It would be interesting to see neural networks that can grow layers as needed using this technique. Perhaps you could have the network occasionally monitor its own progress on a tune set somehow and adjust the model capacity based on the amount of over/underfitting observed.