Machine learning splits into three main paradigms defined by the feedback a model gets. Supervised learning trains on labeled examples to predict a…
Frame the three paradigms by the feedback each one learns from.
The cleanest way to tell the paradigms apart is to ask what feedback the model receives. Supervised learning is given the right answers (labels) and learns to reproduce them. Unsupervised learning is given no answers and must find structure on its own. Reinforcement learning is given delayed rewards for the actions it takes.
Everything else — the algorithms, the metrics — follows from that difference in feedback.
See what each of the two data-driven paradigms does with your data.
Supervised learning trains on examples paired with correct labels, then predicts the label for new inputs. Two flavors: classification predicts a category (spam or not, which digit), and regression predicts a number (house price, temperature).
It is the most common paradigm in practice because many valuable problems come with labels — past outcomes, tagged data, human annotations. Its cost is exactly that: you need labeled data, which can be expensive to gather.
Unsupervised learning works on data with no labels and discovers patterns within it. Clustering groups similar items (customer segments); dimensionality reduction compresses many features into a few while keeping the important variation; anomaly detection flags points that don't fit.
It shines when you have lots of data but no labels, and want to explore, segment, or compress it. The catch is evaluation — with no ground truth, judging whether the structure it found is 'right' takes domain judgment.
Understand learning by trial, error, and reward.
Reinforcement learning has no dataset of answers. An agent observes a situation, takes an action, and receives a reward (or penalty) from the environment; over many trials it learns a policy — a strategy for choosing actions — that maximizes total reward over time.
The hard parts are that rewards can be delayed (a move's value only shows up much later) and that the agent must balance exploring new actions against exploiting what already works. It powers game-playing systems and robotics, and a variant (RLHF) helps align large language models to human preferences.
Match a problem to the right paradigm and avoid the classic mismatch.
The rule follows the feedback. Do you have labeled examples of the answer you want to predict? Use supervised learning. Do you have lots of unlabeled data you want to explore, group, or compress? Use unsupervised learning. Are you optimizing a sequence of decisions where you can score outcomes but not label the 'right' move? Use reinforcement learning.
Many real systems combine them — for instance, unsupervised pretraining followed by supervised fine-tuning.
A frequent mistake is forcing a paradigm onto the wrong data: trying supervised learning without real labels (so the model learns noise), or expecting unsupervised clusters to match a specific business category they were never told about. Match the paradigm to the feedback you actually have, not the one you wish you had.
Machine learning has three paradigms defined by feedback. Supervised learning predicts known labels via classification or regression and needs labeled data. Unsupervised learning finds structure — clusters, compressed representations, anomalies — in unlabeled data. Reinforcement learning improves a policy from rewards through trial and error. Choose by asking what feedback your data provides, and remember real systems often combine paradigms.
Take three problems — flagging fraudulent transactions, segmenting customers with no predefined groups, and teaching a robot arm to stack blocks. Assign each to a paradigm and justify your choice by the feedback available in each case.
What distinguishes the three main machine learning paradigms?
Each paradigm is defined by its feedback signal, which determines the algorithms and evaluation that follow.
What does supervised learning need that unsupervised learning does not?
Supervised learning trains on labels to predict them on new data; unsupervised learning finds structure in unlabeled data instead.
How does reinforcement learning learn?
Reinforcement learning uses trial and error with a reward signal, balancing exploration and exploitation to learn a good policy.
How should you choose a paradigm for a problem?
Start from the feedback your data actually provides; forcing a paradigm onto mismatched data is a common failure.