next up previous
Next: Knowledge Eliciting from Neural Up: Parallel Neural Networks Previous: More tests for parallelism

Conclusions

 

I have shown in this part that shifting towards parallelization is a natural process that should be followed by ANN researchers, because the human brain, that ANNs try to simulate, also makes good use of massive parallelism. I have furthermore presented the advantages that can be obtained by this parallelization and showed that a number of researchers already work towards this goal. In this context I have discussed the main tendencies among parallelization techniques for ANN and detailed the types of parallelizations used. I have also briefly introduced some difficulties that appear when parallelizing ANNs and also indicated some ways of avoiding them.

Furthermore, I described an ideal parallelization and presented the solution of ANN mapping over a general parallel architecture, for generalized parallel computers, and then for the PVM environment and for the UNIX environment (simulation).

For the UNIX simulation, a low level implementation tool for ANN was built, based on parallel design and on the UNIX system shared resources. It was shown how the parallel ANN features can be integrated into the UNIX environment. Also, the advantages of simulated parallelism over sequential programming (one process programming), even in the shared-time UNIX system, were explained.



Alexandra Cristea
Tue Feb 9 20:20:27 JST 1999