A neural network based multi-associative memory model

An improved two-layer neural network model is presented for a multiassociative content-addressable memory. The first layer serves as a preprocessor and computes the association vector which is passed through a set of nonlinear amplifying elements for improved error correction and faster convergence. When compared to a Hopfield model, this model requires
O
(
P
) connections for
P
-bit-long vectors, yields improved error-correction and storage capabilities and faster convergence rate, avoids the storage of complementary and false memories, and possesses analogies to biological neural networks. It is modular and expandable, can be used to storage
N
sets of
M
interassociated sets of vectors, and can retrieve all of the
M
vectors simultaneously by using a vector probe consisting of one or more vectors closest to one or more of M vectors. The use of a bipolar retrieval key has the effect of doubling the length of the probe vector and making the number of +1s equal to the number of -1s in the stored vectors even if the original stored vectors have unequal numbers of +1s and -1s