Definition
A neural network is a computational model built from connected units that transform inputs using learned parameters. A unit typically combines incoming values with weights and a bias, then applies an activation or other operation. Composing many such transformations lets the network represent relationships that a single linear step cannot.
Most modern networks learn through optimization. Training measures a loss, computes how parameter changes would affect that loss, and updates the parameters. The architecture determines which connections and operations are available. The learned parameters determine the behavior produced by a particular trained instance.
"Network" does not require one stack of identical layers. Convolutional, recurrent, transformer, and graph neural networks connect and reuse computations in different ways. Some networks are shallow, while deep networks contain many successive transformations.
Historical foundation
In 1943, Warren McCulloch and Walter Pitts published a mathematical treatment of networks made from simplified all-or-none neurons. They showed that these networks could express logical relationships. Their paper is a foundation of neural-network history, but it did not contain modern gradient training, large datasets, or today's architectures. Neural networks developed through many later contributions, so the modern family should not be credited to one paper alone.
Distinguish it from nearby terms
- A neuron in biology is a living cell. An artificial unit is a mathematical abstraction inspired by selected properties, not a faithful simulation of a brain cell.
- Deep learning usually means training neural networks with many successive learned layers. A neural network can be shallow.
- A model is a broader category. Decision trees and linear regressions are models but not neural networks.
- Parameters are the learned values inside a trained network. The architecture is the design that organizes them.
Operational significance
The architecture name does not establish quality. Data, objective, optimization, compute, evaluation, and deployment conditions all matter. Neural networks can interpolate complex patterns and still fail sharply outside their training distribution. Treat biological metaphors as intuition, not evidence that the model thinks or understands as a person does.
Check your understanding
Two teams use the same network architecture but train on different data. They have the same design, not the same model behavior, because training produced different parameters.