Supervised vs Unsupervised vs Reinforcement Learning

The Three Paradigms

All of machine learning falls into three categories based on how the model learns:

Supervised Learning

Definition: The model learns from labeled examples — input/output pairs where the correct answer is provided.

Training data:
Email: "Win a free iPhone!" → Label: SPAM
Email: "Meeting at 3pm"    → Label: NOT SPAM
Email: "Claim your prize"  → Label: SPAM

Model learns: what patterns predict spam

Two types

Classification — predict a category

# Is this email spam? (yes/no)
# Is this tumor malignant? (yes/no)
# Which digit is this? (0-9)
model.predict(email) → "spam"

Regression — predict a number

# What will this house sell for?
# What will the stock price be tomorrow?
model.predict(house_features) → 450000

Real-world examples

• Email spam detection
• Medical diagnosis
• Credit scoring
• Image classification
• Price prediction

The catch

You need labeled data — someone has to manually label thousands of examples. This is expensive and time-consuming.

Unsupervised Learning

Definition: The model finds patterns in data without labels. No correct answers provided.

Training data:
Customer 1: [age=25, purchases=10, avg_spend=50]
Customer 2: [age=45, purchases=2,  avg_spend=500]
Customer 3: [age=26, purchases=8,  avg_spend=45]
...

Model discovers: there are 3 natural customer groups

Two main types

Clustering — group similar items

from sklearn.cluster import KMeans
 
# Group customers by behavior
kmeans = KMeans(n_clusters=3)
kmeans.fit(customer_data)
# Discovers: budget shoppers, premium buyers, occasional buyers

Dimensionality Reduction — compress data while preserving structure

from sklearn.decomposition import PCA
 
# Reduce 100 features to 2 for visualization
pca = PCA(n_components=2)
reduced = pca.fit_transform(data)

Real-world examples

• Customer segmentation
• Anomaly detection (fraud, network intrusion)
• Topic modeling in documents
• Recommendation systems (find similar users)
• Data compression

Reinforcement Learning

Definition: An agent learns by taking actions in an environment and receiving rewards or penalties.

Agent: AI player
Environment: Chess board
Action: Move a piece
Reward: +1 for winning, -1 for losing, 0 otherwise

Agent learns: which moves lead to winning

The exploration vs exploitation dilemma

• Exploit — do what you know works (get safe rewards)
• Explore — try new things that might work better (risk lower rewards)

Too much exploitation → stuck in local optimum Too much exploration → never converges

Real-world examples

• Game playing (AlphaGo, OpenAI Five)
• Robot locomotion
• RLHF — training ChatGPT to be helpful (human feedback as reward)
• Ad bidding optimization
• Autonomous vehicles

Comparison

Self-Supervised Learning — The Fourth Paradigm

Modern LLMs use a fourth approach: self-supervised learning.

The model creates its own labels from unlabeled data:

Text: "The cat sat on the ___"
Task: Predict the missing word → "mat"

No human labels needed — the text itself provides supervision

This is how GPT, BERT, and all modern LLMs are trained. It scales to internet-scale data without expensive labeling.

Key Takeaway

• Supervised — labeled data, predict output (classification/regression)
• Unsupervised — no labels, find hidden structure (clustering/compression)
• Reinforcement — learn by trial and error with rewards
• Self-supervised — create labels from data itself (how LLMs are trained)
• Most production ML is supervised — it's the most reliable when you have labeled data