A Nodal Growth Algorithm for Concept Discovery

This working paper proposes a new type of algorithm for the discovery of invariant, higher-order concepts in variable, noisy data. The algorithm is novel in that it is based on the unlimited addition of nodes and links to a recurrent connectionist network. Nodes are created to represent patterns of co-activation of existing nodes that are not sufficiently accounted for by existing nodes. This allows the system to produce novel concepts or degrees of freedom that capture the complex correlations over time and/or space of the activations of different sensory elements. The strengths of the links connecting all the nodes in the network (newly created as well as initially given) are determined according to a Hebbian or Delta learning algorithm, thus guaranteeing a network architecture that constantly adapts to its experiences. The paper sketches various potential applications, e.g. in text comprehension, clustering of scientific papers, and situated learning, that can be used to test and optimize the proposed algorithm.