An artificial neural network (ANN) is a machine learning model made of simple connected units, loosely modeled on neurons, arranged in layers, where each unit combines its weighted inputs and passes the result through a nonlinear function, and the weights are learned from training data.
A single unit with a sigmoid activation computes the same function as logistic regression, so a neural network can be seen as many such units stacked in layers. Networks are usually trained with backpropagation, which adjusts the weights to reduce the difference between the network’s output and the desired output. Large language models are very large neural networks.
History
McCulloch and Pitts described networks of idealized neurons as a logical calculus in 1943. Rumelhart, Hinton, and Williams described learning by back-propagating errors in multilayer networks in Nature in 1986.
Sources
- A Logical Calculus of the Ideas Immanent in Nervous Activity, McCulloch and Pitts, Bulletin of Mathematical Biophysics 5(4), 1943. Retrieved October 7, 2026.
- Learning Representations by Back-Propagating Errors, Rumelhart, Hinton, and Williams, Nature 323, 1986. Retrieved October 7, 2026.
- Speech and Language Processing, Chapter 6, Neural Networks, Jurafsky and Martin, Stanford University. Retrieved October 7, 2026.