SMYRF: Efficient Attention using Asymmetric Clustering
This addresses the problem of high memory and computational costs in attention-based models for NLP and computer vision researchers and practitioners, offering a practical improvement over prior methods that required constraints and retraining.
The paper tackles the computational inefficiency of attention mechanisms by proposing SMYRF, an algorithm that reduces attention complexity from O(N^2) to O(N log N) using asymmetric clustering, enabling it to serve as a drop-in replacement without retraining. It reports results such as SMYRF-BERT outperforming BERT on GLUE with 50% less memory and scaling attention to up to 65k tokens in GANs.
We propose a novel type of balanced clustering algorithm to approximate attention. Attention complexity is reduced from $O(N^2)$ to $O(N \log N)$, where $N$ is the sequence length. Our algorithm, SMYRF, uses Locality Sensitive Hashing (LSH) in a novel way by defining new Asymmetric transformations and an adaptive scheme that produces balanced clusters. The biggest advantage of SMYRF is that it can be used as a drop-in replacement for dense attention layers without any retraining. On the contrary, prior fast attention methods impose constraints (e.g. queries and keys share the same vector representations) and require re-training from scratch. We apply our method to pre-trained state-of-the-art Natural Language Processing and Computer Vision models and we report significant memory and speed benefits. Notably, SMYRF-BERT outperforms (slightly) BERT on GLUE, while using $50\%$ less memory. We also show that SMYRF can be used interchangeably with dense attention before and after training. Finally, we use SMYRF to train GANs with attention in high resolutions. Using a single TPU, we were able to scale attention to 128x128=16k and 256x256=65k tokens on BigGAN on CelebA-HQ.