Parallelizing Linear Recurrent Neural Nets Over Sequence Length
This work addresses the computational bottleneck in training RNNs for long sequences, benefiting researchers and practitioners in sequence modeling, though it is incremental as it builds on existing linear RNN frameworks.
The authors tackled the problem of slow RNN training due to sequential dependencies by parallelizing linear RNNs over sequence length using the parallel scan algorithm, achieving up to 9x speedup in training and inference for state-of-the-art architectures and enabling training on sequences with one million timesteps.
Recurrent neural networks (RNNs) are widely used to model sequential data but their non-linear dependencies between sequence elements prevent parallelizing training over sequence length. We show the training of RNNs with only linear sequential dependencies can be parallelized over the sequence length using the parallel scan algorithm, leading to rapid training on long sequences even with small minibatch size. We develop a parallel linear recurrence CUDA kernel and show that it can be applied to immediately speed up training and inference of several state of the art RNN architectures by up to 9x. We abstract recent work on linear RNNs into a new framework of linear surrogate RNNs and develop a linear surrogate model for the long short-term memory unit, the GILR-LSTM, that utilizes parallel linear recurrence. We extend sequence learning to new extremely long sequence regimes that were previously out of reach by successfully training a GILR-LSTM on a synthetic sequence classification task with a one million timestep dependency.