Compressed sensing recovers signals from few measurements, crucial for imaging and communications. SCAD and other nonconvex penalties can surpass ℓ1 limits, but the nonconvexity of the loss induces multiple solutions. We propose an algorithm that addresses these multiple solutions, inspired by replica symmetry breaking from statistical physics, and achieves improved sample efficiency over existing methods.
| 氏名 | コース | 研究室 | 役職/学年 |
|---|---|---|---|
| Xiaosi Gu | データ科学コース | 情報論的学習 | 博士3回生 |
| Ayaka Sakata | その他の専攻・大学 | 准教授 | |
| Tomoyuki Obuchi | システム科学コース | 情報数理システム | 准教授 |