Sheng Zhou

Research profile

Research summary

Sheng Zhou works at the interface of machine learning and wireless communications, with a catalog footprint concentrated in federated learning over wireless edge systems. The current Evolve record contains two 2026 arXiv papers, split evenly (0.5/0.5 in the field distribution) between two closely related subfields: machine learning for federated learning and wireless edge systems, and communication-efficient federated learning / vehicular edge intelligence. FedPGT (2026) proposes progressive transmission of dominant gradient entries slot by slot in vehicular federated learning over time-varying channels, pairing Lyapunov-based resource allocation with a convergence analysis; Evolve's review credits a coherent theory-to-algorithm pipeline and simulation gains while flagging missing variance reporting and narrow sparsification baselines. Sylvas (2026) frames device scheduling for federated continual learning around a synergistic learning value combining drift, collective divergence, and pseudo-label value; the concept is judged sound but the simulation-only evaluation is thin. On both papers Zhou sits in a middle authorship position and is not the corresponding author. The co-authorship graph is compact — 9 distinct coauthors, with Yuxuan Sun, Tan Chen, and Zhisheng Niu recurring across records — and the reproducibility mean across the two reviewed papers is 38, reflecting medium and low readiness ratings respectively.

Research keywords

Highlights

• FedPGT: progressive gradient transmission over time-varying vehicular channels

  In this 2026 arXiv paper, Sheng Zhou contributes in a middle authorship position to a scheme that uploads dominant gradient entries slot by slot in vehicular federated learning over time-varying channels, pairing Lyapunov-based resource allocation with a convergence analysis. Evolve credits the simulation gains but flags no variance reporting and narrow sparsification baselines; reproducibility readiness is rated medium.

• Sylvas: synergistic learning value for device scheduling in federated continual learning

  This 2026 arXiv paper frames federated continual learning device scheduling around a synergistic learning value that combines drift, collective divergence, and pseudo-label value, touching semi-supervised learning and catastrophic forgetting. Evolve judges the magazine-style concept sound but the evaluation thin and simulation-only; reproducibility readiness is rated low.

• A common thread: bandwidth-aware training at the wireless edge

  Across both reviewed papers the same agenda runs: making federated learning practical under wireless constraints — sparing gradient uploads over time-varying channels on one hand, and choosing which devices should train under drift on the other. Both records carry a combined reproducibility mean of 38, leaving clear headroom in variance reporting and evaluation depth.

Timeline

1. 2026FedPGT appears on arXiv: progressive gradient transmission for vehicular federated learning over time-varying channels, with Sheng Zhou in a middle authorship position and not the corresponding author.
2. 2026Sylvas appears on arXiv: synergistic learning value based device scheduling in federated continual learning, again with a middle, non-corresponding authorship position.
3. 2026Both papers enter the Evolve catalog and are reviewed, receiving reproducibility readiness ratings of medium (FedPGT) and low (Sylvas), for a reproducibility mean of 38.

Papers in the Evolve catalog

• 2026-09-17

  FedPGT: Progressive Gradient Transmission for Vehicular Federated Learning over Time-Varying Channels

• 2026-09-14

  Sylvas: Synergistic Learning Value based Device Scheduling in Federated Continual Learning

Frequent co-authors