| 摘要 |
An object of the present invention is to determine the optimal number of neurons in the hidden layers of a feed-forward neural network. The number of the neurons in the hidden layers corresponds to the number of the independent variables of a linear question and the minimum number of the variables required for solving a linear question can be obtained from the rank value in the matrix theory. Therefore, the rank value corresponds to the minimum number of the neurons required for the hidden layers. Accordingly, when the relation between the neural network constructed and trained is memorized in matrix and the rank value is obtained from this matrix, if the number of the neurons in use is larger than the rank value, as it means that redundant neurons exist which correspond to dependent variables, such redundant neurons can be eliminated. In many cases, due to errors in calculation, the diagonal elements of the matrix are not reduced to 0 and consequently the rank value can not be determined. By neglecting the diagonal elements that are smaller than the specific value e, however, the rank value can be estimated within the range of the error.
|
