Hierarchical Learning in Classification of Structured Objects

We discuss the hierarchical learning approach applied to the recognition of structured objects. Learning algorithms for such objects usually display high complexity and typically require a priori assumptions on the subject domain. Hierarchical learning is designed to alleviate many problems associated with structured object recognition. It helps steer searches for solutions toward more promising paths in the otherwise computationally prohibitive search spaces by breaking the original task into simpler, more manageable subtasks. It provides for an effective interactive mechanism to transfer the additional domain knowledge expressed by external human experts into low level operators. The design and the implementation of hierarchical learning and domain knowledge elicitation, based on approximate reasoning and rough mereology constitute an excellent example of Granular Computing at work.