A New Training Method for Feedforward Neural Networks Based on Geometric Contraction Property of Activation Functions
This addresses training efficiency and accuracy for neural network practitioners, but appears incremental as it modifies existing methods rather than introducing a new paradigm.
The authors tackled the problem of training feedforward neural networks by proposing a new method that reduces nonlinearity in the functional, specifically by removing activation function nonlinearity from the output layer, resulting in improved learning speed and better classification error in experiments.
We propose a new training method for a feedforward neural network having the activation functions with the geometric contraction property. The method consists of constructing a new functional that is less nonlinear in comparison with the classical functional by removing the nonlinearity of the activation function from the output layer. We validate this new method by a series of experiments that show an improved learning speed and better classification error.