Understanding the difficulty of training deep feedforward neural networks.

Abstract

Cellular Neural Networks (CNN) [1] main assets are quoted to be their capacity for parallel hardware implementation and their universality. On top, the possibility to add the information of a local sensor on every cell, provides a unique system for massive parallel signal processing responding in hardware time. Image processing has been, for a long time, the main field where the community has focussed its efforts to prove the excellence of CNNs. And, still, they are not used at large scale for image applications, probably because few cases are so demanding in terms of computation complexity and short response time not to be afforded by a standard sequential CPU