Distributed Stable Outlier-Robust Signal Recovery using Minimax Concave Loss

Maximilian H.V. Tillmann, Masahiro Yukawa

研究成果: Conference contribution

抄録

This paper presents a mathematically rigorous framework of remarkably robust signal recovery over networks. The proposed framework is based on the so-called minimax concave (MC) loss, which is a 'hybrid'' between Tukey's biweight loss and Huber's loss in the sense of yielding remarkable outlierrobustness and being able to preserve convexity of the overall cost under an appropriate choice of parameters so that an iterative algorithm could generate a sequence of vectors converging provably to a solution (a global minimizer of the overall cost). We present a formulation which involves an auxiliary vector to accommodate the statistical property of noise explicitly, and we present a condition to guarantee convexity of the local cost. We apply the distributed triangularly preconditioned primal-dual algorithm to our formulation and show by numerical examples that our proposed formulation exhibits remarkable robustness under devastating outliers, and outperforms the existing methods.

本文言語English
ホスト出版物のタイトルProceedings of the 2023 IEEE 33rd International Workshop on Machine Learning for Signal Processing, MLSP 2023
編集者Danilo Comminiello, Michele Scarpiniti
出版社IEEE Computer Society
ISBN(電子版)9798350324112
DOI
出版ステータスPublished - 2023
イベント33rd IEEE International Workshop on Machine Learning for Signal Processing, MLSP 2023 - Rome, Italy
継続期間: 2023 9月 172023 9月 20

出版物シリーズ

名前IEEE International Workshop on Machine Learning for Signal Processing, MLSP
2023-September
ISSN(印刷版)2161-0363
ISSN(電子版)2161-0371

Conference

Conference33rd IEEE International Workshop on Machine Learning for Signal Processing, MLSP 2023
国/地域Italy
CityRome
Period23/9/1723/9/20

ASJC Scopus subject areas

  • 人間とコンピュータの相互作用
  • 信号処理

フィンガープリント

「Distributed Stable Outlier-Robust Signal Recovery using Minimax Concave Loss」の研究トピックを掘り下げます。これらがまとまってユニークなフィンガープリントを構成します。

引用スタイル