摘要
软阈值缩减迭代算法(ISTA)以其简单的操作流程成为了机器学习流行的优化算法,但是收敛速度比较慢,仅为o(1k)。快速软阈值缩减迭代算法(FISTA)通过加速技巧将收敛速度提高了一个数量级,达到了o(1k2)。然而,FISTA将特征向量每一维看成是独立同分布的,丢失了各维之间的相关性,会导致准确率下降和额外的时间开销。为了弥补上述的不足,文中提出了一种相关快速软阈值坐标下降算法(RFTCD)。通过大规模数据库实验证实了RFTCD的正确性和有效性。
Abstract：Although iterative shrinkage-thresholding algorithm （ISTA） becomes popular optimization algorithms of machine learning be- cause of its simple operational processes,but the convergence rate is slow, ordy 0（1/k） . Convergence rate of fast iterative shrinkage-thresholding algorithm （FISTA） by accelerating skills can improve by an order of magnitude,reaching 0（1/k^2） . However each eigenvec-tors dimension is seen by FISTA as independent and identically distributed, which will loss the correlation between each dimension and lead to the decline in accuracy and time overhead. In order to circumvent these drawbacks,present a relative fast soft-threshotding coordi- nate descent algorithm. Extensive experiments on large-scale real database verify the proposed algorithm is correct and effective.
出处
《计算机技术与发展》
2013年第12期55-58,共4页
Computer Technology and Development
基金
国家自然科学基金资助项目(60975040)
关键词
软阈值缩减迭代
机器学习
特征向量
独立同分布
坐标下降
iterative shrinkage-thresholding
machine learning
feature vector
independent and identically distributed
coordinate descent
