Which wearable-pretraining design choices actually win in the Inertia-1 sweep, and by how much?
The Inertia-1 paper (arXiv:2607.06617) reports a systematic sweep over data choices (sensor modality, device placement, sampling rate, window length), model choices (architecture, model size), and training choices (pretraining objective, data scale), evaluated across 15 downstream datasets. See claim-inertia-1-design-sweep-15-downstream-datasets. The abstract states the sweep was run but does not report its results.
What needs answering
- Which specific design choices came out ahead on each axis (which architecture, which pretraining objective, which sampling rate / window length)?
- How much of the downstream transfer gain is attributable to data scale (the headline 18.2M hours — see claim-inertia-1-pretrained-18-million-hours-accelerometer) versus model size? Is there a clean scaling law, or diminishing returns?
- What are the margins — are the "winning" choices decisive, or within noise across the 15 datasets and three task families (activity recognition, gait detection, disease prediction)?
Why it matters
The abstract's "pretraining and scaling principles remain poorly understood" framing is doing rhetorical work; the specific ablation outcomes are what would make any recipe claim genuinely load-bearing in the vault. Until read, the design-sweep note stays seedling with an [unverified-quant/mechanism] flag.
How to answer
Read the paper body directly at https://arxiv.org/abs/2607.06617 (open access; full text / PDF) — specifically the ablation and results tables — and extract the per-axis winners with their margins. Then promote the flagged claim-note and remove the flag, or split the confirmed results into their own claim-notes.
Candidate next moves
- A future hop into the cited adjacent prior art named in the capture — RelCon (ICLR 2025, relative contrastive learning for wearables) and "Scaling wearable foundation models" (ICLR 2025) — would give a baseline for judging whether Inertia-1's winning choices are novel or confirm existing wearable-pretraining findings.
Progress log
- 2026-07-18: answered by claim-inertia-1-data-scale-beats-model-size, claim-inertia-1-triaxial-beats-vector-magnitude-input, claim-inertia-1-no-pretraining-objective-dominates, claim-inertia-1-self-supervised-beats-supervised-gap-widens, claim-inertia-1-downsampling-robustness-is-task-dependent, claim-inertia-1-no-single-window-length-is-optimal — settled by a direct read of the paper body (Results and Analyses, §4, arXiv:2607.06617), which supplied per-axis winners and margins for all three of this question's sub-asks (which choices won, data-scale-vs-model-size, and whether margins are decisive): model size plateaus/degrades while data scale helps; triaxial beats vector-magnitude by 3.2 AUROC overall; SSL beats supervised by 4–20 points depending on task, with objective choice mattering far less than that; sampling-rate and window-length effects are real but task-dependent, not universal. One caveat carried into every answering note: the paper's own per-axis ablations were run only on its two representative architectures (ART, PatchTST), not the full method set.