Reformer: The Efficient Transformer
This addresses the problem of prohibitive training costs for researchers and practitioners using Transformers on long sequences, offering a more efficient alternative.
The paper tackles the high computational cost of training large Transformer models on long sequences by introducing two efficiency techniques: locality-sensitive hashing to reduce attention complexity from O(L^2) to O(L log L) and reversible residual layers to cut memory usage. The resulting Reformer model matches Transformer performance while being significantly more memory-efficient and faster on long sequences.
Large Transformer models routinely achieve state-of-the-art results on a number of tasks but training these models can be prohibitively costly, especially on long sequences. We introduce two techniques to improve the efficiency of Transformers. For one, we replace dot-product attention by one that uses locality-sensitive hashing, changing its complexity from O($L^2$) to O($L\log L$), where $L$ is the length of the sequence. Furthermore, we use reversible residual layers instead of the standard residuals, which allows storing activations only once in the training process instead of $N$ times, where $N$ is the number of layers. The resulting model, the Reformer, performs on par with Transformer models while being much more memory-efficient and much faster on long sequences.