Spectral conjugate gradient methods with sufficient descent property for neural network training
Παναγιώτης Πιντέλας, Ιωάννης Λιβιέρης · 2008
Conjugate gradient methods constitute an excellent choice for efficiently training large neural networks since they don't require the evaluation of the Hessian matrix neither the impractical storage of an approximation of it. Despite the theoretical and practical advantages of these methods their main drawback is the use of restarting procedures in order to guarantee convergence, abandoning second order derivative information. In this work, we propose a neural network training algorithm which preserves the advantages of classical conjugate gradient methods and simultaneously avoids the inefficient restarts. Encouraging numerical experiments verify that the presented algorithm provides fast, stable and reliable convergence.