment. Every ML problem on AIF-C01 — and in the wider AWS ML stack — ultimately reduces to one of these three (plus two hybrids we will cover: self-supervised and semi-supervised).

The AIF-C01 exam guide lists these paradigms explicitly under Task 1.1 and expects candidates to (a) recognize each by keywords in a scenario, (b) pick the correct AWS service that hosts that paradigm, and (c) avoid the trap questions where a plausible-sounding but wrong paradigm is offered.

Why AIF-C01 Obsesses Over the Three-Family Taxonomy

Domain 1 is worth 20% of the AIF-C01 exam and Task 1.1 alone drives roughly six to ten scored questions in a 65-question sitting. Scenario questions almost never ask "define reinforcement learning." Instead they describe a problem ("an engineering team wants a robot to learn to walk without hand-written rules") and expect you to snap to the paradigm. Memorizing keywords is part of it; understanding the shape of the training signal is the durable skill.

How This Topic Links to the Rest of AIF-C01

The supervised, unsupervised, and reinforcement learning vocabulary is the conceptual root of:

• Classification metrics (accuracy, precision, recall, F1) — evaluated in Overfitting, Bias, and Variance
• The ML Development Lifecycle — data labeling cost depends directly on which paradigm you picked
• Foundation models — pre-training is self-supervised; fine-tuning is supervised; RLHF is reinforcement learning from human feedback
• AWS service mapping — SageMaker built-in algorithms partition cleanly into the three families

Why Learning Type Matters: The Label Availability Test

Before any algorithm selection, AWS recommends a single diagnostic question: do you have labels, can you get them cheaply, and what do the labels look like? The answer routes you through supervised, unsupervised, and reinforcement learning in predictable ways.

• Lots of labeled data → supervised learning is default
• No labels but lots of raw inputs → unsupervised learning
• No labels but a simulator/environment with a scoring function → reinforcement learning
• Labels are expensive but unlabeled data is abundant → semi-supervised or self-supervised learning
• Labels emerge from the data itself (predict next token, predict masked pixel) → self-supervised learning

This ordering reappears below and is the single most useful heuristic for supervised, unsupervised, and reinforcement learning scenario questions.

白話文解釋 Supervised, Unsupervised, and Reinforcement Learning

The supervised, unsupervised, and reinforcement learning distinction becomes obvious when you step out of ML jargon and into everyday environments. Three analogies from different domains make the differences unforgettable.

The Open-Book Exam Analogy (Supervised Learning)

Imagine a student preparing for a certification exam. The study materials are a giant question bank where every question comes with the correct answer printed next to it. The student reviews thousands of (question, answer) pairs, notices patterns in how answers are structured, and on exam day is presented with brand-new questions whose answers are hidden. If the student has truly learned, predictions on the new questions will be right most of the time.

That is supervised learning. The labeled question bank is the training set. The hidden-answer exam is the test set. The whole point is generalization from labels you already have to inputs you have not seen. Classification problems (Is this email spam? yes/no) and regression problems (What will this house sell for? a number) are just different output shapes of the same supervised, unsupervised, and reinforcement learning family — the supervised branch.

The Library Analogy (Unsupervised Learning)

Now imagine a new librarian walking into a warehouse full of unshelved, uncatalogued books. Nobody told them which books belong to which genre. All they have is the text itself. Their job is to notice that some books share vocabulary, themes, and writing style — "these fifty books keep mentioning starships; they probably belong together." The librarian is not predicting a pre-existing label. They are discovering structure latent in the data.

That is unsupervised learning. K-Means clustering groups similar customers. Principal Component Analysis squeezes 500 features into 20 while keeping most of the variance. Anomaly detection notices a transaction that does not look like any of the clusters and flags it as potential fraud. Across all three, no ground-truth label exists — the supervised, unsupervised, and reinforcement learning unsupervised branch earns its keep by surfacing patterns no one explicitly taught the model.

The Video-Game Analogy (Reinforcement Learning)

Finally, picture a child learning a new platformer video game. Nobody gives them a textbook of correct moves. They press buttons, fall in lava, die, respawn, press different buttons, reach checkpoints, collect coins, unlock levels. Every reward (coin collected, level cleared) or penalty (death, lost life) adjusts their strategy. Over hours of play, a policy emerges: this state plus this button-combo yields the highest cumulative score.

That is reinforcement learning. The agent is the child, the environment is the game, the state is the current screen, the action is the button press, and the reward is the score. Unlike supervised learning, there is no teacher saying "correct action was X." Unlike unsupervised learning, there is a clear objective (maximize score). Reinforcement learning is the supervised, unsupervised, and reinforcement learning family branch that thrives when you can score outcomes but cannot enumerate correct actions — exactly what AWS DeepRacer simulates on a virtual race track.

Which Analogy to Use on Exam Day

All three analogies describe the supervised, unsupervised, and reinforcement learning taxonomy from a different angle. Pick based on scenario wording:

• Scenario mentions labeled data, historical outcomes, training examples with answers → open-book exam (supervised)
• Scenario mentions discover patterns, group similar items, no labels available → library (unsupervised)
• Scenario mentions agent, environment, reward, policy, trial and error, simulator → video game (reinforcement)

Supervised Learning: Labeled Data, Precise Predictions

Supervised learning is the most common form of supervised, unsupervised, and reinforcement learning in production AWS workloads and the most common AIF-C01 scenario. You have historical data where every row carries a ground-truth answer, and you want the model to predict that answer on future rows.

The Core Assumption: Labels Exist

Supervised learning presumes you either already have labels or can obtain them. Labels might come from human annotators, business outcomes that naturally labeled themselves (the customer did churn, or did not), or automated labeling via Amazon SageMaker Ground Truth. Without labels, you cannot do supervised learning, full stop — that constraint is what separates it from its supervised, unsupervised, and reinforcement learning siblings.

Classification vs Regression: The Output-Type Fork

Supervised learning splits into two sub-categories by output type:

• Classification predicts a category from a discrete set. Is this email spam or not? Is this X-ray showing pneumonia, tuberculosis, or nothing? Classification outputs a class label (sometimes with a probability).
• Regression predicts a continuous numeric value. What will this house sell for? How many units will we ship next month? Regression outputs a real number.

The AIF-C01 exam repeatedly tests this fork. Any scenario whose answer is a number (price, temperature, demand) is regression. Any scenario whose answer is a category (yes/no, A/B/C/D, fraud/not-fraud) is classification.

Training, Validation, and Test Splits

Supervised learning rests on splitting labeled data into three subsets:

• Training set (typically 60–80%) — the model learns patterns here.
• Validation set (typically 10–20%) — used to tune hyperparameters without touching the test set.
• Test set (typically 10–20%) — used exactly once at the end to estimate real-world performance.

Mixing these up leaks information and inflates accuracy. AIF-C01 expects you to know that tuning on the test set is a cardinal sin and the validation set exists precisely to prevent it. The formal name is cross-validation when the splits are rotated, and k-fold cross-validation when the rotation covers k distinct folds.

Common Supervised Algorithms (Conceptual Level)

AIF-C01 does not require mathematical derivations, but expects recognition of the following algorithm families:

• Linear regression — fits a line (or hyperplane) to continuous outputs.
• Logistic regression — despite the name, this is classification, outputting a probability.
• Decision trees — rule-based splits, highly interpretable.
• Random forests — ensembles of decision trees, robust to overfitting.
• Gradient-boosted trees / XGBoost — the workhorse of tabular ML on AWS; SageMaker has a built-in XGBoost algorithm.
• Support vector machines (SVMs) — margin-based classifiers.
• Neural networks — from shallow MLPs to deep CNNs and transformers.

AWS Services That Host Supervised Learning

The supervised branch of supervised, unsupervised, and reinforcement learning is the best-covered by AWS services:

• Amazon SageMaker built-in algorithms — Linear Learner, XGBoost, Factorization Machines, Image Classification, Object Detection, BlazingText (supervised mode), Seq2Seq.
• Amazon SageMaker JumpStart — pre-trained supervised models for text, vision, and tabular tasks.
• Amazon SageMaker Canvas — no-code supervised classification and regression for business analysts.
• Amazon Comprehend Custom Classification and Custom Entity Recognition — supervised NLP on your labeled text.
• Amazon Rekognition Custom Labels — supervised image classification and object detection on your labeled images.
• Amazon Fraud Detector — supervised fraud classifier trained on your historical outcomes.

Unsupervised Learning: Pattern Discovery Without Labels

Unsupervised learning is the second family of supervised, unsupervised, and reinforcement learning and is used whenever labels are missing or expensive. Instead of predicting a target, unsupervised models find structure in the data itself.

Clustering: Grouping Similar Items

Clustering partitions inputs into groups such that items inside a cluster are more similar to one another than to items in other clusters.

• K-Means — pick k, then iteratively assign points to nearest centroid and recompute centroids until stable. AIF-C01's most-tested clustering algorithm.
• Hierarchical clustering — build a tree of nested clusters (agglomerative bottom-up or divisive top-down).
• DBSCAN — density-based, handles arbitrary cluster shapes and marks outliers automatically.

K-Means is available as a SageMaker built-in algorithm, which is the most common AIF-C01 exam-surface mapping.

Dimensionality Reduction: Compression Without Labels

When your data has 500 features, most downstream algorithms slow down or overfit. Dimensionality reduction compresses the feature space while preserving as much signal as possible.

• Principal Component Analysis (PCA) — projects data onto orthogonal axes that capture maximum variance. Available as a SageMaker built-in algorithm.
• t-SNE and UMAP — nonlinear embedding methods mostly used for visualization.
• Autoencoders — neural networks that learn to compress and reconstruct inputs; the bottleneck layer is the low-dimensional representation.

Anomaly Detection: Finding the Weird Ones

Anomaly detection surfaces inputs that do not fit the observed pattern. Fraud detection, infrastructure monitoring, and manufacturing defect detection all lean on this capability.

• Random Cut Forest (RCF) — SageMaker's built-in anomaly algorithm, also used inside Amazon Kinesis Data Analytics and Amazon Lookout.
• Isolation Forest — isolates anomalies by random partitioning; shorter paths correspond to outliers.
• Amazon Lookout for Metrics / Equipment / Vision — managed anomaly-detection services built on unsupervised techniques.

AWS Services That Host Unsupervised Learning

Unsupervised supervised, unsupervised, and reinforcement learning coverage on AWS:

• SageMaker built-in algorithms — K-Means, PCA, Random Cut Forest, IP Insights (anomaly on IP pairs).
• Amazon Lookout for Metrics — anomaly detection on time-series business KPIs.
• Amazon Lookout for Equipment — anomaly detection on industrial sensor data.
• Amazon Lookout for Vision — visual defect detection using unsupervised feature learning.
• Amazon Kinesis Data Analytics — RANDOM_CUT_FOREST SQL function for streaming anomaly detection.

Reinforcement Learning: Learning from Rewards

Reinforcement learning (RL) is the third canonical family of supervised, unsupervised, and reinforcement learning and the one most likely to appear as a distractor when it is actually correct. AIF-C01 covers RL at a conceptual level — enough to identify it from scenario wording and map it to AWS DeepRacer and SageMaker RL.

The Agent-Environment-Reward-Policy Loop

RL formalizes the trial-and-error learning loop with five standard terms:

• Agent — the decision-maker being trained.
• Environment — the world the agent interacts with (real or simulated).
• State — a snapshot of the environment the agent observes.
• Action — what the agent chooses to do.
• Reward — numeric feedback from the environment after the action.
• Policy — the learned mapping from states to actions; the "strategy" of the agent.

Each time the agent picks an action, the environment transitions to a new state and emits a reward. Over many such steps the agent updates its policy to maximize cumulative future reward (formally, the "expected discounted return").

Q-Learning Intuition

Q-Learning is the textbook RL algorithm AIF-C01 sometimes names. Intuitively, Q(s, a) is a table (or neural network) that estimates "how much total reward can I expect if I take action a from state s and then act optimally thereafter?" The agent repeatedly refines Q based on observed rewards, then picks actions by choosing the a that maximizes Q(s, a) in each state. Deep Q-Networks (DQN) replace the table with a neural network so the approach scales to high-dimensional state spaces like pixels.

Exploration vs Exploitation

RL agents face a perpetual tension: should I exploit what I already know (pick the action with the highest estimated reward) or explore (try something new that might turn out better)? The epsilon-greedy strategy — act greedily 1 − epsilon of the time and randomly epsilon of the time — is the classic balance. This tension has no analogue in supervised learning, which is one reason RL sits in its own family of supervised, unsupervised, and reinforcement learning.

On-Policy vs Off-Policy (Conceptual)

On-policy methods (e.g. SARSA) learn about the policy they are currently executing. Off-policy methods (e.g. Q-Learning) can learn about the optimal policy while executing a different (exploratory) one. AIF-C01 rarely drills this distinction — knowing it exists is enough.

AWS DeepRacer: The Exam-Friendly RL Example

AWS DeepRacer is an autonomous 1/18th-scale race car that learns to drive around a track using RL. The student defines a reward function (e.g. reward = track_width − distance_from_center), picks a training algorithm (PPO or SAC), trains in a simulator, and optionally deploys to a physical car. DeepRacer is the AIF-C01 canonical RL example because (a) the five RL terms map cleanly onto it, (b) AWS markets it as an educational tool, and (c) it maps to SageMaker RL under the hood.

Other AWS Services That Host Reinforcement Learning

• SageMaker RL — fully managed RL training jobs with RLlib, Coach, and TensorFlow/PyTorch backends.
• Amazon Bedrock RLHF (Reinforcement Learning from Human Feedback) — used internally by foundation-model providers to align models to human preferences; surfaces as a concept when discussing model fine-tuning.

Self-Supervised Learning: How Foundation Models Pretrain

Self-supervised learning is a relatively recent addition to the supervised, unsupervised, and reinforcement learning family tree and the secret sauce behind every modern foundation model on Amazon Bedrock.

The Trick: Create Labels from the Data Itself

Self-supervised learning is technically a special case of supervised learning — there are labels — but the labels are synthesized from the data itself rather than annotated by humans. For text, the label for each position is simply "what token comes next?" For images, the label might be "which patch was masked out?" For audio, "which chunk was silenced?" You need no human annotators, yet you have infinitely many training examples.

Why It Matters for Foundation Models

Every large language model (LLM) on Amazon Bedrock — Anthropic Claude, Amazon Titan, Meta Llama, Mistral, Cohere — was pre-trained self-supervised on a giant text corpus. The predict-next-token objective gives each model billions of free training examples at essentially zero labeling cost. This is the economic reason foundation models exist: you could never afford to hand-label trillions of tokens, but self-supervision sidesteps the cost entirely.

Other self-supervised objectives:

• Masked language modeling (MLM) — BERT-style models predict randomly masked tokens.
• Masked image modeling — vision models predict masked image patches.
• Contrastive learning — bring similar pairs closer in embedding space, push dissimilar pairs apart (used by CLIP and Titan Multimodal Embeddings).
• Next-token prediction (autoregressive) — GPT-style models predict the next token given all previous tokens; the standard objective for modern LLMs.

Self-Supervised vs Unsupervised: The Fine Line

Some textbooks lump self-supervised under unsupervised learning because neither uses human labels. AIF-C01 treats them as distinct: self-supervised has an explicit prediction target (even if auto-generated), whereas unsupervised learning has no target at all (clustering has no "correct" answer). When in doubt on the exam, remember: if the model is trained to predict something, it is supervised (including self-supervised); if the model is discovering structure with no predicted output, it is unsupervised.

Semi-Supervised Learning: When Labels Are Scarce

Semi-supervised learning is the last paradigm AIF-C01 expects you to recognize. It sits between supervised and unsupervised in the supervised, unsupervised, and reinforcement learning taxonomy and is useful whenever you have a small labeled dataset plus a large pool of unlabeled data.

The Typical Recipe

You train a first-pass model on the small labeled set, use it to generate pseudo-labels on the unlabeled data, keep the high-confidence pseudo-labels, and retrain on the combined set. Alternatives include co-training (two models label each other's unlabeled data) and graph-based label propagation.

When to Pick Semi-Supervised

Semi-supervised is the right pick when:

• Labeling cost is high (medical imaging, legal documents)
• Unlabeled data is abundant (you have millions of product reviews, only thousands labeled)
• Unlabeled distribution matches the labeled distribution (crucial; otherwise pseudo-labels are poisonous)

AWS Services That Support Semi-Supervised Workflows

• Amazon SageMaker Ground Truth — active learning built in; the system automatically labels the "easy" samples and routes hard ones to human annotators, which operationally is a form of semi-supervised labeling.
• Amazon Comprehend Custom — training on small labeled sets with transfer from Amazon's pre-trained base models.

Semi-supervised learning is rarely a direct answer on AIF-C01, but recognizing it lets you eliminate distractors in supervised, unsupervised, and reinforcement learning scenario questions.

When to Pick Each: Problem Shape + Label Availability

A consolidated decision flow for the whole supervised, unsupervised, and reinforcement learning family plus the two hybrids:

1. Do I have an environment with rewards and actions? → Reinforcement learning (DeepRacer, SageMaker RL).
2. Can I generate labels for free from the data itself (predict next token, masked patch)? → Self-supervised learning (foundation-model pre-training).
3. Do I have abundant labels? → Supervised learning.
   • Output is a category → Classification.
   • Output is a number → Regression.
4. Do I have a few labels plus lots of unlabeled data? → Semi-supervised learning.
5. Do I have only unlabeled data and need to discover structure? → Unsupervised learning.
   • Group similar items → Clustering.
   • Compress features → Dimensionality reduction.
   • Flag outliers → Anomaly detection.

Keyword-to-Paradigm Cheat Table

SageMaker Built-in Algorithms Mapped to the Paradigms

Amazon SageMaker ships a suite of built-in algorithms that map cleanly onto supervised, unsupervised, and reinforcement learning. AIF-C01 asks you to recognize which algorithm is used for which paradigm.

Supervised Built-ins

• Linear Learner — classification or regression on tabular data.
• XGBoost — gradient-boosted trees, the tabular workhorse; classification or regression.
• Factorization Machines — for high-dimensional sparse datasets (clickstream, recommendations).
• Image Classification (ResNet-based) and Object Detection (SSD-based) — computer vision supervised tasks.
• Semantic Segmentation — pixel-level classification.
• BlazingText (supervised mode) — text classification.
• Seq2Seq — sequence-to-sequence tasks like translation.
• DeepAR — supervised time-series forecasting.

Unsupervised Built-ins

• K-Means — clustering.
• PCA — dimensionality reduction.
• Random Cut Forest — anomaly detection.
• IP Insights — anomaly detection on IP address pairs.
• BlazingText (unsupervised Word2Vec mode) — word-embedding learning.
• Object2Vec — general-purpose neural embeddings.
• Neural Topic Model (NTM) and LDA — topic modeling over text corpora.

Reinforcement Built-ins

• SageMaker RL — not a single algorithm but a framework supporting PPO, SAC, DQN via RLlib and Coach.
• AWS DeepRacer — packages RL into a gamified consumer-facing product.

Transfer Learning vs Domain Adaptation: The Exam Trap

Community pain-point reports consistently flag the distinction between Transfer Learning and Domain Adaptation as one of the most-missed supervised, unsupervised, and reinforcement learning adjacent concepts on AIF-C01. The two terms look synonymous in casual usage but the exam treats them as distinct.

Transfer Learning: New Task, Reuse Knowledge

Transfer learning reuses a model trained on one task to jump-start a different task. Classic example: a ResNet pre-trained on ImageNet (1000-class general image classification) is fine-tuned on a small dataset of chest X-rays (binary pneumonia detection). The source task (ImageNet classification) and target task (X-ray classification) are different tasks. The feature extractor layers are reused; the classification head is replaced or fine-tuned.

Transfer learning is about changing the task.

Domain Adaptation: Same Task, New Domain

Domain adaptation keeps the same task but adapts the model to a new data distribution. Classic example: a sentiment classifier trained on Amazon product reviews is adapted to classify sentiment on movie reviews. The task (sentiment classification) is identical. What changed is the domain (the statistical distribution of the inputs — vocabulary, style, length).

Domain adaptation is about changing the data distribution while keeping the task.

Side-by-Side Comparison

Why the Exam Loves This Pair

Both concepts describe "use an existing model on a new situation." Without the precise definitions above, candidates shrug and pick whichever word sounds more familiar. The AIF-C01 exam guide explicitly lists both, and community reports confirm they appear as distinct answer options in the same question — forcing you to choose correctly.

Common Exam Traps for Supervised, Unsupervised, and Reinforcement Learning

Beyond the two traps already flagged (binary classification vs regression, clustering vs classification), AIF-C01 fields a handful of recurring supervised, unsupervised, and reinforcement learning trick patterns.

Trap 1: "Unsupervised" Does Not Mean "No Training"

Unsupervised models are still trained — they just train without labels. Do not pick "unsupervised" as a synonym for "rule-based" or "no ML involved."

Trap 2: RL Is Not Always the Right Answer for "Adaptive" Systems

Adaptive recommendation engines are often supervised (collaborative filtering, matrix factorization) or a hybrid. RL shows up only when the system takes sequential actions in an environment with delayed rewards. If the scenario is just "recommend products based on past purchases," the answer is supervised, not reinforcement.

Trap 3: Self-Supervised Pre-Training ≠ Fine-Tuning

Foundation models go through two phases: self-supervised pre-training on a giant corpus, then (optionally) supervised fine-tuning on a smaller labeled set. An AIF-C01 question about fine-tuning on domain-specific labeled data is asking about the supervised fine-tuning step, not the original self-supervised pre-training.

Trap 4: Anomaly Detection Can Be Supervised or Unsupervised

Most anomaly detection on AWS (Random Cut Forest, Lookout family) is unsupervised — you do not have pre-labeled anomalies. But Amazon Fraud Detector uses supervised learning because you label past transactions as fraudulent or legitimate. The paradigm depends on whether the anomalies are pre-labeled.

Trap 5: DeepRacer's Reward Function Is Written by the Human, Not Learned

A common misconception: "DeepRacer learns its own reward." Wrong. The student writes the reward function in Python. The agent learns a policy that maximizes the reward function the student defined. If the reward function is bad, the learned policy will be bad.

Numbers and Constants to Memorize

A handful of numbers show up repeatedly when AIF-C01 exam writers build supervised, unsupervised, and reinforcement learning scenario questions. These are not in any one whitepaper but reflect AWS documentation defaults and typical guidance.

Supervised, Unsupervised, and Reinforcement Learning vs Foundation Models

Modern foundation models blur the old boundaries of supervised, unsupervised, and reinforcement learning. Understanding how they overlap is a fast way to answer comparison questions.

Pre-Training: Self-Supervised

Foundation-model pre-training is self-supervised. Predict the next token, predict the masked patch. No human labels.

Supervised Fine-Tuning (SFT)

After pre-training, providers often fine-tune the model on a smaller dataset of (instruction, ideal-response) pairs. This phase is pure supervised learning — the label is the ideal response.

Reinforcement Learning from Human Feedback (RLHF)

A third alignment phase applies reinforcement learning. A reward model is trained on human rankings of model outputs, and the policy (the LLM) is updated to maximize the reward-model score. The final alignment of models like Claude and Titan uses RLHF, bringing all three families of supervised, unsupervised, and reinforcement learning into a single production model.

Takeaway for the Exam

When a AIF-C01 question asks how foundation models are trained, the correct answer mentions all three phases: self-supervised pre-training, supervised fine-tuning, and reinforcement learning from human feedback. Single-paradigm answers are usually wrong for foundation-model questions.

Question Links: AIF-C01 Task 1.1 Practice Pointers

AIF-C01 task statement 1.1 "Explain basic AI concepts and terminologies" exercises the supervised, unsupervised, and reinforcement learning vocabulary through scenario-based questions. The templates below are the most common shapes you will see. Detailed practice questions with full explanations appear in the ExamHub question bank.

Template A: Label-Availability Routing

A retailer has five years of customer transaction data with no labels indicating which customers are "high value," and wants to discover natural customer segments to design targeted marketing campaigns. Which paradigm applies? Answer: unsupervised learning (clustering). Distractors: supervised classification (wrong because no labels), reinforcement learning (wrong because no environment/reward).

Template B: Output-Type Fork

A real-estate startup wants to predict the sale price of a house from square footage, location, and number of bedrooms. Which paradigm applies? Answer: supervised regression (continuous numeric target). Distractor: supervised classification (wrong because sale price is continuous, not categorical).

Template C: Environment-with-Reward Pattern

A logistics company wants a warehouse robot to learn the fastest path to pick items, adapting to changing warehouse layouts. Which paradigm applies? Answer: reinforcement learning (agent, environment, reward over time). Distractors: supervised learning (wrong because no labeled optimal paths exist).

Template D: Domain Adaptation vs Transfer Learning

A data scientist has a sentiment-analysis model trained on English product reviews and wants to adapt it to classify sentiment on English movie reviews. Which technique applies? Answer: domain adaptation (same task, different domain). Distractor: transfer learning (wrong because the task is unchanged).

Template E: Foundation-Model Pre-Training

An AI team wants to pre-train a language model on a 1-TB corpus of legal documents with no human labels. Which paradigm applies? Answer: self-supervised learning (predict-next-token is the implicit objective). Distractors: unsupervised learning (technically adjacent but does not capture the prediction objective), supervised learning (no human labels).

Supervised, Unsupervised, and Reinforcement Learning Frequently Asked Questions (FAQ)

What is the difference between supervised, unsupervised, and reinforcement learning?

Supervised learning trains a model on (input, label) pairs to predict labels for new inputs. Unsupervised learning trains on inputs alone and finds hidden structure such as clusters or low-dimensional embeddings. Reinforcement learning trains an agent to pick actions in an environment, guided by a reward signal, to maximize cumulative return. These three families of supervised, unsupervised, and reinforcement learning cover most classical ML on AWS and appear explicitly in the AIF-C01 exam guide.

How do I know whether to use supervised or unsupervised learning?

Start with the label-availability test. If you have labeled training data and want to predict the same label for new inputs, use supervised learning. If you have only raw inputs and want to discover structure (groups, compressed features, anomalies), use unsupervised learning. If labels are scarce but unlabeled data is abundant, semi-supervised learning or self-supervised pre-training followed by supervised fine-tuning may beat pure supervised.

Is classification the same as regression?

No. Both are supervised learning, but classification predicts a category from a discrete set (spam / not spam, A / B / C) while regression predicts a continuous numeric value (price, temperature). AIF-C01 asks you to distinguish them by looking at the target variable's type. Numeric → regression. Categorical → classification. Binary yes/no is classification (specifically binary classification), not regression, even though yes/no can be encoded as 0/1.

What is reinforcement learning and when should I pick it on AIF-C01?

Reinforcement learning trains an agent via trial and error to maximize reward in an environment. Pick it when the scenario involves sequential decisions, a simulator or real environment, and a scoring function rather than a labeled dataset. AWS DeepRacer is the AIF-C01 canonical example. Keywords like "agent," "reward," "policy," "simulator," and "trial and error" should snap your attention to reinforcement learning.

What is self-supervised learning and how does it relate to foundation models?

Self-supervised learning creates labels from the data itself — for text, predict the next token; for images, predict the masked patch — so no human annotation is needed. Every modern foundation model on Amazon Bedrock (Anthropic Claude, Amazon Titan, Meta Llama, Mistral) is pre-trained self-supervised on a giant corpus. Self-supervised pre-training is what makes foundation models economically feasible — human-labeled data at the required scale would cost billions. AIF-C01 expects you to recognize self-supervised learning as the mechanism behind foundation-model pre-training.

What is the difference between transfer learning and domain adaptation?

Transfer learning changes the task (e.g. ImageNet classifier → chest X-ray classifier) by reusing learned feature representations. Domain adaptation keeps the task and adapts the model to a new input distribution (e.g. product-review sentiment classifier → movie-review sentiment classifier). Both concepts appear in AIF-C01 Task 1.1 and 3.3, and community reports consistently flag this pair as a high-trap distinction. Memorize: transfer learning = new task; domain adaptation = new data distribution.

Which AWS services implement each learning paradigm?

For supervised learning, use SageMaker built-in algorithms (XGBoost, Linear Learner, Image Classification), SageMaker JumpStart, Comprehend Custom, Rekognition Custom Labels, Amazon Fraud Detector, and Amazon Personalize. For unsupervised learning, use SageMaker K-Means, PCA, Random Cut Forest, and the Amazon Lookout family for anomaly detection. For reinforcement learning, use AWS DeepRacer and SageMaker RL. For self-supervised learning, Amazon Bedrock continued pre-training lets you continue a foundation model's self-supervised training on your own unlabeled data.

Does AIF-C01 require me to code any of these algorithms?

No. AIF-C01 is a foundational certification that tests conceptual understanding, vocabulary, and AWS service mapping. You need to recognize the paradigms, distinguish their sub-types, match scenarios to the correct family, and identify which AWS service hosts which approach. You do not need to implement K-Means or gradient descent. That level of depth is reserved for the AWS Certified Machine Learning Engineer – Associate (MLA-C01) and AWS Certified Machine Learning – Specialty (MLS-C01) certifications.

Further Reading

• AWS AIF-C01 Exam Guide: https://d1.awsstatic.com/training-and-certification/docs-ai-practitioner/AWS-Certified-AI-Practitioner_Exam-Guide.pdf
• Amazon SageMaker Built-in Algorithms: https://docs.aws.amazon.com/sagemaker/latest/dg/algos.html
• AWS DeepRacer Developer Guide: https://docs.aws.amazon.com/deepracer/latest/developerguide/what-is-deepracer.html
• Amazon SageMaker K-Means: https://docs.aws.amazon.com/sagemaker/latest/dg/k-means.html
• Amazon SageMaker XGBoost: https://docs.aws.amazon.com/sagemaker/latest/dg/xgboost.html
• Amazon SageMaker JumpStart Fine-Tuning: https://docs.aws.amazon.com/sagemaker/latest/dg/jumpstart-fine-tune.html
• Amazon Bedrock Custom Models (continued pre-training and fine-tuning): https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html
• AWS Cloud Adoption Framework for AI, ML, and Generative AI: https://docs.aws.amazon.com/whitepapers/latest/aws-caf-for-ai/aws-caf-for-ai.html
• Machine Learning Best Practices for Public Sector Organizations: https://docs.aws.amazon.com/whitepapers/latest/ml-best-practices-public-sector-organizations/introduction.html

Related ExamHub topics: AI and ML Core Concepts, Overfitting, Underfitting, Bias, and Variance, ML Development Lifecycle, Practical AI and ML Use Cases.