What is a neural network and what does a single artificial neuron compute?

A neural network is a function approximator built from layers of connected neurons. A single neuron computes a weighted sum of its inputs plus a bias, then applies a non-linear activation: output = f(w·x + b). Stacking neurons into layers lets the network learn hierarchical representations, where early layers capture simple features (edges, tokens) and deeper layers combine them into complex concepts. A network with at least one hidden layer and non-linear activations is a universal approximator, meaning it can approximate any continuous function given enough units. Training adjusts the weights and biases via gradient descent to minimize a loss. The perceptron is the earliest single-neuron model with a step activation; modern networks replace the step with differentiable activations so gradients can be computed.