Parallelization of SINNs for Shift and Rotation Invariant Pattern Recognition

In this paper an algorithm is presented for the parallel implementation and data distribution of structured invariant neural networks (SINN) for invariant pattern recognition. Structured invariant neural networks have shown convincing results for the task of shift and rotation invariant pattern recognition. They perform the invariant feature extraction and adaptive classi cation simultaneously in a sparsely connected architecture with shared weights, thus leading to a fast invariant recognition of gray scale images. Additional speed up of the image processing system can be achieved by parallelization. Image capturing and preprocessing with parallel computers can lead to an enormous acceleration. Here we present a parallel implementation of the SINN that is necessary to meet real time demands also for the task of invariant pattern recognition. First we give a reordering algorithm for the highly irregular and complex data access patterns of the SINN minimizing the communication overhead. This algorithm accepts a wide class of possible input distributions determined by the implementation at the preceeding preprocessing step. The resulting data distribution allows the local computation of parts of the network architecture in parallel on a large number of processors. Our experimental results show a signi cant speed up of this parallelization compared to a sequential computation of the SINN.