arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1506.09016cs.LGcs.CVcs.NAmath.NAmath.OCstat.ML
收藏

在线学习采样

Online Learning to Sample

University College London(伦敦大学学院);Xerox Research Center Europe(施乐欧洲研究中心);UGA(格勒诺布尔-阿尔卑斯大学)

Guillaume Bouchard, Théo Trouillon, Julien Perez, Adrien Gaidon

AI总结:

本文提出AW-SGD算法,通过在线学习优化采样策略,提升在线优化效率,应用于图像分类、矩阵分解和强化学习。

摘要:

随机梯度下降(SGD)是机器学习中用于在线优化最广泛使用的技术之一。在本工作中,我们通过适应性地学习如何在每个时间步选择最有用的训练示例来加速SGD。首先,我们证明SGD可以用于学习重要采样估计器的最佳可能采样分布。其次,我们证明SGD算法的采样分布可以通过逐步最小化梯度的方差来在线估计。所得到的算法——自适应加权SGD(AW-SGD)——维护一组用于优化的参数,以及一组用于采样学习示例的参数。我们证明AWSGD在三个不同的应用中实现了更快的收敛:(i)使用深度特征的图像分类,其中图像的采样取决于其标签,(ii)矩阵分解,其中行和列不是均匀采样的,以及(iii)强化学习,其中优化和探索策略同时被估计,其中我们的方法对应于一个off-policy梯度算法。

英文摘要:

Stochastic Gradient Descent (SGD) is one of the most widely used techniques for online optimization in machine learning. In this work, we accelerate SGD by adaptively learning how to sample the most useful training examples at each time step. First, we show that SGD can be used to learn the best possible sampling distribution of an importance sampling estimator. Second, we show that the sampling distribution of an SGD algorithm can be estimated online by incrementally minimizing the variance of the gradient. The resulting algorithm - called Adaptive Weighted SGD (AW-SGD) - maintains a set of parameters to optimize, as well as a set of parameters to sample learning examples. We show that AWSGD yields faster convergence in three different applications: (i) image classification with deep features, where the sampling of images depends on their labels, (ii) matrix factorization, where rows and columns are not sampled uniformly, and (iii) reinforcement learning, where the optimized and exploration policies are estimated at the same time, where our approach corresponds to an off-policy gradient algorithm.

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。

↑