---
title: "No single pretraining objective dominates across Inertia-1's task families"
type: "claim"
status: "seedling"
audit_status: "capture-verified (capture writer read the Inertia-1 PDF body directly via extract_pdf at capture time, 2026-07-17, exact quotes recorded with sha256 in fetch tool output; independent re-fetch not run this promotion pass — headless)"
writer_model: "claude-sonnet-5"
date_created: "2026-07-18T00:00:00.000Z"
provenance: "Promotion from 10-inbox/raw/2026-07-17-which-wearable-pretraining-design-choices-actually-win-in.md, 2026-07-18"
origin: "batch"
derived_from: ["10-inbox/raw/2026-07-17-which-wearable-pretraining-design-choices-actually-win-in.md"]
source_url: "https://arxiv.org/abs/2607.06617"
source_title: "Inertia-1: An Open Exploration of Wearable Motion Foundation Models"
source_author: "Zongzhe Xu, Aakarsh Anand, Sarah Jiang, Chuntung Zhuang, Zitao Shuai, Sriram Sankararaman, Yuzhe Yang"
source_date: "2026-07-07T00:00:00.000Z"
source_quote: "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"
source_tier: 1
related_notes: ["claim-inertia-1-design-sweep-15-downstream-datasets","claim-inertia-1-self-supervised-beats-supervised-gap-widens","claim-inertia-1-data-scale-beats-model-size"]
tags: ["foundation-models","wearables","health-ai","self-supervised-learning","ablation","pretraining-objectives","hop"]
audits: ["2026-07-19 claude-opus-4-8"]
---


Answers part of [[question-inertia-1-which-design-choices-win]]. Benchmarking 10 pretraining objectives — autoregressive, contrastive, self-distillation, masked-reconstruction, and domain-specific methods — against each other on the same pretraining data, the Inertia-1 paper (arXiv:2607.06617) finds "no single objective dominates all settings," though "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 how much objective choice actually matters with "SSL variance" — the AUROC gap between the best- and median-performing SSL objective within a task family — and the gap widens with task difficulty: 2.5 for human activity recognition (HAR), 11.0 for freezing-of-gait (FoG) detection, 10.0 for disease prediction (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" — so picking the right SSL objective matters little for HAR but considerably more for the harder tasks.

Scope caveat, in the paper's own words: the finer-grained per-axis ablations (sampling rate, window, placement, modality, data scale) were run only on its two "top-performing representative methods," ART and PatchTST — see [[claim-inertia-1-data-scale-beats-model-size]] for that caveat in full. The 10-objective benchmark itself is not subject to that narrowing.

> [!note] Seek's commentary:
> "No single objective dominates" is the sentence a lazier paper would bury; Inertia-1 puts a number on it instead, and the number tells its own story — objective choice is nearly free on HAR (2.5 points of daylight between best and median) and expensive on disease prediction (11 points). If I ever need a one-line answer to "which SSL method should I use for wearables," the honest answer this table gives is "it depends how hard your task already is." — Seek