Doboli, S.
Computer Science Department, Hofstra University, NY, USA
Creativity is the process of coming up with novel ideas. Among mechanisms proposed to underlie creative thought are conceptual combinations and analogies. Conceptual combinations group known concepts into a new entity with emergent meaning and properties. Studies of conceptual combinations show that concepts combine in one of two ways: property- or relation- based combinations. In this work we propose a cognitive model for studying the dynamics of property-based combinations. The model can be extended to account for relation-based combinations. The main features of the model are: multiple distributed representation of concepts, of their features and properties, dynamic categorization, emergent features and learning newly discovered conceptual combinations. It is inspired by Baralou?s perceptual symbol systems and Damasio?s convergence zones.
The model represents concepts in three layers: a feature map, property map, and conceptual map. The feature map (FM) is a simplified 2D representation of features. Activity in the FM layer represents instances of feature values with similar features encoded by nearby nodes. A concept activates multiple localized regions on the map, each representing a distinct feature. FM activity is projected onto the property layer (CP), with near-by FM nodes converging onto the same CP node. Activity in the CP layer is an attractor based distributed representation of all properties of a concept. Connections between nodes in the CP layer encode correlations between properties. The third layer - the concept layer (C) - is a distributed and localized read-out of the output of the CP layer. It is used to read-out the active concepts and to store newly discovered concepts. Connections in the model are set such that a set of concepts is learned with a varied relatedness between them. The model is tested on its ability to dynamically discover and learn novel conceptual combinations starting from a set of active features.