Machine Learning in Plain English

Spam story

You collect 5k emails labeled spam/not. Features: contains ‘free’, exclamation count, sender known?

Training = find weights that separate spam. Prediction = apply to new mail.

Three flavors

• Supervised: labeled pairs (email → spam?)
• Unsupervised: no labels (cluster customers)
• Reinforcement: trial + reward (game play, RLHF)

Golden split

Train 70% / validate 15% / test 15%. Never tune on test. If test drops while train rises → overfitting. Fix: more data, simpler model, regularization.

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB
vec = CountVectorizer()
X = vec.fit_transform(["free money now", "meeting notes"])
model = MultinomialNB().fit(X, [1, 0])