Which wearable-pretraining design choices actually win in the Inertia-1 sweep, and by how much?

This capture answers question-inertia-1-which-design-choices-win by reading the body (Results and Analyses, §4) of the Inertia-1 paper directly — arXiv:2607.06617, "Inertia-1: An Open Exploration of Wearable Motion Foundation Models" (Xu, Anand, Jiang, Zhuang, Shuai, Sankararaman, Yang; submitted 7 July 2026), fetched via extract_pdf from https://arxiv.org/pdf/2607.06617 (28 pages, TLS verified, sha256 recorded in fetch tool output). This supersedes the abstract-only [unverified-quant/mechanism] flag on claim-inertia-1-design-sweep-15-downstream-datasets for the specific numbers below. No safety flags fired — this is a standard academic preprint, third-person throughout, no addressed-to-AI or override language.

Claim: Data scale, not model size, is the reliable lever for downstream transfer in Inertia-1

Holding the pretraining corpus and input configuration fixed while scaling the ART encoder from ~10M to ~30M to ~100M parameters, "linear-probe performance largely plateaus across model sizes and task families, suggesting that additional capacity does not translate into better representations without additional data or stronger task alignment." Under full finetuning, "larger models can even degrade performance, likely due to overfitting on smaller downstream datasets." By contrast, expanding pretraining data along multiple axes (more individuals, more segments per individual, and mixing UK Biobank with NHANES) "improves AUROC, suggesting that wearable disease representations benefit from population diversity as well as richer behavioral coverage within each person." The paper's own summary: "data volume and diversity provide a more reliable path to improved transfer than increasing model size alone." This is a technical-mechanism + quantitative claim from a controlled ablation (Fig. 3, Fig. 4) in the primary paper — Tier 1, floor cleared.

Claim: Triaxial input consistently beats collapsed vector-magnitude (uniaxial) input, across all three task families

Comparing full triaxial acceleration against vector-magnitude summaries (the compressed format many large cohorts release) under matched pretraining, triaxial input wins on every task family: HAR 95.7 vs. 90.6 AUROC, FoG 79.8 vs. 76.7, disease prediction 68.8 vs. 67.7, for an overall 81.5 vs. 78.3. Quote: "triaxial inputs consistently outperform uniaxial inputs across task families... directional and orientation-dependent motion patterns carry transferable information that cannot be fully recovered after magnitude aggregation." The paper adds one architecture-dependent nuance: "ART is relatively more resilient to this loss than PatchTST... suggesting that architecture can influence robustness to compressed input representations, but preserving raw axes remains the stronger default choice." Tier 1 quantitative claim, directly from Table/Fig. 6 in the primary paper.

Claim: No single pretraining objective dominates, but self-supervised pretraining beats supervised training from scratch, with the gap widening sharply from HAR to FoG to disease prediction

Across 10 pretraining objectives (autoregressive, contrastive, self-distillation, masked-reconstruction, and domain-specific methods) benchmarked against supervised baselines on the same data: "self-supervised pretraining consistently outperforms supervised training from scratch across task categories... While no single objective dominates all settings, specialized or predictive objectives are generally more robust for short-term recognition tasks than generic augmentation-based baselines, while these differences are less pronounced in disease prediction." The paper quantifies this with two metrics per task family — "SSL variance" (gap between best and median SSL AUROC) and "SSL gain" (gap between best SSL and best supervised AUROC): HAR variance 2.5 / gain 4.0; FoG variance 11.0 / gain 16.4; disease prediction variance 10.0 / gain 19.7 (Table 6). In the paper's words, "basic HAR is closer to saturation, whereas clinical and disease-oriented tasks still benefit substantially from representations that capture subtle motion and physiological signatures." Tier 1 quantitative claim, directly from the primary paper's Table 6 and Tables 3–5.

Claim: Sampling-rate and window-length effects are task-dependent, not universal — pretraining buys robustness to downsampling for activity recognition but not for disease prediction

Varying sampling rate {20Hz, 5Hz, 1Hz, 0.2Hz} and window length {10s, 30s, 60s, 2h} under a fixed ART/PatchTST pretraining setup: overall (HAR+FoG+DP) AUROC rises from 71.7 at 1Hz to 76.4 at 5Hz to 81.5 at 20Hz. But the paper notes "pretrained representations remain competitive even at 1Hz" for HAR specifically ("unlike prior studies that report sharp drops below 10Hz"), while "this robustness is weaker for disease prediction, where higher sampling rates provide clearer gains, indicating that clinical outcomes may depend on subtle biomechanical signatures that are attenuated by aggressive downsampling." Window length shows a similar task-split pattern rather than one global optimum: overall AUROC is 78.9 (10s), 81.5 (30s), and 81.2 (60s), with the paper concluding "no single duration is optimal across HAR, FoG, and Disease Prediction... short-window activity labels favor localized motion structure, whereas gait and health tasks may benefit from longer behavioral context." Tier 1 quantitative claim, directly from Table/Fig. 5 in the primary paper.

Further leads

• Time-domain representations beat frequency-domain (spectrogram-style) ones for FoG and disease prediction specifically (overall AUROC 83.0 time vs. 76.9 frequency), with little difference on HAR — Inertia-1 paper, Fig. 6, arXiv:2607.06617.
• Multi-stream fusion (combining sensor modalities or body placements) consistently improves performance over single-stream, and a model pretrained only on wrist accelerometry still transfers to unseen placements/modalities in both single- and multi-stream settings — Inertia-1 paper, §4.4/Fig. 7–8, arXiv:2607.06617.
• The paper's own stated limitation: the detailed per-axis ablations (sampling rate, window, placement, modality, scale) were run only on "top-performing representative methods" (ART, PatchTST), so objective-specific sensitivity to each design axis is explicitly not fully characterized — Inertia-1 paper, §5 Limitations, arXiv:2607.06617.
• Named comparison methods worth their own notes if the vault goes deeper on wearable SSL: RelCon (contrastive, wearable-specific), Self-PAB (masked reconstruction, frequency-domain), DINO (self-distillation) — all benchmarked in Tables 3–5, arXiv:2607.06617.
• Code and project site (not yet fetched/verified this session): https://github.com/yang-ai-lab/Inertia-1 and https://yang-ai-lab.github.io/Inertia-1 — could carry additional ablation detail or model checkpoints.

Safety flags

None. The source is a standard arXiv preprint PDF (extract_pdf reported tls: "verified"), written entirely in third-person academic register. No addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing were present.

Entity candidates

• Yuzhe Yang — person — corresponding author (UCLA), leads the Inertia-1 project; recurring name if the vault tracks wearable/health foundation-model authors.
• Inertia-1 — concept — the paper/model/benchmark itself; already has two claim-notes, could anchor a MOC on wearable foundation models.
• UK Biobank — concept — recurring large-cohort accelerometer dataset used across wearable foundation-model literature (also appears in the existing Inertia-1 pretraining-scale note).
• NHANES — concept — the primary pretraining source in Inertia-1 (14,000 participants, raw high-frequency accelerometer + health/survey variables); recurring health-data cohort.
• RelCon — concept — wearable-specific contrastive pretraining objective (ICLR 2025), cited as prior art in both the design-sweep note's open question and this paper's benchmarks.
• freezing-of-gait detection — concept — clinical Parkinson's-disease task family used as one of Inertia-1's three evaluation regimes; distinct enough from generic "gait detection" to warrant its own definitional note (already flagged once in the design-sweep note's audit).
• SSL variance / SSL gain — concept — the paper's own metrics for quantifying how much pretraining-objective choice matters per task family; could be a useful general methodology note for evaluating "does design choice X matter" questions beyond wearables.