# Neural Network An artificial neural network is a computing architecture loosely inspired by the brain: layers of simple units ("neurons") connected by weighted links, where each neuron sums its inputs, applies a nonlinear function, and passes the result on. By adjusting the connection weights, the network learns to map inputs to outputs. A single layer can only learn simple patterns; stacking many layers ([[Deep Learning]]) lets the network learn hierarchical, abstract features. Neural networks are the foundational architecture beneath essentially all modern AI. The idea is old — it traces to the 1943 McCulloch-Pitts neuron and the 1958 Perceptron — but it spent decades in the wilderness, twice abandoned (the [[AI Winter]]s). It became dominant only when Backpropagation (a practical training method), big data, and [[Graphics Processing Unit]] compute converged around 2012. The brain analogy is real but loose: artificial neurons are vastly simpler than biological ones, and the resemblance is more inspirational than literal. ## Why it matters here The neural network is mostly foundational technical literacy — the architecture other pages refer to — with one conceptual stake worth naming: neural networks are **black boxes.** Their knowledge is distributed across millions or billions of numerical weights with no human-readable rules, so even their creators cannot fully explain why a given network produced a given output. This opacity is the technical root of the AI **explainability and accountability** problem that runs through the vault's governance concerns: when an inscrutable network decides a loan, a sentence-length recommendation, or a content-moderation call, there is often no auditable reasoning to contest — the opposite of the transparent [[Symbolic AI]] tradition. Everything downstream (the [[AI Safety]] worry that we can't verify what a system "wants," the bias problems, the regulatory push for explainability) starts from this property. ## See Also - Perceptron — the original single-layer network - Backpropagation — how networks are trained - [[Deep Learning]] — many-layered networks - Connectionism — the paradigm neural networks embody