Timezone: »

 
Poster
Stable, Fast and Accurate: Kernelized Attention with Relative Positional Encoding
Shengjie Luo · Shanda Li · Tianle Cai · Di He · Dinglan Peng · Shuxin Zheng · Guolin Ke · Liwei Wang · Tie-Yan Liu

Wed Dec 08 12:30 AM -- 02:00 AM (PST) @
in Poster Session 3 »
The attention module, which is a crucial component in Transformer, cannot scale efficiently to long sequences due to its quadratic complexity. Many works focus on approximating the dot-then-exponentiate softmax function in the original attention, leading to sub-quadratic or even linear-complexity Transformer architectures. However, we show that these methods cannot be applied to more powerful attention modules that go beyond the dot-then-exponentiate style, e.g., Transformers with relative positional encoding (RPE). Since in many state-of-the-art models, relative positional encoding is used as default, designing efficient Transformers that can incorporate RPE is appealing. In this paper, we propose a novel way to accelerate attention calculation for Transformers with RPE on top of the kernelized attention. Based upon the observation that relative positional encoding forms a Toeplitz matrix, we mathematically show that kernelized attention with RPE can be calculated efficiently using Fast Fourier Transform (FFT). With FFT, our method achieves $\mathcal{O}(n\log n)$ time complexity. Interestingly, we further demonstrate that properly using relative positional encoding can mitigate the training instability problem of vanilla kernelized attention. On a wide range of tasks, we empirically show that our models can be trained from scratch without any optimization issues. The learned model performs better than many efficient Transformer variants and is faster than standard Transformer in the long-sequence regime.
Keywords: Optimization · Machine Learning · Transformers · Vision

Author Information

Shengjie Luo (Peking University)
Shanda Li (Peking University)
Tianle Cai (Princeton University)
Di He (Peking University)
Dinglan Peng (University of Science and Technology of China)
Shuxin Zheng (Microsoft)
Guolin Ke (Microsoft Research)
Liwei Wang (Peking University)
Tie-Yan Liu (Microsoft Research Asia)

More from the Same Authors