On the Optimality of Incremental Neural Network Algorithms