Optimal Mapping of Neural Networks onto FPGAs

The paper shows how a feedforward neural network defined by a set of m binary examples of n bits each, can be determined by a novel constructive algorithm (which determines the number of layers, the number of neurons in each layer and the synaptic weights of a particular neural network). For doing that, the optimisation criteria of the new algorithm can be chosen from the following: (i) the area of the circuit A; (ii) the AT 2 complexity measure of VLSI; (iii) the delay T; or (iv) the maximum fan-in for a gate Δ. As a result the neural network which is build can be optimised for mapping onto a FPGA. By considering the maximum fan-in of one neuron as a parameter, we proceed to show its influence on the area, and suggest how to obtain a full class of solutions. We also compare our results with other constructive algorithms and benchmark it on the classical “two spirals problem.” Conclusions and some open problems are closing the paper.