Learning in neural networks by eliminating frustrated bonds
Harm J. J. Jonker, A C C Coolen · Journal of Physics A Mathematical and General · 1993
The authors study neural network models in which the initial interaction matrix elements are drawn from an arbitrary probability distribution and in which patterns are subsequently stored by eliminating the frustrated bonds, generalizing a proposal by Kinzel (1985). They show that the optimal choice for the a priori distribution corresponds to choosing uniform ferromagnetic initial interactions. For the optimal model they study analytically the dynamical behaviour, the equilibrium properties, the sizes of the domains of attraction and a number of information-theoretical performance measures.