ML problems are grouped by the feedback the learner gets: labels (supervised), no labels (unsupervised), a few labels (semi-supervised), labels made from the data itself (self-supervised) or rewards (reinforcement learning).

Type	Training data	Goal	Typical tasks & algorithms
Supervised	Inputs `X` with labels `y`	Predict `y` for new `X`	Classification, regression — linear/logistic regression, trees, SVM, boosting
Unsupervised	Inputs `X` only	Find structure	Clustering (k-means, DBSCAN), dimensionality reduction (PCA), anomaly detection
Semi-supervised	Few labelled + many unlabelled rows	Use unlabelled data to improve a supervised model	Self-training, label propagation
Self-supervised	Raw data; labels created from the data (the next word, a masked patch)	Learn general representations	Pre-training language and vision models
Reinforcement	States, actions, rewards from an environment	A policy maximising cumulative reward	Game playing, robotics, recommendations as sequential decisions

Supervised vs unsupervised on the same data

The iris dataset has 150 flowers of 3 species. A classifier learns from the species labels; k-means never sees them but still finds groups that largely match them:

Python

from sklearn.cluster import KMeans
from sklearn.datasets import load_iris
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import adjusted_rand_score

X, y = load_iris(return_X_y=True)

clf = LogisticRegression(max_iter=1000).fit(X, y)                 # supervised: uses y
print(clf.predict(X[[0, 50, 100]]))                               # [0 1 2] ← species 0, 1, 2

km = KMeans(n_clusters=3, n_init=10, random_state=0).fit(X)       # unsupervised: never sees y
print(km.labels_[[0, 50, 100]])                                   # [1 0 2] ← cluster ids are arbitrary names
print(round(adjusted_rand_score(y, km.labels_), 2))               # 0.73 ← 1.0 = perfect match, ≈0 = random; versicolor/virginica overlap

Supervised: classification vs regression

	Classification	Regression
Target	A category (spam / not spam; cat / dog / bird)	A number (price, temperature, demand)
Output	Class label, usually with a probability or score	A real value
Loss (training)	Log-loss (cross-entropy), hinge	Squared error, absolute error, Huber
Metrics	Accuracy, precision, recall, F1, ROC-AUC, PR-AUC	MAE, RMSE, R²
Variants	Binary, multi-class (one of k), multi-label (any subset of k)	Single- or multi-output

Other ways to slice ML

Distinction	One side	Other side
Batch vs online	Batch: train on the full dataset, retrain periodically	Online: update incrementally as data arrives (`partial_fit` in `SGDClassifier`, `MultinomialNB`)
Model-based vs instance-based	Model-based: learn parameters, then discard the data (linear models)	Instance-based: keep the training data and compare new points to it (k-NN)
Parametric vs non-parametric	Parametric: a fixed number of parameters regardless of data size (linear/logistic regression, Naive Bayes)	Non-parametric: complexity grows with the data (k-NN, decision trees, kernel SVMs)
Discriminative vs generative	Discriminative: model `P(y \| x)` or the boundary directly (logistic regression, SVM)	Generative: model how data is produced, `P(x \| y)·P(y)` (Naive Bayes, Gaussian mixtures)

Reinforcement learning in one paragraph

An agent observes a state, takes an action, receives a reward and a new state from the environment, and learns a policy (state → action) that maximises total future reward. Unlike supervised learning, nobody gives the correct action — only delayed, scalar feedback — so the agent must balance exploration and exploitation. Q-learning, MDPs and value iteration are worked through in the Artificial Intelligence notes.

Is anomaly detection supervised or unsupervised?

What is self-supervised learning and why does it matter?

Parametric vs non-parametric — give an example of each.