Symmetry breaking in nonmonotonic neural networks
G. Boffetta, Rémi Monasson, Riccardo Zecchina · Journal of Physics A Mathematical and General · 1993
The optimal performance of a nonmonotonic neural network is studied by the replica method. In analogy to what happens in multi-layered networks, the authors show that replica symmetry breaking (RSB) is required. The distribution of the patterns stabilities, the correlations in the distribution of the internal representation and the optimal capacity per synapse ( alpha c approximately=4.8) are computed with one step of RSB.