arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2403.01327cs.CGcs.DS
收藏

通过深度随机特征进行欧氏距离压缩

Euclidean distance compression via deep random features

Brett Leroux, Luis Rademacher

AI总结:

本文提出由随机特征映射复合而成的深度随机非线性映射,可将点集直接编码为离散立方体中的二值草图并保持成对平方距离,在部分参数下减少存储比特。

摘要:

受将点集压缩为尽可能少的比特,同时保留点之间近似距离信息这一问题的启发,我们构造了随机非线性映射 $φ_\ell$,以下述方式压缩点集。对于点集 $S$,映射 $φ_\ell:\mathbb{R}^d \to N^{-1/2}\{-1,1\}^N$ 具有如下性质:只要最小距离相对于 $ε$ 不太小,存储 $φ_\ell(S)$($S$ 的一个“草图”)就能以高概率报告 $S$ 中各点之间的成对平方距离,误差至多为某种乘性 $(1\pm ε)$。映射 $φ_\ell$ 是某类随机特征映射的 $\ell$ 重复合。此外,我们确定了 $N$ 需要多大,将其表示为 $ε$ 和点集其他参数的函数。与现有技术相比,我们的映射具有若干优势。通过随机映射压缩点集的标准方法依赖 Johnson-Lindenstrauss 引理,该引理意味着:如果将一个含 $n$ 个点的集合用高斯随机矩阵映射到 $\mathbb{R}^k$,且 $k=Θ(ε^{-2}\log n)$,那么点之间的成对距离会以高概率得到保持,误差至多为乘性 $(1\pm ε)$。我们的映射 $φ_\ell$ 相对于随机线性映射的主要优势在于,它将点集直接映射到离散立方体 $N^{-1/2}\{-1,1\}^N$ 中,因此不需要额外步骤将草图转换为比特。在某些参数范围内,我们的映射 $φ_\ell$ 生成的草图所需的比特存储空间更少。

英文摘要:

Motivated by the problem of compressing point sets into as few bits as possible while maintaining information about approximate distances between points, we construct random nonlinear maps $φ_\ell$ that compress point sets in the following way. For a point set $S$, the map $φ_\ell:\mathbb{R}^d \to N^{-1/2}\{-1,1\}^N$ has the property that storing $φ_\ell(S)$ (a \emph{sketch} of $S$) allows one to report pairwise squared distances between points in $S$ up to some multiplicative $(1\pm ε)$ error with high probability as long as the minimum distance is not too small compared to $ε$. The maps $φ_\ell$ are the $\ell$-fold composition of a certain type of random feature mapping. Moreover, we determine how large $N$ needs to be as a function of $ε$ and other parameters of the point set. Compared to existing techniques, our maps offer several advantages. The standard method for compressing point sets by random mappings relies on the Johnson-Lindenstrauss lemma which implies that if a set of $n$ points is mapped by a Gaussian random matrix to $\mathbb{R}^k$ with $k =Θ(ε^{-2}\log n)$, then pairwise distances between points are preserved up to a multiplicative $(1\pm ε)$ error with high probability. The main advantage of our maps $φ_\ell$ over random linear maps is that ours map point sets directly into the discrete cube $N^{-1/2}\{-1,1\}^N$ and so there is no additional step needed to convert the sketch to bits. For some range of parameters, our maps $φ_\ell$ produce sketches which require fewer bits of storage space.

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。

↑