상세 보기
Communication-Efficient Federated Learning with Local-Reconstruction Error-Feedback-Based Rescaled 1-Bit Compressive Sensing
- Oh, Junsuk;
- Won, Dongwook;
- Truong, Thanh Phung;
- Cho, Sungrae
SCOPUS
0초록
Federated learning (FL) is well-suited for the next-generation networks, but it is often constrained by the communication cost of repeatedly transmitting high-dimensional updates. Compressive sensing (CS) alleviates this limitation by utilizing a short measurement vector; however, low-dimensional projection and coarse quantization, especially the 1-bit CS (1BCS), can lead to reconstruction errors that slow down convergence. Therefore, we propose local reconstruction error feedback (LREF), where each device performs 1BCS reconstruction locally to feed back not only the sparsification residual but also the reconstruction discrepancy, and introduce the rescaled 1BCS (R1BCS), which uses a magnitude scalar to rescale the unit sparse gradient estimated from the 1-bit measurement vector. We provide upper bounds on the convergence rate, along with the required constraints to ensure the performance of the proposed design. In experiments, R1BCS-LREF approaches the ideal case (i.e., no reconstruction error) while significantly reducing the cost, confirming that the proposed design effectively counters the reconstruction error.
키워드
- 제목
- Communication-Efficient Federated Learning with Local-Reconstruction Error-Feedback-Based Rescaled 1-Bit Compressive Sensing
- 저자
- Oh, Junsuk; Won, Dongwook; Truong, Thanh Phung; Cho, Sungrae
- 발행일
- 2026
- 유형
- Conference Paper
- 언어
- ENG
- 출판사
- Institute of Electrical and Electronics Engineers Inc.
- ISSN
- P 1550-3607