Neural Computation
May 1993, Vol. 5, No. 3, Pages 363-366
(doi: 10.1162/neco.1993.5.3.363)
Backpropagation with Homotopy
Article PDF (255.26 KB)
Abstract
When training a feedforward neural network with backpropagation (Rumelhart et al. 1986), local minima are always a problem because of the nonlinearity of the system. There have been several ways to attack this problem: for example, to restart the training by selecting a new initial point, to perform the preprocessing of the input data or the neural network. Here, we propose a method which is efficient in computation to avoid some local minima.