Characterization of artificial neural network algorithms

Tradeoffs must be made when artificial neural network models are implemented efficiently. One popular artificial neural network model, the back-propagation algorithm, promises to be a powerful and flexible learning model. The effects on its performance when the model is modified for efficient hardware implementation are discussed. The modifications examined concern limited precision architectures, sign/threshold propagation, sum weight changes, and the addition of noise. It is found that reduced precision computation can be used successfully for the back-propagation algorithm, the communication between processors can be reduced when propagating the weights, accumulating the weight changes can improve the execution time of the algorithm, and noise can have a positive effect on the learning algorithm.<
>