A comparison of separating capabilities of the feed-next and the feed-forward three layer neural networks
Ryuzo Takiyama, Kimitoshi Fukudome · 2005
To implement every dichotomy on a 2M-point set in R/sup n/, the feed-next three layer neural net needs 2M-1 signum neurons in its hidden layer. It is shown that the feedforward neural net needs only M-1 hidden neurons to do it. An active use of an equivalent transformation theorem on the feedforward net to the feed-next net is discussed.