Artificial Neuron
An artificial neuron is the basic computational unit of a neural network: it takes a vector of numeric inputs, multiplies each by a learned weight, adds the products together with a learned bias, and passes the total through a nonlinear activation function to produce a single output. Written compactly, the output is y = f(w1·x1 + w2·x2 + … + wn·xn + b), and everything a modern network does — recognizing a face, predicting the next word, steering a robot arm — is built by wiring enormous numbers of these tiny weighted-sum-and-squash operations into layers and adjusting the weights until the overall behavior is useful.
The three parts of the unit each do a distinct job. The weights set how strongly each input pushes the output up or down (a negative weight suppresses, a positive one excites); the bias shifts the point at which the unit starts to respond, so that it can fire on a weak signal or demand a strong one; and the activation function bends the result so that stacking units does something a single linear equation could not. Without the nonlinearity, any number of layers collapses into one big matrix multiplication, which is why the choice of activation is not cosmetic: the step function of early models gave way to smooth curves such as the logistic sigmoid and the hyperbolic tangent, and then, in deep networks of the 2010s, to the rectified linear unit (ReLU), which simply outputs zero for negative totals and passes positive totals through unchanged, a form that trains far more reliably in deep stacks.
The lineage starts with the 1943 McCulloch-Pitts model, but the modern unit differs from it in the one respect that matters most: it learns. Warren McCulloch and Walter Pitts, in "A Logical Calculus of the Ideas Immanent in Nervous Activity" (Bulletin of Mathematical Biophysics, 1943), described an idealized neuron that is either on or off, sums excitatory inputs, is vetoed by any inhibitory input, and fires if a fixed threshold is reached; wired together, such units could compute logical propositions, but every connection was set by hand. Frank Rosenblatt's perceptron (1958) made the weights adjustable by an error-correcting rule, Bernard Widrow and Marcian Hoff's ADALINE (1960) trained its weights by gradient descent on squared error, and the 1986 backpropagation paper by David Rumelhart, Geoffrey Hinton and Ronald Williams showed how to compute gradients through many layers of units with smooth activations, which is the step that turned the neuron from a logic element into a trainable function approximator.
A worked example shows what one unit can and cannot do. With two inputs, weights of 1 and 1, a bias of −1.5 and a hard step activation, the unit outputs 1 only when both inputs are 1, so it computes logical AND; changing the bias to −0.5 turns it into OR. Geometrically, a single neuron draws one straight line (a hyperplane in higher dimensions) through its input space and classifies points by which side they fall on, so XOR, whose true cases cannot be separated from its false cases by any single line, is out of reach for one unit but easy for two layers of them. This is the reason depth matters, and the universal approximation results of George Cybenko (1989) and Kurt Hornik and colleagues (1989) formalized the converse: a network with even one hidden layer of sigmoidal units can approximate any continuous function on a bounded region to arbitrary accuracy, given enough units.
The word neuron is a metaphor, and treating it as a biological claim is the most common misunderstanding. A real cortical neuron has branching dendrites that perform local computation, spikes at specific moments, many types of synapse with different timescales, and chemical modulation; the artificial unit keeps one abstraction — many weighted inputs, one output, strength of connection as the carrier of learning — and discards the rest. Backpropagation itself has no accepted counterpart in the brain. The metaphor was productive for guiding early designs, but modern practitioners tend to see the unit as a parameterized differentiable function, and a deep network as a function assembled from them, rather than as a simulation of tissue.
Today the artificial neuron is largely invisible as a separate object because it has been absorbed into matrix operations. A layer of a thousand neurons is one matrix multiplication followed by an elementwise activation, which is exactly the operation GPUs and tensor accelerators are built to do at scale; the feed-forward blocks inside a Transformer language model are two such layers with a smooth activation such as GELU between them, repeated dozens of times. Convolutional units share one set of weights across positions in an image, recurrent units feed their output back in, and gated units (LSTM, GRU) add multiplicative switches, but each remains the same weighted sum, bias and nonlinearity at heart. Pitfalls follow from the simplicity: saturating activations such as the sigmoid produce vanishing gradients in deep stacks, ReLU units can die if their inputs stay negative, and a single neuron's weights are rarely interpretable on their own, since meaning in a trained network tends to be spread across many units at once.
Read artificial neuron as the smallest learnable decision in machine learning: a weighted vote, shifted by a bias and bent by a nonlinearity, whose power comes not from any single unit's cleverness but from the way the weights can be tuned and the units stacked.