Types of Learning & Sampling • Victor Guerra

Learning paradigms

Supervised — every input has a ground-truth label; learn f(x)→y. Classification, regression, sequence prediction.
Unsupervised — no labels; find structure. Clustering (k-means), dimensionality reduction (PCA), anomaly detection. Good for pretraining / feature extraction.
Weakly supervised — noisy/limited/imprecise labels: inexact (coarse, e.g. image-level not boxes), inaccurate (errors, e.g. crowdsourced), incomplete (only some labeled).
Semi-supervised — small labeled + large unlabeled set; leverage the unlabeled data to improve generalization when labels are expensive. Methods: self-training, consistency regularization (penalize different predictions for the same input under noise → robustness), pseudo-labeling (model labels unlabeled data, trains on its confident predictions as if ground truth).
Active learning — the model selects the most informative unlabeled samples and queries an oracle (human) for labels → high performance with fewer labels. Selection strategies: uncertainty sampling (lowest confidence), margin sampling (top-2 class probs close), entropy-based (highest output entropy), query-by-committee (max disagreement across models), diversity/core-set (represent the data distribution). Used in labeling tools / human-in-the-loop.

Is a language model supervised or unsupervised? Technically unsupervised (no manual labels — it learns from raw text). Mechanistically supervised — each token becomes an (input, target) pair trained with cross-entropy. The distinction is about the source of labels, not the training mechanism. See embeddings.

Sampling & creating training data

With replacement — items can repeat, samples independent, infinite outcomes. Used in bootstrapping (bagging / random forests — ensembles).
Without replacement — each item at most once, samples dependent, limited to dataset size. Used in train/test splits (no overlap) and epochs (see pytorch-training-loop).
MCMC (Markov Chain Monte Carlo) — explore a complex, high-dimensional distribution by a guided random walk: a Markov chain (each state depends only on the previous) + Monte Carlo (random sampling to estimate integrals/expectations). Used for Bayesian inference and sampling from intractable distributions.