1.2 Supervised, Unsupervised, and Reinforcement Learning
Key Takeaways
- Supervised learning trains on labeled input-output pairs and is the right answer whenever the business question is prediction or classification against a known ground truth.
- Unsupervised learning finds structure in unlabeled data through clustering, dimensionality reduction, and anomaly detection, and is what powers customer segmentation and embedding spaces.
- Reinforcement learning optimizes a policy through reward signals rather than labeled answers, and RLHF is the stage that turns a raw pretrained LLM into a helpful, instruction-following assistant.
- Foundation models are pretrained with self-supervised learning, which manufactures its own labels from raw text and is why they can absorb web-scale corpora without human annotation.
- Exam scenarios select the paradigm from the data available: labels present means supervised, no labels means unsupervised, feedback-from-consequences means reinforcement.
1.2 Supervised, Unsupervised, and Reinforcement Learning
Executive Summary: The Generative AI Leader exam expects you to identify the machine learning approach behind a business scenario. The discriminator is almost always what kind of data the organization already has. If historical examples carry known correct answers, the approach is supervised. If the data has no answers attached and the goal is to discover structure, the approach is unsupervised. If the system learns by acting and receiving reward or penalty, the approach is reinforcement learning. Foundation models are built by a fourth variant, self-supervised learning, and then aligned with reinforcement learning from human feedback (RLHF).
Why a Leader Needs This Distinction
Leaders are rarely asked to choose an optimizer. They are asked questions like "Can we build this?", "How long will data preparation take?", and "What will it cost?" The learning paradigm answers all three, because it determines whether the organization must first pay for labeling.
A supervised project cannot start until someone has produced ground truth: thousands of loan applications tagged defaulted or repaid, thousands of support tickets tagged with the correct department. That labeling effort is frequently the single largest line item and the longest pole in the schedule. An unsupervised project can begin against raw data immediately, but it produces groupings, not decisions, and a human must still interpret what each group means. A reinforcement learning project needs neither labels nor groupings, but it does need a reward function and a safe environment in which to make mistakes.
The Three Approaches Side by Side
| Dimension | Supervised Learning | Unsupervised Learning | Reinforcement Learning |
|---|---|---|---|
| Training data | Labeled pairs: input plus the known correct output | Unlabeled data only | No dataset; an environment plus a reward signal |
| Question it answers | "What is the correct output for this input?" | "What natural structure exists in this data?" | "What sequence of actions maximizes long-run reward?" |
| Core tasks | Classification, regression, forecasting | Clustering, dimensionality reduction, anomaly detection | Policy optimization, sequential decision-making |
| Business examples | Credit default prediction, invoice field extraction, churn scoring | Customer segmentation, fraud outlier detection, document topic discovery | Data-center cooling control, dynamic pricing, robotics, LLM alignment |
| Google Cloud fit | Agent Platform AutoML (Tabular, Vision), BigQuery ML, Agent Platform Managed Training | BigQuery ML k-means, Agent Platform Vector Search embeddings, clustering pipelines | Reinforcement Learning on Gemini Enterprise Agent Platform; RLHF in Gemini preference tuning |
| Main cost driver | Human labeling effort | Human interpretation of the resulting clusters | Environment/simulator construction and reward design |
Supervised Learning: Learning From Answered Examples
In supervised learning the model sees an input X together with the correct output Y and adjusts its parameters until its predictions match. Two sub-types matter commercially:
- Classification predicts a discrete category. Is this transaction fraudulent? Which of eleven departments should own this ticket? Is this chest X-ray normal or abnormal?
- Regression predicts a continuous number. What will this property sell for? How many units will this SKU move next quarter?
Worked business case. A regional insurer wants to route inbound claims automatically. It has 400,000 historical claims, each already tagged by a human adjuster with one of nine claim types. That existing tagging is the label set, so the insurer can train a supervised classifier without commissioning new annotation. This is the single most common "quick win" pattern in enterprise AI: find a business process where humans have been silently generating labels for years.
The trap the exam sets: a scenario describes rich historical data with known outcomes, and one distractor option proposes clustering. Clustering is wrong when the answer is already known - you would be discarding the most valuable asset in the dataset.
Unsupervised Learning: Finding Structure Without Answers
Unsupervised learning receives only inputs. It looks for structure that was never explicitly described.
- Clustering groups similar records. A retailer with no predefined customer segments runs k-means over purchase histories and discovers six behavioral cohorts it had never articulated.
- Dimensionality reduction compresses high-dimensional data into a smaller representation that preserves relationships. This is the conceptual ancestor of embeddings, which is why unsupervised thinking underpins semantic search and retrieval-augmented generation.
- Anomaly detection identifies records that do not resemble anything else - a strong fit for fraud, industrial fault detection, and network security, where you cannot label every possible failure in advance.
Worked business case. A manufacturer has ten years of sensor telemetry from 4,000 machines but no record of which readings preceded failures. Nobody labeled anything. An unsupervised anomaly detector can still flag statistically unusual vibration signatures for engineer review, and those reviews then create the labels that make a future supervised model possible. Unsupervised first, supervised second, is a very common real-world sequence.
Reinforcement Learning: Learning From Consequences
Reinforcement learning (RL) has no labeled dataset at all. An agent takes an action in an environment, receives a reward or penalty, and updates its policy to earn more reward over time. It suits problems where the correct action at each step is unknown but the eventual outcome is measurable.
Google's own showcase example is data-center cooling: no engineer can label the optimal setting for every combination of load, weather, and equipment state, but energy consumed is directly measurable, so an RL agent can learn a control policy that reduces it.
The framing that matters for the exam: RL is defined by delayed, evaluative feedback. Supervised learning is told the right answer immediately. RL is only told how good the outcome was, sometimes long after the decision.
How Foundation Models Are Actually Trained
Generative AI does not fit neatly into the classic three, and the exam guide's phrasing rewards candidates who can say why.
- Self-supervised pretraining. The model is shown enormous volumes of unlabeled text, code, images, and audio and is trained to predict masked or subsequent tokens. The label is manufactured from the data itself - the next word is both the input's continuation and its own ground truth. This is what makes web-scale training economically possible: no human annotates a trillion tokens. Self-supervised learning is best described as a special case of supervised learning in which the labels are generated automatically rather than by people.
- Supervised fine-tuning (SFT). A comparatively small, high-quality set of human-written prompt/response pairs teaches the model to follow instructions rather than merely autocomplete.
- Reinforcement learning from human feedback (RLHF). Human raters rank competing model outputs. Those rankings train a reward model, and reinforcement learning then optimizes the LLM against that reward model. RLHF is the stage that produces helpfulness, harmlessness, and tone. On Google Cloud this alignment family is exposed as Gemini preference tuning and, more broadly, Reinforcement Learning on Gemini Enterprise Agent Platform.
A leader who can say "the base model came from self-supervised pretraining, the assistant behavior came from SFT plus RLHF" can explain to a board why a raw open-weights base checkpoint behaves so badly in a chatbot, and why instruction-tuned variants exist.
Choosing the Paradigm in an Exam Scenario
Read the scenario for the data condition, not the industry:
| Signal in the scenario | Correct approach |
|---|---|
| "We have 12 years of resolved cases tagged by outcome" | Supervised |
| "We have no categories yet and want to discover natural groupings" | Unsupervised |
| "The system should improve as it observes the results of its own decisions" | Reinforcement learning |
| "We trained on billions of unlabeled web documents" | Self-supervised pretraining |
| "Human reviewers ranked pairs of answers to make the model more helpful" | RLHF |
Strategic Leadership Guidance: Exam Tips and Common Pitfalls
[!TIP] Exam Tip: When a scenario mentions the cost or delay of annotation, the examiner is signalling supervised learning. When it stresses that the organization "does not know what it is looking for," the answer is unsupervised.
Pitfall 1: Calling generative AI "unsupervised." It is not. Pretraining is self-supervised: labels are derived automatically from the data, which is a different claim from having no labels at all.
Pitfall 2: Treating RLHF as a data-collection exercise. RLHF collects preferences (which of these two answers is better), not correct answers. That distinction is why it can improve subjective qualities like tone that no label set could capture.
Pitfall 3: Assuming more labels always beat fewer. For generative use cases, a few hundred high-quality, on-brand examples used for supervised fine-tuning routinely outperform tens of thousands of noisy ones.
A logistics company has ten years of GPS and engine telemetry from its truck fleet. No one has ever recorded which readings preceded a breakdown, and the company cannot describe in advance what a pre-failure pattern looks like. It wants to surface unusual vehicle behavior for mechanic review. Which machine learning approach fits this data condition?
A leader is asked why a large language model can be trained on trillions of words of internet text when the company's own supervised projects stall for months waiting on human annotation. What is the accurate explanation?
An enterprise has fine-tuned a model on high-quality written examples, but reviewers still find its tone inconsistent and occasionally unhelpful. The team collects thousands of comparisons in which raters chose the better of two candidate answers, trains a reward model on those comparisons, and optimizes the assistant against it. Which technique is being described?