Parallel Grid-aware Library for Neural Networks Tr.. (PaGaLiNNeT)
Parallel Grid-aware Library for Neural Networks Training
Start date: Apr 1, 2009,
End date: Mar 31, 2011
The proposed research is focused on the software library development for parallel neural networks training on computational Grids. The main scientific reason of the proposed research is to develop enhanced parallel neural network training algorithms which provide better parallelization efficiency on heterogeneous computational Grids in the contrast to existing algorithms. The objectives of the proposed research are: 1. to adapt the computational cost model of parallel neural network training algorithms within single pattern, batch pattern and modular approaches to heterogeneous computational Grid resources of host institution; 2. to develop enhanced single pattern and batch pattern parallel neural network training algorithms based on improved communication and barrier functions; 3. to develop a method of automatic matching of parallelization strategy to architecture of appropriate parallel computing system; 4. to develop parallel Grid-aware library for neural networks training capable to use heterogeneous computational resources; 5. to test experimentally parallel Grid-aware library for neural networks training on heterogeneous computational Grid system of host institution within the tasks of one of its active projects; 6. to deploy parallel Grid-aware library for neural networks training on the computational Grid of return host; 7. to test experimentally parallel Grid-aware library on computational systems of both host institution and return host. The cost models of the algorithms will be developed using computational complexity approaches, improved barrier and reducing function will be adapted to neural network parallelization schemes, optimization strategies will be used to find best matching “architecture of parallel system – neural network parallelization scheme”, software library will be implemented on C programming language and MPI parallelization, the efficiency of parallel algorithm will be assessed in comparison with sequential implementation.
Get Access to the 1st Network for European Cooperation