A Locally-Distributed Associative Memory Network

摘要

The principal purpose of this report is to propose a mathematical model for an associative memory network. A network of mathematical neurons is presented which is capable of storing the information patterns which arrive through specific collections of neurons. The neurons of the model resemble biological neurons in many ways, and it is shown that in a network the size of the cerebral cortex, there is sufficient capacity to store the images accumulated during an average human lifetime. The storage network is based on the principle of 'matched filtering.' The recognition of current information is accomplished by crosscorrelating the current input information with previously stored information. This crosscorrelation occurs simultaneously at every storage location in the memory network whenever an input pattern arrives at the memory network. The recalled pattern from a particular memory location is a copy of the information stored within that memory location. Computer simulations of the memory network indicate that for patterns comprised of 'fine lines,' the recognition signal is stronger than for patterns composed of 'broad lines.' Simulations also show that the memory network functions adequately well even if there is a large amount of background noise. (Author)