Skip to content

‹ Project Plan index

Reading list

Tags: [STD] re-implement · [EXT] extend to our setting · [NEW] genuine contribution. The N-x claims are stated in full in novelty.md.

Position against (work we must distinguish ourselves from)

Grouped by the part of our approach each bears on.

  • The coupling as a live loop (N-A): DiSF (2025) — selection→collapse, offline · SOFed/FedCoCo (Shi 2022) — streaming-SSL selection, single criterion, no loop (also the closest prior for N-D).
  • Budget flip under co-adaptation (N-B): CCS (Zheng 2023) · Sorscher (2022) · D2 Pruning (Maharana 2024) — offline, fixed model.
  • Selection-loop (in)stability (N-C): one-sided-feedback (2020) · recsys/bandit sampling-bias loops · LLM self-consuming / model collapse (Shumailov 2023) — other domains.
  • Selection × aggregation skew in FL (N-D): FedU (Zhuang 2021) · Orchestra (Lubana 2022) — federated SSL without active selection bias.
  • Health-monitor → control loop (N-E): RankMe (Garrido 2023) · LiDAR (Thilak 2024) · dimensional collapse (Jing 2022) · ADWIN (Bifet 2007) — health metrics / drift as offline diagnostics, not live control.
  • Federation-level health analytics (N-F): Krum (Blanchard 2017) · trimmed-mean/median (Yin 2018) — bad-client detection by update geometry, not representation health.
  • Health-steerable filter (N-G): D2 Pruning (2024) · CCS (2023) — coverage/difficulty mix fixed offline, never steered live.
  • Collapse prevention (context for N-A/N-E): VICReg (Bardes 2022) · IConE (2026) · AdaDim (2025) · CMP (2025) — via loss/architecture, not selection, not a loop.
  • Online-SSL degradation premise (Phase 1 / motivates B5): continual-SSL line · RanDumb (2024) — streaming degrades SSL; fixed features can rival online-learned ones.
  • Noise-robust selection (temporal-robustness layer, [EXT]): Co-teaching · DivideMix · FALSE · RSS-MGM — supervised label noise; we transpose to label-free sensor corruption.
  • Generative replay (replay codebook, [EXT]): CAN (2025) · diffusion-as-replay — federated generative replay; we differ by inversion + embedding-space + temporal scheduling.
  • Automotive FL-SSL (the use case): federated SSL for AV depth (2023).

Toolbox we build on (work we use, does not threaten novelty of our work)

  • SSL backbones [STD]: MAE (He 2022) · DINO (Caron 2021) · SwAV (Caron 2020) · SimSiam (Chen & He 2021) · BYOL/SimCLR/MoCo (2020) · CaSSLe (Fini 2022).
  • Selection [EXT]: Core-Set (Sener & Savarese 2018) · Herding/iCaRL (Welling 2009 / Rebuffi 2017) · SemDeDup (Abbas 2023) · GSS/MIR (Aljundi 2019) · reservoir (Vitter 1985).
  • OOD scoring (F-a) [EXT]: kNN-OOD (Sun 2022) · Mahalanobis (Lee 2018) · Deep SVDD (Ruff 2018).
  • Health / collapse metrics [EXT]: effective rank (Roy & Vetterli 2007) · RankMe (Garrido 2023) · LiDAR (Thilak 2024) · VICReg variance+covariance (Bardes 2022) · alignment & uniformity (Wang & Isola 2020) · dimensional collapse (Jing 2022) · representation drift via CKA (Kornblith 2019) · ADWIN (Bifet 2007).
  • CL & evaluation metrics [STD]: linear probe + kNN eval (Wu 2018) · Backward Transfer (Lopez-Paz & Ranzato 2017) · Forgetting Measure (Chaudhry 2018).
  • Control / replay [EXT]: Deep Generative Replay (Shin 2017) · Brain-inspired replay (van de Ven 2020) · DeepInversion (Yin 2020).
  • FL [STD]: FedAvg (McMahan 2017) · FedProx (Li 2020) · Marfoq (2023) · Scaleout FedN.
  • Datasets: STL-10 (Coates 2011) · BDD100K (Yu 2020) · nuScenes (Caesar 2020) · ZOD.