Parallel computer structure for modeling and training of artificial of neural networks

摘要

A parallel processor structure for modelling and training artificial neuronal networks is connected to a host computer and constructed as two-dimensional matrix of simple identical processor elements. The processor elements are supplied with a command stream by a sequencer in accordance with the SIMD principle. The processor elements arranged on the diagonals of the matrix are allocated to the nodes of the neuronal network and intended for carrying out the neuronal functions. The non-diagonal processor elements handle the logic combinations between the nodes and are selected for the function of the variable synaptic weightings. The matrix firstly has a local neighbourhood networking to the four next neighbouring processors in each case. In addition, lines come from the neuronal processors, separated in x and y direction, which drive the non-diagonal synapse processors in parallel. In one direction, these lines are used for accelerating the distribution of the calculation results of the neuronal processors to the synapse processors. In the other direction, the lines are used for the accelerated distribution of correction data during the learning process. IMAGE