This work discusses some aspects of the relationship between connectivity and the capability to store information in neural networks. In these non-linear systems, memories are dynamically stable attractors and the recognition process is identified with the convergence of the sensory input to one of the attractors. It is shown that the strength of the basins of attraction depends on the connective morphology. Networks with strong basins of attraction are used to construct systems in which the dynamics are not limited to the monotonic convergence to one of the memories, but describe trajectories in the memory space. The elements of a trajectory are determined by memory content; upon presentation of an input the system does not recognize just a single configuration but explores a portion of the memory space
Memory and Connectivity
1997-01-01
Abstract
This work discusses some aspects of the relationship between connectivity and the capability to store information in neural networks. In these non-linear systems, memories are dynamically stable attractors and the recognition process is identified with the convergence of the sensory input to one of the attractors. It is shown that the strength of the basins of attraction depends on the connective morphology. Networks with strong basins of attraction are used to construct systems in which the dynamics are not limited to the monotonic convergence to one of the memories, but describe trajectories in the memory space. The elements of a trajectory are determined by memory content; upon presentation of an input the system does not recognize just a single configuration but explores a portion of the memory spaceI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.