In a previous simulation study, the complexity of neural networks for limited cases of binary and normally-distributed variables based the null distribution of the likelihood ratio statistic and the corresponding chi-square distribution was characterized. This study expands on those results and presents a more general formulation for calculating degrees of freedom.
"Estimating Model Complexity of Feed-Forward Neural Networks,"
Journal of Modern Applied Statistical Methods:
2, Article 13.
Available at: http://digitalcommons.wayne.edu/jmasm/vol8/iss2/13