Towards an event-space self-configurable neural network

An approach to neural network design based on processing and representing data at the event-space domain is presented. Analysis is then extended to processing structured objects. Decision functions based on nominal-valued structured objects can be evaluated at the event level, and parameters (e.g., weights) can be estimated analytically or adaptively. A type of this network on continuous-valued structured objects is presented based on partitioning of the outcome space such that each partitioned subspace corresponds to an input node in the network. The partitioning uses the maximum entropy criterion, applied iteratively on selected subspaces to produce a hierarchical partitioning of the outcome space. The criterion for the selection of subspaces can be based on class-entropy value such that subspaces with high degrees of class discrimination power are generated. The approach is illustrated in data analysis and image analysis problems