This paper puts forward the need for neural evolution schemes that reduce the volumes of synchronization and communication required by current neural models in order to obtain efficient implementations in the parallel machines and networks of computers which are available today. In this respect, a parallel implementation of a modified Boltzmann machine is considered as an example. The neurons of the machine are distributed among the processors of the multicomputer, which asynchronously computes the evolution of their subset of neurons. In this evolution, the processors use values which may or may not be updated for the neuron states, and furthermore they do not have to wait for these values to come from other processors, thus reducing the communication requirements. Nevertheless, this lack of coherence between the neuron states in different processors is corrected by the way the processors update them with the information coming from other processors.
展开▼
