Abstract
A method of estimating the number of hidden units required by a two-layer perception learning binary mappings using back-propagation of error signals is presented. A two-layer perception actually has three discernable layers. One is a layer of inputs, each of which is connected to every unit found on the next or hidden layer. These units are called the hidden units and each hidden unit, in turn, is connected to every output unit on the output layer. No intralayer connections are used. Usually the application dictates the number of input units and the number of output units in a rather obvious fashion. Specifying the number of hidden units however is more difficult and yet very important. We consider an example using a net that classifies 2004 input vectors. It would require only a set of 4 such vectors to present the net with a conflict similar to the exclusive-OR problem.
Original language | English (US) |
---|---|
Title of host publication | IEE Conference Publication |
Publisher | Publ by IEE |
Pages | 120-124 |
Number of pages | 5 |
Edition | 313 |
State | Published - 1989 |
Event | First IEE International Conference on Artificial Neural Networks - London, Engl Duration: Oct 16 1989 → Oct 18 1989 |
Other
Other | First IEE International Conference on Artificial Neural Networks |
---|---|
City | London, Engl |
Period | 10/16/89 → 10/18/89 |
ASJC Scopus subject areas
- Electrical and Electronic Engineering