Reconfigurable Hardware Implementation of Neural Networks for Humanoid Locomotion

An artificial neural network (ANN) is a parallel distribution of linear processing units arranged as layers. Parallelism, modularity and dynamic adaptation are computational characteristics associated with networks. These characteristics support FPGA implementation of networks, because parallelism takes advantage of FPGA concurrency, and modularity and dynamic adaptation benefit from network reconfiguration. The most important aspects of FPGA implementation of neural networks are: the benefits of reconfiguration, the representation of internal data and implementation issues like weight precision and transfer functions. This paper proposes a number of internal data formats that optimize the network precision and a way of implementing sigmoid transfer functions to make the most of FPGA implementation.