Implementation of a digital radial basis function neural net

Abstract

The main idea behind neural networks is the creation of a system capable of executing complex functions, such as recognition, as similar as possible to the human brain. Neural networks are becoming more popular in today's technologies as problems are becoming more complex and high computation speeds are required. Radial Basis Function (RBF) neural network is a special type of neural network, embedded in a three layer network including the input layer, one hidden layer and an output layer. Each neuron in the hidden layer employs a radial activated function and the output neurons implement a weighted sum of the hidden neurons. RBF neural networks have been successfully applied to a large variety of applications ranging from consumer electronics like automobiles and "intelligent" home appliances to classification systems like speech and face recognition and many other applications. In this final year project the study, design and implementation of a RBF neural network were carried out. The application to which the RBF neural network was designed is handwritten digit recognition. Two implementations were used to implement the RBF neural network. The first implementation was in software, where the algorithm was tried and tested and the second implementation was in hardware, where the algorithm was tested for speed and area. The testing for the recognition rate was performed in software for three different sets of hidden layer neurons' centres set by random clustering, using the Self Organising Map clustering algorithm using the Euclidean distance measure and the Manhattan distance measure. The highest recognition rates obtained were 51.67%, 68.8% and 57.48% respectively. Afterwards the RBF neural network was implemented in hardware. Simulations and synthesis reports show that the hardware implementation executes as required. A problem was encountered when the codes were synthesized, because of the limited hardware resources available on the FPGA. For this reason, a simplified version of the RBF architecture was implemented using only 6 hidden layer neurons rather than 20.