A study on generalization ability of 3-layer recurrent neural networks
Hiroshi Ninomiya, Akinori Sasaki · 2003
In this paper, we report a study on the generalization ability of 3-layer recurrent neural networks (3LRNN). 3LRNN are composed of the both of the feed-forward and feedback connections. The generalization ability of 3LRNN is compared with one of 3-layer feed-forward neural networks through the computer simulations. It is shown that 3LRNN are not only almost equivalent to 3LFNN but also much superior to one on a certain condition from the viewpoint of the generalization capability. Furthermore, we investigate the generalization ability of 3LRNN with the neurons that have the step functions as the input-output property.