arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2007.14209stat.MLcs.LG
收藏

Langevin Monte Carlo:随机坐标下降与方差缩减

Langevin Monte Carlo: random coordinate descent and variance reduction

University of Wisconsin-Madison(威斯康星大学麦迪逊分校);Wisconsin Institutes for Discovery(威斯康星发现研究院)

Zhiyan Ding, Qin Li

AI总结:

本文研究了将随机坐标下降(RCD)与方差缩减技术(SAGA、SVRG)结合以提升Langevin Monte Carlo的计算效率,在欠阻尼框架下实现了每次迭代仅需单方向导数的最优计算成本。

摘要:

Langevin Monte Carlo (LMC) 是一种流行的贝叶斯采样方法。对于对数凹分布函数,该方法以指数级速度收敛,且仅存在可控的离散化误差。然而,该方法在每次迭代中需要计算全梯度,对于 $\mathmathbb{R}^d$ 上的问题,这意味着每次迭代需要进行 $d$ 次偏导数计算。当 $d\gg1$ 时,计算成本很高。在本文中,我们研究了如何通过将 RCD(随机坐标下降)应用于 LMC 来提高计算效率。理论分为两个方面:1. 盲目地将 RCD 应用于 LMC 时,每次迭代仅用随机选择的方向导数代替全梯度。尽管单次迭代成本降低,但为达到预设的误差容限,总迭代次数增加。最终并未带来计算收益;2. 随后,我们将方差缩减技术,如 SAGA(随机平均梯度)和 SVRG(随机方差缩减梯度),引入 RCD-LMC。将证明,与经典 LMC 相比其成本有所降低;且在欠阻尼情况下,以相同的迭代次数实现收敛,同时每次迭代仅需计算一个方向导数。这意味着我们在 underdamped-LMC 框架下获得了最优的计算成本。

英文摘要:

Langevin Monte Carlo (LMC) is a popular Bayesian sampling method. For the log-concave distribution function, the method converges exponentially fast, up to a controllable discretization error. However, the method requires the evaluation of a full gradient in each iteration, and for a problem on $\mathbb{R}^d$, this amounts to $d$ times partial derivative evaluations per iteration. The cost is high when $d\gg1$. In this paper, we investigate how to enhance computational efficiency through the application of RCD (random coordinate descent) on LMC. There are two sides of the theory: 1 By blindly applying RCD to LMC, one surrogates the full gradient by a randomly selected directional derivative per iteration. Although the cost is reduced per iteration, the total number of iteration is increased to achieve a preset error tolerance. Ultimately there is no computational gain; 2 We then incorporate variance reduction techniques, such as SAGA (stochastic average gradient) and SVRG (stochastic variance reduced gradient), into RCD-LMC. It will be proved that the cost is reduced compared with the classical LMC, and in the underdamped case, convergence is achieved with the same number of iterations, while each iteration requires merely one-directional derivative. This means we obtain the best possible computational cost in the underdamped-LMC framework.

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。

↑