Investigating The Effect Of Data Complexity On Artificial Neural Network Architecture And Performance

Abstract

Within pattern recognition, Artificial Neural Networks (ANNs) are very powerful tools used to classify data according to patterns which may or may not be apparent within the data itself. This ability to link pieces of data allows for trained networks to have predictive abilities based upon new data, which can allow for advance warning of potential hazards, such as the appearance of tumorous growths within the body. However, this power comes at a cost of the design and architecture of the network itself, since the capabilities are directly linked to the amount of processing power (the number of neurons) available. Too few neurons within the network, and it will be unable to find a suitable pattern. Too many neurons, and the network will over-fit the data, effectively memorising the training data and being unable to generalise the pattern. This project explored the effect of the complexity of data on the architecture of the network used to predict it, with a view to minimise training time on real data (which can take additional time to learn) by creating architectures on simplified, simulation data. A link was established between input complexity and architecture, assuming the format of the data remains unchanged, however while the link exists, it does not affect the architecture greatly.