The dynamic threshold of semi-orthogonal associative memory model

Discusses the fundamental properties of state evolution in the recalling processes of the semi-orthogonal associative memory (SAM) model and derives the optimum dynamic threshold of a SAM. In a probabilistic sense, there is a convergence criterion Q_ν in the SAM such that, for arbitrary initial input, the recalling outputs converge to the desired pattern on this input when the effective Hamming distance between the desired pattern and the initial input is no larger than (1-Q_ν)N, but the attracting basins of stable states in this model are strange. It is proved that the strange attracting basins can be improved by introducing a dynamic threshold into this model. Making use of the statistical neurodynamics, the optimum dynamic threshold is given and its efficiency is investigated