# Machine Learning Machine learning (ML) is the dominant paradigm of modern [[Artificial Intelligence]]: instead of a programmer writing explicit rules for every case (the [[Symbolic AI]] approach), the system *learns* patterns from data and improves its performance with experience. You give the algorithm examples, it finds the statistical regularities, and it generalizes to new cases. This shift — from hand-coded rules to learned patterns — is the single most important idea in the field's modern history. ML is conventionally divided by how the system learns: - **Supervised learning** — trained on labeled examples (images tagged "cat"/"dog," emails tagged "spam"/"not spam"); learns to predict the label for new inputs. The workhorse of practical AI. - **Unsupervised learning** — finds structure in unlabeled data (clustering customers, compressing data); no "right answers" provided. - **Reinforcement learning** — learns by trial and error against a reward signal; covered in Reinforcement Learning. The most powerful ML today uses [[Deep Learning]] — many-layered [[Neural Network]]s — trained on enormous datasets, but ML also includes simpler, widely-used methods (decision trees, support vector machines, linear regression) that remain better suited to small or structured data. ## Why it matters here Machine learning is the substrate beneath most of the vault's AI concerns, and two of its properties have direct political stakes. **It learns society's patterns, including its biases.** Because ML extracts patterns from historical data, it inherits and can amplify the biases in that data — the core of [[Algorithmic Bias]]. A hiring or lending or policing model trained on biased history will reproduce that history while wearing a veneer of mathematical objectivity ("the algorithm decided"). This is the mechanism behind documented harms in predictive policing, credit scoring, and résumé screening, and it cuts across partisan framings: the left emphasizes disparate impact, while some on the right object to models being deliberately re-weighted to correct it (the "woke AI" critique). **It is opaque.** Many ML models, especially deep ones, are "black boxes" whose decisions cannot be fully explained — a serious problem when they are used for consequential decisions about people (the explainability and accountability debates in AI governance). [[Neurosymbolic AI]] is the main architectural response: pair the learned model with a symbolic component whose reasoning can actually be inspected. ## See Also - [[Deep Learning]] — the dominant ML technique today - [[Symbolic AI]] — the rule-based paradigm ML displaced - Reinforcement Learning — the trial-and-error branch - [[Algorithmic Bias]] — what ML inherits from its training data - [[Neural Network]] — the architecture behind modern ML - [[Neurosymbolic AI]] — the hybrid that tries to keep ML’s power and recover symbolic transparency