Foundations Of Machine Learning
The Three Paradigms
Every ML technique in this entire curriculum falls into one of three paradigms, distinguished by what kind of feedback the learning process receives.
Jr Codex ML Notes
Level: Beginner Prerequisites: Chapter 1 Time to complete: ~25 minutes
Table of Contents
- The Map of Machine Learning
- Supervised Learning
- Supervised Learning's Two Branches: Regression & Classification
- Unsupervised Learning
- Reinforcement Learning
- A Side-by-Side Comparison
- How This Curriculum Is Organized Around This Map
- Summary & Next Steps
1. The Map of Machine Learning
Every ML technique in this entire curriculum falls into one of three paradigms, distinguished by what kind of feedback the learning process receives.
The Three Paradigms, At a Glance
─────────────────────────────────────────
Supervised Learning: learns from LABELED examples
("here's the input, here's the
CORRECT answer")
Unsupervised Learning: finds PATTERNS in UNLABELED data
("here's the input — find
structure yourself, no answer
key given")
Reinforcement Learning: learns from REWARDS resulting
from ACTIONS ("try things,
get feedback on how good the
outcome was, no direct
'correct answer' given")
─────────────────────────────────────────
Visual Map of This Entire Curriculum
─────────────────────────────────────────
Machine Learning
/ | \
Supervised Unsupervised Reinforcement
/ \ / \ |
Regression Classif. Cluster. DimRed. RL (Module 8)
(Module 3) (Module 4) (Module 6's two halves)
│
Ensemble Methods
(Module 5 — techniques
that BOOST classification
AND regression algorithms)
─────────────────────────────────────────
2. Supervised Learning
Supervised learning trains a model on data where each example already has the "correct answer" (a label) attached — the model learns to map inputs to these known outputs, so it can predict labels for new, unlabeled inputs.
# Supervised learning ALWAYS has this shape: (input, correct_output) pairs
# Example: predicting house prices
training_data = [
({"sqft": 1200, "bedrooms": 2}, 250000), # input features → KNOWN correct price
({"sqft": 1800, "bedrooms": 3}, 340000),
({"sqft": 2400, "bedrooms": 4}, 410000),
]
# The model LEARNS the relationship between features and price,
# then can PREDICT the price for a brand new house it's never seen:
# model.predict({"sqft": 2000, "bedrooms": 3}) → some predicted priceTerminology You'll See Constantly From Here On
─────────────────────────────────────────
Features (X): the input variables (sqft, bedrooms)
Label/Target (y): the correct output being predicted (price)
Training set: the (features, label) pairs used to
teach the model
Test set: held-out (features, label) pairs used
to HONESTLY evaluate the trained
model (Chapter 5 covers this properly)
─────────────────────────────────────────
3. Supervised Learning's Two Branches: Regression & Classification
Supervised learning splits further, based on what kind of value is being predicted:
Regression Classification
───────────────────────────── ─────────────────────────────
Predicts a CONTINUOUS NUMBER Predicts a CATEGORY/CLASS
(Module 3) (Module 4)
"What will this house SELL FOR?" "Is this email SPAM or NOT?"
→ $340,250.50 (any real number) → "spam" (one of a fixed
set of categories)
"How many units will we SELL "Which of these 3 species
next month?" is this flower?"
─────────────────────────────────────────
# Regression: output is a NUMBER, on a continuous scale
def predict_house_price(features) -> float:
... # e.g. returns 340250.50
# Classification: output is a CATEGORY, from a fixed, discrete set
def predict_email_category(email) -> str:
... # e.g. returns "spam" or "not_spam"A quick test to tell them apart: if the answer could sensibly be "342.7" or "$1,204,551.23," it's regression. If the answer must be one of a fixed set of labels ("cat"/"dog"/"bird," or "yes"/"no"), it's classification. This distinction determines which algorithms apply (Modules 3 vs 4) and which evaluation metrics make sense (very different metrics, covered in each module's final chapter).
4. Unsupervised Learning
Unsupervised learning works with data that has no labels at all — no "correct answer" is provided. Instead, the goal is to discover inherent structure, patterns, or groupings within the data itself.
# Unsupervised learning has ONLY inputs — no correct-answer labels at all
customer_data = [
{"age": 25, "annual_spend": 1200},
{"age": 52, "annual_spend": 8500},
{"age": 31, "annual_spend": 1350},
{"age": 48, "annual_spend": 7900},
# ... thousands more, with NO pre-existing "correct group" label
]
# The algorithm might discover, e.g., that these customers naturally
# form TWO groups (young/low-spend, older/high-spend) — a pattern
# NO ONE explicitly labeled in advance (Module 6 covers this: clustering)Two Main Unsupervised Tasks (Module 6)
─────────────────────────────────────────
Clustering: group similar data points together
(e.g. customer segments, Module 6,
Ch.1-2)
Dimensionality Reduction: compress many features into fewer,
while preserving the important
structure (e.g. visualizing
100-dimensional data in 2D,
Module 6, Ch.3-4)
─────────────────────────────────────────
The absence of labels is the key distinguishing feature: there's no "correct" clustering to check your answer against the way there's a correct house price — success is judged by how useful or coherent the discovered structure turns out to be, a genuinely different evaluation philosophy from supervised learning (revisited in Module 6).
5. Reinforcement Learning
Reinforcement learning (RL) is covered in depth in the AI Notes' dedicated RL module — this curriculum's Module 8 builds directly on that foundation rather than re-deriving it. The core idea, briefly:
RL's Distinct Feedback Signal
─────────────────────────────────────────
Not labeled examples (supervised) — no one tells the agent
"the correct action here was X"
Not just unlabeled data (unsupervised) — the agent takes
ACTIONS in an ENVIRONMENT and receives REWARDS
An agent LEARNS, through trial and error, which actions lead
to good outcomes over time — exactly the framework built in
the AI Notes' Module 5 (agent types, Markov Decision
Processes, Q-learning)
─────────────────────────────────────────
# RL's shape: (state, action, reward, next_state) — NOT (input, correct_label)
# An agent playing a game:
# state = current game board
# action = a move it tries
# reward = +1 for winning, -1 for losing, 0 otherwise
# → learns, over MANY games, which actions tend to lead to rewardModule 8 of this curriculum picks up exactly where the AI Notes' Q-learning chapter left off — covering policy gradient methods, actor-critic algorithms, and the bridge into deep reinforcement learning (which uses neural networks, covered in the Deep Learning Notes, to handle RL problems too large for the tabular methods in the AI Notes).
6. A Side-by-Side Comparison
| Supervised | Unsupervised | Reinforcement | |
|---|---|---|---|
| Feedback | Labeled correct answers | No labels at all | Rewards from actions |
| Goal | Predict a label for new inputs | Discover structure/patterns | Learn a strategy (policy) maximizing reward |
| Example Task | Predicting house prices, spam detection | Customer segmentation, data compression | Game-playing, robotics control |
| This Curriculum | Modules 3-5 | Module 6 | Module 8 (builds on AI Notes) |
7. How This Curriculum Is Organized Around This Map
Following This Chapter's Map, Module by Module
─────────────────────────────────────────
Module 2: Data Preparation — needed for ALL three paradigms
Module 3: Supervised Learning — REGRESSION
Module 4: Supervised Learning — CLASSIFICATION
Module 5: Ensemble Learning — techniques that BOOST
both regression and classification models
Module 6: Unsupervised Learning — clustering &
dimensionality reduction
Module 7: Model Selection & Tuning — applies
across supervised techniques
Module 8: Reinforcement Learning — the third
paradigm, building on the AI Notes
Module 9: ML in Production — deploying
whatever paradigm you've built
─────────────────────────────────────────
Every module from here forward fits cleanly into this three-paradigm map — when you encounter a new algorithm later in this curriculum, the first useful question is always "which of these three paradigms does this belong to, and why?"
8. Summary & Next Steps
Key Takeaways
- Machine learning splits into three paradigms based on the type of feedback available: supervised (labeled examples), unsupervised (no labels, find structure), and reinforcement (rewards from actions).
- Supervised learning further splits into regression (predicting a continuous number) and classification (predicting a category from a fixed set).
- Unsupervised learning's defining trait is the absence of a "correct answer" to check against — success is judged by usefulness or coherence of discovered structure, not accuracy against a known label.
- Reinforcement learning is covered in depth in the AI Notes; this curriculum's Module 8 extends that foundation with policy gradients and a bridge to deep RL.
Concept Check
- A model predicting "will this customer churn: yes/no" — is this regression or classification, and why?
- Why can't unsupervised learning be evaluated the same way as supervised learning (checking predictions against known correct labels)?
- What's the key difference in the feedback signal between supervised learning and reinforcement learning?
Next Chapter
→ Chapter 3: The Machine Learning Workflow
Jr Codex — 1-on-1 Personalized Coaching | Back to Module Index