Neural network

A neural network is built from layers of very simple units. Each unit takes numbers in, multiplies them by weights, adds the results, applies a simple function, and passes the answer on. One unit does almost nothing. Stacked into layers and given enough data, the arrangement can represent relationships far too complicated for anyone to write down as rules. Training means adjusting those weights repeatedly and by tiny amounts until the outputs match the examples, which is what backpropagation does.

Depth is what makes them powerful. Each layer builds on the representations of the layer below it, so the network composes simple detected features into complex concepts without being told the intermediate steps. Almost everything currently described as artificial intelligence in public discussion, large language models included, is some arrangement of this idea. The convolutional neural network is the variant built for images.

The cost is a cliff in explainability, and it is steeper than people expect. The knowledge in a trained network is distributed across millions or billions of weights, none of which means anything on its own. There is no path to print, as there is with a decision tree, and no coefficient to inspect, as there is with a linear regression. You can observe that the network produced an answer and measure how often it is right, but the reason for any individual answer is not available in a form a person or an auditor can read. Techniques exist to approximate an explanation after the fact, and they remain approximations. Where a decision has to be justified rather than merely made, that constraint belongs at the top of the model selection conversation, not the end of it.