Adaptive Attention Span in Transformers
This addresses the computational and memory bottlenecks in large-scale language modeling, though it is incremental as it builds on existing Transformer architectures.
The authors tackled the problem of limited context size in Transformers by proposing a self-attention mechanism that learns its optimal attention span, enabling an extension to 8k characters and achieving state-of-the-art performance on text8 and enwiki8 datasets.
We propose a novel self-attention mechanism that can learn its optimal attention span. This allows us to extend significantly 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 and enwiki8 by using a maximum context of 8k characters.