An Optimal Reduction of TV-Denoising to Adaptive Online Learning

An Optimal Reduction of TV-Denoising to Adaptive Online Learning
复制标题

DOI:
--
复制
发表时间:
2021-01
期刊:
--
影响因子:
--
通讯作者:
Dheeraj Baby;Xuandong Zhao;Yu-Xiang Wang
Dheeraj Baby;Xuandong Zhao;Yu-Xiang Wang
中科院分区:
其他
文献类型:
--
作者:
Dheeraj Baby;Xuandong Zhao;Yu-Xiang Wang

文献摘要

相似文献

本文考虑了从n个噪声样本中估计一个函数的问题,其离散全变差(TV)由C_n$限定。我们揭示了与强自适应在线学习(Daniely et al,2015)的看似不同的问题的深刻联系,并提供了一个时间复杂度为O(n \log n)的算法,该算法在平方误差损失下获得了O(n^{1/3}C_n^{2/3})的近似最小最大最优速率。由此产生的算法在线运行,并最佳地适应未知的平滑参数$C_n$。这导致了一个新的和更通用的替代小波为基础的方法(1)自适应估计电视有界函数;(2)在线预测电视有界的趋势,在时间序列。
We consider the problem of estimating a function from $n$ noisy samples whose discrete Total Variation (TV) is bounded by $C_n$. We reveal a deep connection to the seemingly disparate problem of Strongly Adaptive online learning (Daniely et al, 2015) and provide an $O(n \log n)$ time algorithm that attains the near minimax optimal rate of $\tilde O (n^{1/3}C_n^{2/3})$ under squared error loss. The resulting algorithm runs online and optimally adapts to the unknown smoothness parameter $C_n$. This leads to a new and more versatile alternative to wavelets-based methods for (1) adaptively estimating TV bounded functions; (2) online forecasting of TV bounded trends in time series.