Cheatsheet: What machine learning actually is

Rules vs learning from data

	Traditional programming	Machine learning
You provide	The rules (logic)	Examples (inputs plus answers)
Machine provides	Execution	The learned rules
Intelligence lives in	Your code	The data
Best when	Rule is simple and known	Rule is too many, fuzzy, or unknown, and you have data

The two families

Family	Has labels?	Goal	Sub-types
Supervised	Yes	Predict the answer for new inputs	Regression (number), Classification (category)
Unsupervised	No	Find structure in the data	Clustering, Dimensionality reduction
Reinforcement (out of scope)	No (uses reward)	Learn by trial and error	n/a here

Supervised: regression vs classification

	Regression	Classification
Answer type	A number	A category
Examples	House price, temperature, units sold	Spam / not-spam, which digit, fraud / legit

The two-question classification test

Step	Ask	If…
1	Do I have labeled answers?	No -> unsupervised, or not machine learning
2	Is the answer a number or a category?	Number -> regression; Category -> classification

When NOT to use machine learning

Situation	Why a rule wins
A simple known rule works	Exact, fast, fully explainable (e.g. sales tax)
No data	Nothing to learn from
Unexplainable mistake unacceptable	A statistical black box cannot be trusted

The one rule that governs everything

Claim	Verdict
”Perfect on the training data”	Proves nothing
”Holds up on new, unseen data”	The only test that counts
Model memorizes training data	Learned the noise, not the pattern

Worked classifications

Problem	Answer
Predict tomorrow’s temperature	Supervised, regression
Is this transaction fraud?	Supervised, classification
Group shoppers, no categories given	Unsupervised, clustering
Compute sales tax	Neither (write the rule)
Compress 200 survey questions to a few themes	Unsupervised, dimensionality reduction