Machine Learning: The Foundations

Supervised learning maps inputs to labeled outputs and accounts for the majority of deployed ML — image classification, spam detection, credit scoring, medical triage. It depends on large, accurately labeled datasets and well-defined targets.

Unsupervised learning discovers structure without labels — clustering customers, learning embeddings, compressing data. Self-supervised learning, a special case where the data supplies its own labels, is the regime that produced modern foundation models like GPT and CLIP.

Reinforcement learning optimizes behavior through reward feedback in sequential decision problems. It underpins game-playing systems (AlphaGo, AlphaZero), robotics control, recommender ranking, and the RLHF stage of frontier LLM training.

The central scientific challenge of ML is generalization — performing well on data not seen during training. A model that perfectly memorizes the training set but fails on new inputs is useless.

The bias-variance tradeoff frames this: high-bias models underfit (too simple to capture the pattern); high-variance models overfit (capturing noise as if it were signal). Regularization (L1/L2, dropout, early stopping, data augmentation) and held-out validation are the standard tools for navigating it.

• Training/validation/test split prevents optimistic estimates.
• Cross-validation is the gold standard on small datasets.
• Modern over-parameterized deep networks generalize despite memorizing — the 'double descent' phenomenon documented by Belkin et al. (2019).

Production ML systems are infrastructure, not just models. They combine data collection, labeling, feature engineering or representation learning, model selection, training, evaluation, monitoring, and continuous retraining.

Google's widely cited 'Hidden Technical Debt in Machine Learning Systems' (2015) showed that the ML model itself is typically a small fraction of total system code. Most production failures occur at the data and deployment stages — distribution shift, label drift, broken pipelines — not the model.

Classical methods — linear/logistic regression, decision trees, gradient-boosted trees (XGBoost, LightGBM), SVMs, random forests — still dominate tabular data and many enterprise problems where datasets are small and interpretability matters.

Deep learning dominates perception (vision, speech) and language because it learns representations end-to-end from raw, high-dimensional inputs. The two paradigms increasingly coexist in hybrid stacks.