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

We study the storage and retrieval of spatiotemporal patterns in a network of realistic neurons. These neurons are simplified versions of the Hodgkin-Huxley model. Their time dynamics are exploited to represent information in terms of oscillatory spatiotemporal patterns. The proposed scheme uses excitatory connections to represent the spatial content of information (i.e., at which neurons the spikes are generated) and inhibitory connections for the temporal content (i.e., the timing of the spikes generated at different neurons). Two self-organizational rules are presented for the construction of a content addressable system. Computer simulations demonstrated successful retrieval of spatial oscillatory patterns in the presence of noise and time-dependent fluctuations of the inputs. The results suggest an encoding/decoding scheme in terms of the spatiotemporal distribution of action potentials, which is consistent with recent experimental results.