A variant of backpropagation through time that is useful for training RNNs on long sequences by limiting the number of timesteps for which gradients are backpropagated. 27.07.2023 17:54 aior