Parallel Implementation of Labeled Graph Matching using a Biologically Inspired Model

Although a number of Artiicial Neural Network models have been used for object recognition tasks, most do not provide desirable invariances to various transformations of the image. The Dynamic Link Architecture (DLA) provides these desired properties by representing objects as labeled graphs. Object recognition therefore amounts to a labeled graph matching task. The DLA is a biologically inspired model, and is therefore inherently parallel at a very ne grain. In this paper, we present a high level computational description of the model, and investigate the inherent amount of parallelism in this neural model. We will see how the DLA can be partitioned at the nest grain, and how this partitioning can be mapped onto a model of parallel computation. We also present an analysis of the parallel computational complexity of the implementation.