Neural Networks

Source: NESA Software Engineering Course Specifications (page 29).

Neural networks were designed to mimic the processing inside the human brain. They consist of a series of interconnected nodes (artificial neurones). Each neurone can accept a binary input signal and potentially output another signal to connected nodes.

Training cycle

Internal weightings and threshold values for each node are determined in the initial training cycle for each neural network. The system is exposed to a series of inputs with known responses. Linear regression with backward chaining is used to iteratively determine the set of unique values required for output. Regular exposure to the training cycle results in improved accuracy and pattern matching.

Execution cycle

In the diagram, signal strength between nodes with the strongest weightings are thicker representing a higher priority in determining the final output. The execution cycle follows the training cycle and utilises the internal values developed during the training cycle to determine the output.