Perturbative treatments and learning techniques

L. DIAMBRA; A. PLASTINO

Physical Review E

AMERICAN PHYSICAL SOCIETY

Año: 1996 vol. 53 p. 3970 - 3975

A second order approximation for studying the thermodynamics of the learning process in neural networks is proposed. Particular attention is paid to noise effects. We show that the inclusion of pair interactions between replicas considerably improves upon the well-known first-order approach.