Quicknet on MultiSpert: Fast Parallel Neural Network Training

TitleQuicknet on MultiSpert: Fast Parallel Neural Network Training
Publication TypeTechnical Report
Year of Publication1997
AuthorsFärber, P.
Other Numbers1111

The MultiSpert parallel system is a straight-forward extension of the Spert workstation accelerator, which is predominantly used in speech recognition research at ICSI. In order to deliver high performance for Artificial Neural Network training without requiring changes to the user interfaces, the existing Quicknet ANN library was modified to run on MultiSpert. In this report, we present the algorithms used in the parallelization of the Quicknet code and analyse their communication and computation requirements. The resulting performance model yields a better understanding of system speed-ups and potential bottlenecks. Experimental results from actual training runs validate the model and demonstrate the achieved performance levels.

Bibliographic Notes

ICSI Technical Report TR-97-047

Abbreviated Authors

P. Färber

ICSI Publication Type

Technical Report