This is an API for evaluating and training RNNs on memory-constrained systems. It provides APIs for some of the memory-saving algorithms described in Gruslys et al. and Chen et al..
Traditional back-propagation requires O(T) memory for sequences of length T. This makes RNNs difficult to train on certain tasks, even when the RNN could theoretically learn the task. One example of such a task is meta-learning with long episodes.
With less naive back-propagation techniques, memory consumption can be reduced to O(log(T)) without significant performance sacrifices. This makes RNNs suitable for a much wider range of applications.