Stability and attractivity in associative memory networks (Q1101376)

The problem is addressed how to get for a net of N bipolar neurons a connection matrix C so that domains of attraction of learned patterns S m (1\(\leq m\leq p)\) are as great as possible. As known, if C is given by \[ C=(A_ 1s\quad 1,...,A_ ps\quad p)(\Sigma '\Sigma)^{-1}\Sigma '+\tilde C, \] where \(\Sigma\) is formed by the patterns S m (linearly independent), A m are positive diagonal and \(\tilde C\Sigma\) \(=0\), then stability (perfect recall) of the patterns is guaranteed. The paper shows, how the radii of Hamming domains of attraction of the patterns depend on the A m and suggests the construction of the A m which leads to semioptimal C. For equal-sized domains, each pattern attracts vectors from at least 1/(2 max \(C_{ii})\) neighbourhoods; for the orthogonal case (the Hebb rule) the radius is confirmed to be at least N/2p.