‹ Project Plan index

Novelty & positioning

Tags: [STD] re-implement · [EXT] extend to our setting · [NEW] genuine contribution. The one-line version of the claims below lives in the index; this file is the full statement + the closest prior work each [NEW] claim stands against.

★ What is [NEW], and the prior work each claim stands against

The space is crowded along each axis; the assembly is not. Every claim below is paired with the closest prior work on that axis — what it covers, and what it leaves open — so the mapping is explicit. Terminology: an edge is a literal connection in the loop (model→filter, or monitor→filter); the claims below are the distinct axes of contribution, only two of which (inside N-A) are literal edges. The contribution is the closed loop as a live system, not any single axis.

N-A — The coupling, run as a live loop [NEW]. Claim: the two loop edges operate together and are studied live — the fast edge (model→filter: the filter scores informativeness against the continuously-adapting backbone, so what counts as novel shifts as the model learns) and the slow edge (monitor→filter: the health monitor re-aims selection based on the degradation regime). Closest prior: DiSF (2025) links selection→collapse but offline, fixed budget, no model in the loop; SOFed/FedCoCo (2022) does selection in streaming SSL but with one fixed criterion and no health coupling. Open: nobody runs selection ↔ representation-health as a closed feedback loop with a co-adapting model.

N-B — Does the budget flip survive co-adaptation? [NEW] Terms: storage budget = how many frames the buffer may keep; coverage selection = pick frames that represent/span the data distribution; novelty selection = pick hard/atypical frames. The budget flip: at small budgets coverage wins, at large budgets novelty wins — the best rule "flips" coverage→novelty as the budget grows. Closest prior: CCS (Zheng 2023) — coverage beats hard-example selection at high pruning rates (small budgets); Sorscher (2022) — keep prototypical when data is scarce, hard when abundant; D2 Pruning (2024) — balances diversity vs difficulty. All offline, fixed model, pool-based. Open: whether the flip holds, moves, or inverts under single-pass causal selection against a model the selection is itself changing.

N-C — Self-reinforcing selection (in)stability [NEW]. Claim: selection drives the model, which changes what is selected next — a loop that may self-reinforce toward a degenerate diet, or be stabilized. Closest prior: one-sided-feedback learning (2020), recommender/bandit sampling-bias loops, and LLM self-consuming "model collapse" (Shumailov et al. 2023) — the same pathology, in other domains. Open: the streaming-SSL-selection instance, its damping, and how FL averaging amplifies or damps it.

N-D — Selection × aggregation skew in FL [NEW]. Claim: independent clients each over-selecting their local tail may skew the aggregate's effective training distribution; global-aware selection may correct it. Closest prior: SOFed (2022) notes overlapping client selections in passing; federated SSL — FedU (2021), Orchestra (2022) — aggregates unsupervised representations but without active per-client selection bias. Open: characterizing and correcting the selection-induced aggregate drift.

N-E — Online health-monitor → control loop [NEW]. Claim: a multivariate health signal used as a live controller — modulating selection (the monitor→filter edge) and scheduling replay/rollback — not merely a diagnostic. Closest prior: RankMe (Garrido 2023), LiDAR (Thilak 2024), dimensional collapse (Jing 2022) supply the health metrics, but as offline diagnostics / checkpoint selection; ADWIN (Bifet 2007) detects drift but is not tied to SSL-health control. Open: closing the loop live; payoff is conditional — does closed-loop beat open-loop (B-open)?

N-F — Federation-level health analytics [NEW]. Claim: flag a degraded/corrupted client via representation-health diagnostics before its update poisons the global model — the server-side twin of the on-device monitor. Closest prior: robust-FL aggregation — Krum (Blanchard 2017), trimmed-mean/median (Yin 2018) — detects bad updates via update geometry, not via the client model's representation health. Open: using the same health analytics (effective rank, etc.) as the pre-aggregation gate.

N-G — Health-steerable filter (the actuator) [NEW]. Claim: a selection policy whose coverage↔novelty operating point (mix α) and acceptance rate are a live function of the health signal, so the monitor can steer it — the actuator that makes the loop closable. Closest prior: the static filters we would otherwise steer — reservoir (Vitter 1985), SemDeDup (Abbas 2023), loss-gate (SOFed 2022), core-set (Sener & Savarese 2018) — expose no coverage↔novelty knob; the hybrid filters that do have the mix — D2 Pruning (2024), CCS (2023) — fix it offline, never driven by a live signal. Open: making the mix (and acceptance rate) a controlled variable set by the monitor. Validation: only via closed-loop vs. open-loop (B-open) on health + downstream — never a static-benchmark win.

Summary

• The loop, not just the leaderboard, is what makes this filter design novel
• This informs our approach: we can implement the loop with existing pieces, and analyze system behavior. This study is novel.
• Afterward: work toward novelty along any of the axes above.

Everything else is [STD] or [EXT]. If a claim is not in the list above, we are not claiming it as novel. The full reference list (including the [STD]/[EXT] toolbox) is in the reading list.

The prior work grouped by the claim it bears on is in the reading list → Position against.