Hybrid pattern recognition system capable of self-modification

Systems capable of recognizing and learning patterns in two-dimensions can be used in imaging systems and robotic perception systems. The symbolic and neuromorphic methods for pattern processing problems of this type are complementary in character. We present a hybrid system that utilizes components of symbolic and neuromorphic type; we employ two hybrid components that simultaneously operate up on the same data to produce hypotheses about the data. To resolve the potential conflicts in these hypotheses, we propose a method that learns a combination rule based on a set of examples. We employ the method of empirical risk minimization and does not require knowledge about the error probability distributions of the modules. We are building a prototype system to recognize control panels using a vision system.