“Adaptive Attention Span in Transformers”, Sainbayar Sukhbaatar, Edouard Grave, Piotr Bojanowski, Arm Holdings, Joulin2019-05-19 (sparse Transformers; similar)⁠:

We propose a novel self-attention mechanism that can learn its optimal attention span. This allows us to extend the maximum context size used in Transformer, while maintaining control over their memory footprint and computational time.

We show the effectiveness of our approach on the task of character level language modeling, where we achieve state-of-the-art performances on text8/enwik8 by using a maximum context of 8k characters.

Similar Links: