Network training. When using traditional statistical techniques such as linear or logistic regression, the P-value for each input variable can be used to judge its usefulness within a regression model, once the model has been created. The weaker non-significant variables can then be removed, with the assumption that this will make the model either less complex whilst maintaining accuracy, or more accurate due to reduced confounding noise (110). Within ANNs, no such P-values are provided. Instead, initially an untrained XXX has neutral weightings to all the pathways within it. The network will then start to run predictions or classifications with all the variables carrying equal importance. Positive or negative feedback will then propagate through the network depending on if the result was either further or nearer towards the correct outcome (back propagation). This will gradually increase or decrease the weightings each node has within the network at each nodal layer. In Error! Reference source not found. the steps towards network optimization are r epresented; (top left) Untrained XXX with standardised weightings between nodes, (top right) XXX producing correct outcome, thus leading to increased weights in those pathways, (bottom left) incorrect outcome leading to corresponding reduction or negative weights, (bottom right) trained XXX where the number of nodes, layers and relative weights have been set. Unlike traditional regression, it is not possible to give a definite significance that each variable has upon the outcome, as its weighted path through the network is not revealed. However, what can be represented is the relative weight that a variable has upon the network. During the training process, variables with low weight can be removed and the network retrained, with the eventual efficacy on the target outcome assessed. This process can be repeated until no further improvement upon the network is produced, with this technique known as the ‘embedded technique’ (109).
