A data-driven implementation of back propagation learning algorithm
AliM. Alhaj, Hiroaki Terada · 2003
Data-driven computers are scalable, highly concurrent machines. They have been proposed as an alternative to the conventional von Neumann computers to allow for maximal exploitation of parallelism in large-scale computations. The authors describe a parallel implementation of the backpropagation learning algorithm on a data-driven computer using the Q-v1, a general-purpose data-driven processor. The implementation is successful as the parallelism of the neural network is explicitly expressed by the functional and asynchronous data-driven program and naturally exploited by the pipelined and scalable data-driven processors. The suitability of applying data-driven multiprocessors for efficient simulation of neural networks is demonstrated.>