Parallel Recursive LSTM

Tristan Gaudreault, Yongyi Mao

Transformers have become the dominant architecture for sequence modeling by using self-attention to enable expressive and highly parallel processing. However, the resulting quadrat…