Cross-validation and Bootstrap for Optimizing the Number of Units of Competitive Associative Nets

Cross-validation and bootstrap, or resampling methods in general, are examined for applying them to optimizing the number of units of competitive associative nets called CAN2. There are a number of resampling methods available, but the performance depends on the neural network to be applied and functions to be learned. So, we apply several resampling methods to the CAN2 which has been shown effective in many areas so far. By means of numerical experiments, we have observed that a modified bootstrap method and the Lendasse’s bootstrap estimator work well for selecting the number of units. We also describe a new method and its performance for estimating generalization error via resampling methods.