In plain words
A neural network is a machine-learning model that transforms inputs through connected layers of mathematical operations. It learns by adjusting parameters that influence those transformations.
A closer look
A typical unit combines input values using weights, adds a bias, and applies an activation function. Nonlinear activations let a network represent relationships that a simple linear formula cannot. Layers can build progressively useful internal representations, although individual units do not always have an easy human interpretation.
During training, an objective measures how well outputs match the task. Backpropagation efficiently computes how parameters affect this objective, and an optimizer uses that information to update them. Networks with many layers are often called deep neural networks; “deep” refers to structure, not depth of understanding.
In practice
An image classifier converts pixel values into intermediate features and then class scores. Training adjusts the network so that useful visual patterns influence the final prediction.
A useful distinction
A neural network is inspired in part by biological ideas but is not a detailed simulation of a human brain. Terms such as “neuron” and “learning” describe technical mechanisms and do not establish consciousness.