The units are organized into layers or other structured connections. Weights influence how signals are combined, and transformations let the network represent complex relationships. Different architectures suit different tasks, such as processing images, language or sequences of actions.
For example, a network can learn from labeled images to predict a category. Its internal parameters do not usually provide an easy human-readable explanation of every decision. Performance must be assessed using suitable new examples and the actual intended conditions.
ELI5
A neural network is a set of small mathematical steps connected together. Training changes the numbers in those steps to help it give useful answers.
For example, a network can learn from labeled pictures to tell cats from dogs. It may still make mistakes on blurry pictures or animals unlike those it saw during training.
Is a neural network a biological brain?
No. It is a computational model inspired by connected processing units, not a reproduction of a biological brain.
Are the weights the whole model?
No. Architecture, computation rules and other components also determine how a network processes information.
