A new method to evaluate a trained artificial neural network
conference contributionposted on 27.01.2009, 09:33 by Yingjie Yang, Chris J. Hinde, D Gillingwater
In comparison with traditional local sample testing methods, this paper proposes a new approach to evaluate a trained neural network. A new parameter is defined to identify the different potential roles of the individual input factors based on the trained connections of the nodes in the network. Compared with field-specific knowledge, the dominance of individual input factors can be checked and then false mappings satisfying only the specific data set may be avoided.
- Computer Science