Intelligent Engineering Systems through Artificial Neural Networks Volume 18
62 Weight Search Space with Riemannian Geometry for Acceleration of Backpropagation in a Multilayer Perceptron
Download citation file:
- Ris (Zotero)
- Reference Manager
The convergence of a Multilayer Perceptron (MLP) to a global minimum error in the backpropagation algorithm can be accelerated by introducing the representation of a weight search space, according to the Riemannian geometry, which considers the curvature of this space. The trajectory of the search vector at the training phase is modified by a “force” originated by this curvature. The main concept is developed from the Einstein's General Relativity theory, and the search space (or manifold) must be controlled by parameters like spatial density and scale factor, modifying the gradient descent in the backpropagation algorithm. Some preliminary results show the weights initialization much closer from the final solution, at few epochs, rather than the normal MLP, for an application related here.