Many applications require the extraction of spatiotemporal correlations among dynamically emergent features of non-stationary distributions. In such applications it is not possible to obtain an a priori analytical characterization of the emergent distribution. This paper extends the Growing Cell Structures (GCS) network and presents two novel (GIST and GEST) networks, which combine unsupervised feature-extraction and Hebbian learning, for tracking such emergent correlations. The networks were successfully tested on the challenging Data Mapping problem, using an execution driven simulation of their implementation in hardware. The results of the simulations show the successful use of the GIST and GEST networks for extracting spatiotemporal correlation information among emergent features of previously unknown distributions and, indicate the feasibility of hardware implementation for online use. Of the two networks, the GEST network evinced better performance in terms of the network map stability, feature/correlation tracking ability and network sizes evolved.
Tumuluri, Chaitanya, "Unsupervised Algorithms for Learning Emergent Spatio-Temporal Correlations" (1996). Electrical Engineering and Computer Science Technical Reports. 145.