Thanks to a wonderful patch supplied by Raphael Manfredi, the author of Storable, AI::NeuralNet::Simple now supports multiple networks, additional activation functions and (drumroll, please) persistence! You can save and restore multiple networks to your heart's content. You can also train towards a target error rate rather than a set number of iterations. That helps to avoid overtraining.
Given that the core is written in C, we have a fast, lightweight backprop neural network available for anyone who's interested. Raphael's changes have changed my small module from a toy into something which might actually be useful for production work.
Version 0.10 should soon be on a mirror near you.