Temporal Sequencing via Supertemplates
暂无分享,去创建一个
A category-theoretic account of neural network semantics has been used to characterize incremental concept representation in neural memory. It involves a category of concepts and concept morphisms together with categories of objects and morphisms representing the activity in connectionist structures at different stages of weight adaptation. Colimits express the more specialized concepts as combinations of abstract concepts along shared subconcept relationships specified in diagrams. This provides a mathematical model of concept blending, in which designated relationships among concepts are preserved in a combination. Structure-preserving mappings called functors from the concept to neural categories provide a mathematical model of incremental concept representation through stages of adaptation. The work reported here extends these ideas to express temporal sequences of events, such as episodic memories. This requires an extended notion of neural morphism and a design principle for diagrams involving concepts in a temporal sequence. This is tested in a new architecture that involves a notion of supertemplates, which are ART network templates extending over a multi-level ART hierarchy with an interposed temporal integrator network.
[1] Peter Zimmer,et al. Comparison of Adaptive Resonance Theory Neural Networks for Astronomical Region of Interest Detection and Noise Characterization , 2007, 2007 International Joint Conference on Neural Networks.