Nyströmformer: Approximating self-attention in linear time and memory via the Nyström method

Hugging Face · 4 years ago · 5

Nyströmformer approximates the quadratic-complexity self-attention mechanism in standard Transformers by applying the Nyström matrix approximation method to queries and keys rather than directly to the attention matrix. The approach reduces complexity from O(n²) to O(n) while maintaining competitive performance with just 32 or 64 landmarks across sequences of 4,096 to 8,192 tokens. Four pre-trained checkpoints are available on HuggingFace for masked language modeling at different sequence lengths, allowing practitioners to trade off between model capacity and computational efficiency.