Abstract:
This research focuses on parallel implementation of image recognition by adopting an
Optimal Mapping and Hybrid Parallel Training (HPT) scheme to train Multilayer Per ceptron (MLP) neural network with Back Propagation (BP) algorithm on High Perfor mance Cluster (HPC). Network decomposition technique is used to break down dataset
over the available processors and to train each part-network with our proposed HPT
scheme to drive out the communication overhead during parallel training phase of BP
algorithm. We used Message Passing Interface (MPI) and Posix Threads (PThreads)
for HPT of BP algorithm at distributed and shared memory architecture level respec tively. We derived the closed form expression of parallel training time, based on time
expression computed performance parameters like speedup and efficiency and com pared our results with the ones with previously reported Hybrid Partitioning (HP)
schemes. Various well known image recognition benchmarking problems have been
simulated to evaluate the proposed scheme. Analytical and experimental results show
that our proposed scheme takes less training time, is more efficient and required less
memory space than other HP schemes.
The functional verification of simulated image recognition problem has also been
done using Light Efficient Network Simulator (LENS). We devised a novel pre-processing
scheme to reduce the size of input dataset, thus reducing the size of underlying neu ral network. Each pixel at the input layer is represented with a single value, which
requires less memory space and thus is computationally more efficient. We evaluated
the training performance by changing number of neurons in input layer, with varying
number of hidden layers and by altering the learning rate.
We also collaborated with the SpiNNaker Project at the University of Manchester
to simulate our proposed technique on SpiNNaker Massively Parallel CMP System.
