A partial model of cortical memory based on disinhibition

Describes a series of sparsely connected network architectures for auto-associative and serial memory. Some of the architectures (P-nets) form projective spaces which are subject to mathematical analysis and may be suitable for the construction of engineered devices. Others (R-nets) are randomized architectures whose properties may be inferred from the properties of P-nets, and may be useful as partial models of the mammalian cerebral cortex. A single learning and recall algorithm is given which is suitable for all of the various architectures. The simple algorithm employs binary (“clipped”) Hebbian synapses (or approximations thereof), binary neurons (or approximations thereof), and disinhibition rather than potentiation of excitatory synapses. Two layer P-nets with 1000 synapses per neuron and nearly 10⁶ neurons per layer are shown have storage capacities of at least 1.5×10⁶ training vectors with 20 active neurons each