Supervised, Unsupervised, and Reinforcement Learning for the Generative AI Leader Exam

One of the foundational topics on the Generative AI Leader exam is the set of learning paradigms that machine learning models fall under. There are three you need to know: supervised learning, unsupervised learning, and reinforcement learning. Each one represents a different relationship between the model and the data it learns from, and each one is suited to different kinds of problems. In this post I will walk through what each paradigm is, what it is used for, and the model families and use cases that show up under each category.

Supervised Learning

Supervised learning is a category of machine learning where the model learns how to map inputs to outputs using labeled data. During training, each input is paired with the correct answer, and the model gradually learns the relationship between them. Most supervised learning tasks fall into one of two groups. Classification, where the goal is to predict a category. Regression, where the goal is to predict a continuous value.

Image classification is a classic example. The model is given many labeled images of cats and dogs, and it learns to predict whether a new image shows a cat or a dog. The label is what makes the learning supervised. The model is generalizing from labeled examples.

There are several model families commonly used in supervised learning:

• Linear models such as linear regression and logistic regression are foundational supervised learning models. They work well when the output variable is continuous.
• Tree-based models like decision trees and ensembles such as random forests or gradient boosted trees split the input space based on decision rules. These are useful for capturing nonlinear relationships and tend to work well on tabular data.
• Support vector machines are margin-based classifiers. They work by finding the optimal boundary that separates classes in the feature space, and can be extended with kernels to handle more complex cases.
• K-nearest neighbors takes a different approach. It classifies a new example by looking at the labels of its closest neighbors in the training set.

You do not need to memorize each of these model families for the Generative AI Leader exam. The takeaway is that supervised learning maps input labels to outputs, and the label is then predicted for new data.

Unsupervised Learning

In supervised learning, labeled data provides a clear objective. The model is trained to approximate a known mapping from inputs to outputs. But not all datasets come with that structure. In many real-world scenarios you are given inputs without labels, and the task becomes discovering the structure that underlies the data itself. That is the world of unsupervised learning.

Unsupervised learning refers to a class of techniques where the goal is to identify patterns, groupings, or structure within unlabeled data. The model is not guided by known outcomes. Instead, it has to make sense of the data based solely on the relationships between the inputs.

There are three core applications you should be familiar with:

• Clustering attempts to group data points based on similarity. Techniques like K-means or hierarchical clustering operate by evaluating distances between points and organizing them into coherent groups. These methods are frequently used in exploratory data analysis, recommendation systems, or market segmentation, where understanding the structure of a dataset is often more valuable than making predictions.
• Anomaly detection identifies outliers, data points that differ substantially from the norm. This is essential in domains where rare events are important to know about, such as fraud detection or system failure monitoring.
• Principal component analysis, or PCA, is concerned with reducing dimensionality rather than predicting labels. It finds directions, or components, along which the data varies most, and uses these to project the data into a smaller number of dimensions. This allows for more efficient storage, visualization, and often improves the performance of downstream models by eliminating noise.

Where supervised learning requires labeled data to train predictive models, unsupervised learning is focused on discovering latent structure. It often serves as a foundation for understanding complex datasets in the absence of predefined targets.

Reinforcement Learning

Reinforcement learning, often abbreviated RL, is a different kind of learning approach. Unlike supervised learning, which learns from labeled examples, or unsupervised learning, which detects patterns inherent in the data, reinforcement learning is based on interaction. It uses trial and error based on a reward function.

Reinforcement learning shows up in a few common scenarios: robotics, AIs playing games, recommendation systems, and some generative AI use cases. It is particularly useful in situations where explicit training data is hard to come by, or where outcomes depend on a sequence of decisions.

The basic feedback loop looks like this. There is an agent or AI, and there is an environment. The agent takes an action. The environment responds. The agent receives a reward or a penalty. That reward signal then guides the agent's learning process.

Each action results in some form of feedback from the environment. A reward if the action moves the agent toward its goal, a penalty if it does not. The agent uses this feedback to gradually improve its behavior over time. It is similar to how humans or animals learn by trial and error, exploring, making mistakes, and adjusting based on what works.

How to Think About These for the Exam

The Generative AI Leader exam will not ask you to implement any of these algorithms, but it does expect you to recognize which paradigm matches a given scenario. The mental model I find most useful is to ask what the data looks like and what kind of feedback is available.

• If the data has labels and the task is to predict those labels for new examples, that is supervised learning.
• If the data has no labels and the goal is to find structure or group similar items, that is unsupervised learning.
• If learning happens through interaction with an environment that returns rewards or penalties, that is reinforcement learning.

Once you have that framing locked in, the model families and use cases under each paradigm become much easier to keep straight.

My Generative AI Leader course covers learning paradigms in depth alongside the rest of the foundational material you need to pass the exam.