Evolving Neural Update Rules for Sequence Learning

We consider the problem of searching, end to end, for effective weight and activation update rules governing online learning of a recurrent network on problems of character sequence memorisation and prediction. We experiment with a number of functional forms and find that the performance depends on them significantly. We find update rules that allow us to scale to a much larger number of recurrent units and much longer sequence lengths than has been achieved with this approach previously. We also find that natural evolution strategies significantly outperforms meta-gradients on this problem, aligning with previous studies suggesting that such evolutionary strategies are more robust than gradient back-propagation over sequences with thousands(s) of steps.