Research map / Time series / Foundation models for time series
RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RelCon, a self-supervised relative contrastive learning method with a learnable motif-based distance, used to pretrain a 3.9M-parameter accelerometry foundation model on Apple Heart and Movement Study data.
Problem
Building a general-purpose foundation model for wearable accelerometry that transfers across HAR and gait regression tasks, where hard positive/negative contrastive labels are prone to false positives and negatives.
Approach
Learn a motif-based distance (cross-attention reconstruction with accel augmentations, sparsemax, and RevIN) and use it to define soft relative rankings of candidates in a contrastive loss, with candidates drawn both within-subject across time and between subjects.
Key results
- RelCon achieves the best double support time results in all 5 gait evaluations (corr .7559 vs .7074 for SimCLR) and 1st in 3/5 stride velocity evaluations.
- On AHMS workout-level HAR, RelCon reaches F1 55.28 vs 54.63 (AugPred) and 50.06 (SimCLR).
- With an MLP probe on Opportunity and PAMAP2, RelCon FM reaches F1 69.1 and 85.4 vs 57.0 and 72.5 for the UK Biobank FM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RelCon rather than REBAR's hard contrastive loss | RelCon · Stride Velocity and Double Support Time | Correlation | Helped |
| Use RelCon rather than SimCLR | AHMS Activity Classification · Motionsense (Wrist→Waist) | F1: 83.93 ± 1.78 vs 80.35 ± 0.71 | Hurt |
| Add accelerometry-specific augmentations during distance-function learning (ablation) | RelCon · AHMS Activity Classification, subsequence level | F1: -12.5% | Helped |
| Add accelerometry-specific augmentations during distance-function learning (ablation) | RelCon · AHMS Activity Classification, workout level | F1: -17.93% | Helped |
| Retain RevIN in the distance function (ablation) | RelCon · Stride Velocity | Correlation: -11.66% | Helped |
| Use SparseMax rather than softmax in the distance function (ablation) | RelCon · AHMS Activity Classification, subsequence level | F1: -5.06% | Helped |
| Sample candidates within-subject across time (ablation) | RelCon · AHMS Activity Classification, workout level | F1: -7.25% | Helped |
| Replace log-ratio metric loss with relative contrastive loss (ablation) | RelCon · Stride Velocity | Correlation: -3.64% | Helped |
| Replace log-ratio metric loss with relative contrastive loss (ablation) | RelCon · AHMS Activity Classification, workout level | F1: 1.39% | Mixed |
| Replace hard binary contrastive loss with relative contrastive loss (ablation) | RelCon · AHMS Activity Classification, workout level | F1: -9.03% | Helped |
Surprising findings
- SimCLR achieves the highest stride velocity correlation (.8454) despite not being designed for gait analysis.
- On Motionsense, SimCLR and a fully supervised model outperform RelCon FM under linear probing.
Limitations
- Training cost scales with candidate count due to per-candidate distance computation.
- Evaluated only on single wrist/accelerometer-only data; multi-sensor extension left to future work.
- Underperforms SimCLR on Motionsense and some stride velocity metrics.
Experimental setup
Datasets: Apple Heart and Movement Study (AHMS) IMU data Opportunity PAMAP2 HHAR Motionsense Capture-24
Benchmarks: Gait metric regression (stride velocity, double support time) AHMS field HAR (16 classes, subsequence and workout level) Opportunity PAMAP2 HHAR Motionsense
Metrics: MSE SDSE MAE SDAE Pearson correlation F1 Cohen's kappa accuracy macro AUC
Baselines: Accel SimCLR Augmentation Prediction REBAR UK Biobank foundation model (Yuan et al.) SimSiam BYOL MAE CPC Autoencoder DeepConvLSTM
Builds on: SimCLR REBAR Accel SimCLR sparsemax reversible instance normalization
Closest papers by meaning
- Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.88)
- Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.88)
- UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.85)
- MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition (NeurIPS 2025 · similarity 0.85)
- Limitations in Employing Natural Language Supervision for Sensor-Based Human Activity Recognition - And Ways to Overcome Them (AAAI 2025 · similarity 0.85)
- Spatial-Related Sensors Matters: 3D Human Motion Reconstruction Assisted with Textual Semantics (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.