Correlation-based associative memory and its MOS implementation

We propose an associative memory based on minimization of a free-energy function determined by the library vectors to be stored. When the library vectors are bipolar, the energy function contains minima at the library vector location. The minima can be sought by search techniques such as gradient descent. Significantly, if the correlation nonlinearity is chosen to be sufficiently strong, then convergence occurs in a single step. We demonstrate how the algorithm can be implemented using MOS circuitry.