引用cgreerrun · 2022-06
“Maddness is their more recent work and yields 100x speedups: Multiplying Matrices Without Multiplying”
指出 Maddness 是他们更新的工作,称有 100 倍加速
用学习得到的哈希与查表近似矩阵乘法,比精确乘法快约 100 倍
论文提出一种基于学习的近似矩阵乘法算法,在来自不同领域的数百个矩阵上的实验里,往往比精确矩阵乘快 100 倍,比现有近似方法快 10 倍。当其中一个矩阵事先已知(常见情形)时,该方法还有一个有趣的性质:不需要任何乘加运算。作者认为,其核心运算哈希、平均与字节混洗的组合,可能比近年受到大量研究和硬件投入的稀疏化、分解和标量量化矩阵乘更有前景。
HN 评论者围绕它的后续工作 MADDNESS 展开讨论,作者本人也在评论里说明速度提升在矩阵冗余越多时越大,并称方法大幅超过稀疏截断 PCA 基线。
“Maddness is their more recent work and yields 100x speedups: Multiplying Matrices Without Multiplying”
“We characterize this for various problems in the paper, but tl;dr is that it speeds things up more at a given level of error when there's more redundancy in your matrices.”
“We beat it by a lot (see the paper [1])”
“has been done in the preceding paper: Multiplying Matrices Without Multiplying”