No label defined (Q4237930)

The storage capacities of the generalized Hopfield model \[ H_G (\sigma)= -{1\over 2N} \sum^N_{i,j=1} \sum^m_{\mu, \nu=1} \xi_i^\mu \xi_j^\nu G_{\mu\nu} \sigma_i \sigma_j \tag{1} \] are studied. Here the random variables \(\xi_i^\mu\) take values \(\pm 1\) with probability 1/2 and form a Markov chain with transition probabilities \[ P\bigl(\xi_i^\mu =x_i^\mu \mid\xi_i^{\mu-1}= x_i^{\mu-1} \bigr)= Q(x_i^{\mu-1}, x_i ^\mu), \] where \(Q\) is a symmetric \(2 \times 2\) matrix with entries \(p\) and \(1-p\), \(p\neq 1/2\). The model is chosen in the way that matrix \(G\) is inverse to the covariance matrix \(C_{\mu\nu} =E \xi_1^\mu \xi_1^\nu\). The main result is that the storage capacities of this model coincide with those of the regular Hopfield model determined by matrices (1) with \(G_{\mu\nu}= \delta_{\mu\nu}\) and jointly independent \(\{\xi_i^\mu\}\). This behaviour is different to the case of the Hopfield model with weakly correlated patterns \(\xi^\mu\) previously studied by the author [Ann. Appl. Probab. 8, No. 4, 1216-1250 (1998)].