SPATIOTEMPORAL INFORMATION-STORAGE IN A CONTENT-ADDRESSABLE MEMORY USING REALISTIC NEURONS

We study the storage and retrieval of spatiotemporal patterns in a net work of realistic neurons. These neurons are simplified versions of th e Hodgkin-Huxley model. Their time dynamics are exploited to represent information in terms of oscillatory spatiotemporal patterns. The prop osed scheme uses excitatory connections to represent the spatial conte nt of information (i.e., at which neurons the spikes are generated) an d inhibitory connections for the temporal content (i.e., the timing of the spikes generated at different neurons). Two self-organizational r ules are presented for the construction of a content addressable syste m. Computer simulations demonstrated successful retrieval of spatial o scillatory patterns in the presence of noise and time-dependent fluctu ations of the inputs. The results suggest an encoding/decoding scheme in terms of the spatiotemporal distribution of action potentials, whic h is consistent with recent experimental results.