| Representation and modeling is a fundamental problem in building pattern recognition and machine cognitive systems. While classification and matching have been the most prevalent approaches, the concept of similarity has been recognized as useful in representing domains which cannot be adequately characterized through labels or classes. In this thesis, similarity is formulated as a context-driven cognitive and perceptual process which associates distinct but related patterns. It is argued that a computational realization of similarity includes the following conceptual components: (1) a method for acquiring the similarity relationships, (2) a similarity representation, and (3) a mapping from the input pattern space to the representation. These requirements are addressed through the creation of a set of conceptual and computational tools. A new similarity representation using a novel graph-based approach featuring sparse, locally connected neighborhoods is formulated. A computational realization of the mapping from the input feature space to the similarity domain is realized using a new neural network architecture and novel training algorithms. The structure of this multilayer network is derived from the graph similarity structure constructed for a specific cognitive system and context. A set of software tools for similarity modeling has been developed which includes modules for acquiring similarity relationships using interactive user interfaces, constructing and visualizing similarity graphs, and simulating and training large-scale neural networks for graph-based similarity models. A similarity representation for a three-dimensional object recognition problem was used to demonstrate the approach and methodology, and the resultant model is shown to capture important similarity properties. |
