A model for cortical memory

Abstract

Information in the nervous system is coded into patterns of firing in specific collections of neurons. We assume without proof that memorization is the conversion of these firing patterns into spatially distributed patterns (such as changes in the transmission properties of the synapses or cell membranes). We also assume that recall is the synthesis of a copy of the firing patterns which occurred during storage. It is the purpose of this paper to develop a network called a memory block which can store, search for, and recall these firing patterns. The storage mechanism of the memory block closely resembles the holographic storage process. Information which is stored in a particular memory location is distributed throughout all the neurons of that location. Damage to large numbers of neurons in a given memory location does not destroy the stored information but merely reduces the fidelity of the recalled patterns. When a specific piece of information is required from the memory block, a similar information pattern is sent to every memory location. Each memory location correlates the test pattern with its stored patterns and returns a signal which is proportional to the similarity between the test pattern and the corresponding stored pattern. A high signal indicates a close match and locates the stored information. Recall is immediately possible. To convert between temporal firing patterns and spatial neuronal changes, a memory timing mechanism is introduced. This mechanism is a wave of activity in the nonneuronal medium which inactivates all but a thin layer in the network. During storage, the active layer moves through the appropriate memory location. In this way, the time of input is converted into position within the storage network. During recall, the memory timing mechanism is again used, this time to convert between position of storage and time of output. A learning network is then presented which utilizes the memory block as a storage network. This network differs from other learning models in several prominent ways. Not only are all input patterns permanently stored, but the memory traces remain unchanged for the life of the system. Learning and forgetting are independent of the memory traces themselves but depend on the threshold of recallability of information from the individual memory locations. We conclude with a brief discussion of some physiological and biological questions posed by the model.