arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.09268cs.LG
收藏

Transformers、并行计算与对数深度

Transformers, parallel computation, and logarithmic depth

Columbia University(哥伦比亚大学);New York University(纽约大学)

Clayton Sanford, Daniel Hsu, Matus Telgarsky

AI总结:

本文证明常数个自注意力层与Massively Parallel Computation的常数个通信轮次可相互高效模拟,由此得出对数深度足以让transformers解决其他神经序列模型及次二次近似无法高效求解的计算任务,确立了并行性为其关键区分属性。

摘要:

我们证明,常数个自注意力层能高效模拟,且可被Massively Parallel Computation的常数个通信轮次所模拟。由此,我们证明对数深度足以使transformers解决基本的计算任务,而这些任务无法被其他几种神经序列模型和次二次transformer近似高效解决。因此,我们确立了并行性是transformers的关键区分属性。

英文摘要:

We show that a constant number of self-attention layers can efficiently simulate, and be simulated by, a constant number of communication rounds of Massively Parallel Computation. As a consequence, we show that logarithmic depth is sufficient for transformers to solve basic computational tasks that cannot be efficiently solved by several other neural sequence models and sub-quadratic transformer approximations. We thus establish parallelism as a key distinguishing property of transformers.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑