Fundamentals of gen AI · Topic 2 of 6
Machine Learning Approaches and the ML Lifecycle
Supervised, unsupervised and reinforcement learning, and the ML lifecycle from data ingestion to model management with the Google Cloud tools for each stage.
- 4 min
- Includes a diagram
- Supervised: labeled data, predict known outcomes. Unsupervised: unlabeled data, find structure. Reinforcement: learn from rewards.
- Lifecycle: ingestion, preparation, training, deployment, management, then back again.
- Agent Platform covers training, deployment and management; BigQuery and Cloud Storage cover data.
What it is
Machine learning models learn in different ways. The exam names three approaches: supervised, unsupervised and reinforcement learning. It also expects you to know the stages of the ML lifecycle (data ingestion, data preparation, model training, model deployment, model management) and the Google Cloud tools used at each stage.
How it works
The three approaches
| Approach | How it learns | Typical business use | Example |
|---|---|---|---|
| Supervised learning | From labeled examples: inputs paired with the correct answer | Classification and prediction | Emails labeled spam or not spam train a spam filter |
| Unsupervised learning | From unlabeled data, finding structure on its own | Grouping, pattern finding, anomaly detection | Segmenting customers by purchase behavior |
| Reinforcement learning | By trial and error, receiving rewards or penalties for actions | Decision making in changing environments | Optimizing warehouse robot routes; improving model responses using human feedback |
Look for the clue. "Historical data with known outcomes" is supervised. "Find natural groups" is unsupervised. "Learns by reward" is reinforcement.
Generative AI models use a mix: they learn language patterns from huge amounts of unlabeled text (a form of self-supervised learning), are then refined with labeled examples, and are often improved with reinforcement learning from human feedback.
The ML lifecycle
| Stage | What happens | Google Cloud tools |
|---|---|---|
| Data ingestion | Collect data from sources such as apps, files, databases and streams | Cloud Storage, Pub/Sub, BigQuery |
| Data preparation | Clean, transform, label and split data; create features | BigQuery, Dataflow, data labeling, Agent Platform Feature Store |
| Model training | Train or adapt a model, and evaluate it | Gemini Enterprise Agent Platform (custom training, AutoML, model tuning), BigQuery ML |
| Model deployment | Make the model available to applications | Agent Platform endpoints for online and batch predictions, Model Registry |
| Model management | Monitor quality, detect drift, retrain, version and govern | Agent Platform model monitoring, Agent Platform Pipelines, Model Registry |
The lifecycle is a loop, not a line. Data changes, models drift, and teams retrain. Managing that loop with automation is often called MLOps.
Why the lifecycle matters to leaders
- Most effort goes into data, not into the model. Budget and plan for data work.
- A model that is never deployed creates no value.
- A deployed model that is not managed slowly gets worse as the world changes.
- Using a managed platform (Agent Platform) for the whole loop reduces cost and risk.
Foundation models and the lifecycle
With foundation models, many organizations skip training from scratch. They start from a pre-trained model, adapt it with prompts, grounding or tuning, then deploy and manage it. The lifecycle still applies: data preparation (for grounding or tuning), evaluation, deployment and monitoring remain.
A worked scenario
An insurer wants to flag likely fraudulent claims. Past claims marked fraud or genuine are labeled data, so it uses supervised learning. Claims data flows into BigQuery (ingestion), is cleaned and turned into features (preparation), a model is trained with AutoML on Agent Platform (training), deployed to an endpoint that scores each new claim (deployment), and monitored for drift and retrained every quarter with Agent Platform Pipelines (management).
How the exam asks about this
- A learning situation is described and you choose the approach.
- A lifecycle stage is described and you name it, or choose the Google Cloud tool for it.
- "What happens after a model is deployed?" It is monitored and managed, and retrained when it drifts.
What to remember for the exam
- Supervised: labeled data, predict known outcomes. Unsupervised: unlabeled data, find structure. Reinforcement: learn from rewards.
- Lifecycle: ingestion, preparation, training, deployment, management, then back again.
- Agent Platform covers training, deployment and management; BigQuery and Cloud Storage cover data.
- Foundation models shorten the training stage but do not remove the rest of the lifecycle.
"Deployed" is not "done". Without monitoring, model quality falls silently as data changes.
Common mistakes
Calling clustering supervised. Clustering finds groups without labels, so it is unsupervised.
Forgetting data preparation. It is usually the largest part of the work.
Treating the lifecycle as one pass. It is a continuous loop.
Check yourself
Which ML approach learns from labeled examples?
Supervised learning.
Name the five stages of the ML lifecycle.
Data ingestion, data preparation, model training, model deployment and model management.
Which approach learns by receiving rewards for actions?
Reinforcement learning.
Topic 2 of 6 in Fundamentals of gen AI
Test yourself on Fundamentals of gen AI
The full Google Cloud Generative AI Leader question bank has 512 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Google Cloud. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.