Variable Input Masked Autoencoders for Wrist-Worn Accelerometry

Harish Haresamudram; Wanting Mao; Aditya Radhakrishnan; Maxwell A. Xu; James M. Rehg

cs.LGOct 12, 2026

Variable Input Masked Autoencoders for Wrist-Worn Accelerometry

Authors: Harish Haresamudram, Wanting Mao, Aditya Radhakrishnan, Maxwell A. Xu, James M. Rehg

Organizations: University of Illinois Urbana-Champaign · Google

Abstract

Large-scale repositories of wearable sensor data have enabled the development of foundation models for human activities and behavior. However, existing models are pretrained using fixed input durations and sampling rates, despite substantial heterogeneity in recording conditions, downstream tasks, and deployment constraints. We introduce VarMAE, a variable input masked autoencoder that learns a single representation across sampling rates and temporal contexts. During pretraining we construct two views of the same signal: a long duration high frequency view, and a short duration lower frequency view, obtained via random cropping and downsampling. We perform masked reconstruction within each view and introduce a cross-resolution masked reconstruction objective that uses visible tokens from the long view to predict masked tokens from the short view. Cropping and downsampling impose complementary information bottlenecks, encouraging the model to learn representations that remain effective across input durations and sampling rate changes. We pretrain VarMAE on the large-scale NHANES dataset, producing a single pretrained encoder matching or outperforming controlled foundation model baselines on eight benchmark datasets. With only 3.82M parameters, VarMAE is parameter efficient and shows that a single foundation model can achieve broad transfer capabilities across sensing configurations.

Figures & tables

Figure 1: VarMAE is a variable input MAE trained using two views: first, a long duration, high frequency view, and second, lower frequency, short duration view. In addition to within-view masked reconstruction, we also perform cross-resolution masked reconstruction.

Approach	Enc #Params	1 s		2 s		5 s		10 s
Approach	Enc #Params	Open Watch	Watch Hand	Motion Sense	Mobi Act	HARTH	Real World	PAAWS- SimFL	Capture- 24
End-to-end training
Conv encoder	80k	0.5785	0.6846	0.8850	0.8125	0.7679	0.7420	0.7817	0.4177
ResNet encoder	2.61M	0.6398	0.6286	0.8775	0.8389	0.7933	0.7046	0.8364	0.4516
TSFMs + linear eval
Chronos-T5-small	21M	0.1182	0.4467	0.8303	0.5891	0.5876	0.6417	0.5894	0.4074

Table 1: Activity recognition performance: we contrast VarMAE against end-to-end training, general purpose TSFMs, as well as accelerometer-based FMs. VarMAE matches or outperforms other FMs on six out of eight datasets. Bold is best and underline second best per dataset. † indicates settings where the window duration differs from that used during pretraining.

Figure 2: Few Shot Learning: During linear evaluation, we extract representations only for {2, 5, 10, 25, 50, 100} randomly sampled windows per activity, for five random seeds. Our approach matches or outperforms the specialized FM baselines.

Figure 3: Impact of the pre-training losses: removing the short-view and cross-resolution losses has substantial impact at low frequencies.

Mask ratio
Long view	Short view	50 Hz	25 Hz ∗	10 Hz	5 Hz ∗	Avg.
0.5	0.75	0.7730	0.7681	0.7564	0.6874	0.7462
0.25	0.75	0.7739 (+0.0009)	0.7707 (+0.0026)	0.7565 (+0.0001)	0.6709 ( $-$ 0.0165)	0.7430 ( $-$ 0.0032)
0.75	0.75	0.7703 ( $-$ 0.0027)	0.7628 ( $-$ 0.0053)	0.7503 ( $-$ 0.0061)	0.7166 (+0.0292)	0.7500 (+0.0038)
0.5	0.25	0.7668 ( $-$ 0.0062)	0.7547 ( $-$ 0.0134)	0.7466 ( $-$ 0.0098)	0.5425 ( $-$ 0.1449)	0.7027 ( $-$ 0.0435)
0.5	0.5	0.7751 (+0.0021)	0.7718 (+0.0037)	0.7616 (+0.0052)	0.4934 ( $-$ 0.1940)	0.7005 ( $-$ 0.0457)

Table 2: Varying the masking ratio during pre-training: we tabulate the mean test performance over four evaluation datasets, namely WatchHand, MobiAct, RealWorld, and SimFL. The short view’s masking ratio has substantial impact at low sampling rates. ∗ Rate not used in pretraining.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Win.	#Users	#Classes	Classes
OpenWatch		50	6	double clench, double pinch, pinch down, pinch up, slide, negative
WatchHand	1 s	23	3	wrist up, wrist down, wrist left
MobiAct		61	11	standing, walking, jogging, jumping, stairs up, stairs down, stand-to-sit, sitting, sit-to-stand, car step-in, car step-out
MotionSense	2 s	24	6	stairs down, stairs up, walking, jogging, standing, sitting
RealWorld		15	8	climbing down, climbing up, jumping, lying, running, sitting, standing, walking
HARTH	5 s	22	9	walking, running, shuffling, stairs up, stairs down, standing, sitting, lying, cycling

Appendix

Table 3: Overview of the evaluated datasets, grouped by window length. All signals are tri-axial accelerometry, downsampled to 50Hz (if their original frequencies are higher). The overlap is set to 50% for the 2s and 5s datasets, 20% for the 1s datasets, and 0 for Capture-24 and SimFL since they have a larger subject count.

Figure 4: Visualizing the class proportions of the downstream datasets. Many datasets have substantial class imbalance.

Foundation Model	Window (s)	Sampling rate (Hz)
Harnet-Oxford ( Yuan et al., 2024a )	5/10	30
BioPM ( Tarale et al., 2026 )	10	30
REBAR ( Xu et al., 2024 )	2.56	100
RelCon ( Xu et al., 2025b )	2/2.56	100
Oxford-Sleep FM ( Yuan et al., 2024b )	30	30

Appendix

Table 4: Input configurations for accelerometer FMs.

	Encoder				Decoder				Parameters
Model	Layers	Width	MLP	Heads	Layers	Width	MLP	Heads	Encoder	Total
Small	4	128	256	4	2	64	128	2	0.70M	0.95M
Medium	6	256	512	8	2	128	256	4	3.82M	4.81M
Large	8	384	768	8	2	192	384	4	10.94M	13.17M

Appendix

Table 5: Model size variants of VarMAE . All sizes use 0.2 s patches (10 samples at 50 Hz) and a single-query tokenizer with 4 heads.

	50 Hz	25 Hz †	10 Hz	5 Hz †
OpenWatch (1 s)
All losses	0.626	0.595	0.581	0.410
w/o $\mathcal{L}_{\mathrm{Cross}}$	0.606 ( $-$ 0.020)	0.534 ( $-$ 0.061)	0.571 ( $-$ 0.010)	0.332 ( $-$ 0.078)
w/o $\mathcal{L}_{\mathrm{Short}}$	0.632 (+0.006)	0.528 ( $-$ 0.067)	0.435 ( $-$ 0.146)	0.147 ( $-$ 0.263)
w/o $\mathcal{L}_{\mathrm{Long}}$	0.612 ( $-$ 0.014)	0.588 ( $-$ 0.007)	0.569 ( $-$ 0.012)	0.456 (+0.046)
MobiAct (2 s)

Appendix

Table 6: Evaluating the impact of each loss component on performance. Bold: best within each dataset and rate. † Rate not used in pretraining.

	50 Hz	25 Hz ∗	10 Hz	5 Hz ∗
WatchHand (1 s)
default (long 0.5, short 0.75)	0.668	0.664	0.647	0.620
long 0.25	0.667 ( $-$ 0.001)	0.661 ( $-$ 0.003)	0.644 ( $-$ 0.003)	0.609 ( $-$ 0.011)
long 0.75	0.667 ( $-$ 0.001)	0.653 ( $-$ 0.011)	0.642 ( $-$ 0.005)	0.633 (+0.013)
short 0.25	0.651 ( $-$ 0.017)	0.637 ( $-$ 0.027)	0.626 ( $-$ 0.021)	0.450 ( $-$ 0.170)
short 0.5	0.666 ( $-$ 0.002)	0.663 ( $-$ 0.001)	0.648 (+0.001)	0.381 ( $-$ 0.239)

Appendix

Table 7: Evaluating the impact of the masking ratio of the long and short views. Bold: best within each dataset and rate. ∗ Rate not used in pretraining.

Explore similar work

Inertia-1: An Open Exploration of Wearable Motion Foundation Models
Jul 7, 2026Zongzhe Xu, Aakarsh Anand, Sarah Jiang +4Human Activity Recognition Wearable Sensing
University of California, Los Angeles · Duke University · John Hopkins University
Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data
Mar 11, 2026Prithviraj Tarale, Kiet Chu, Abhishek Varghese +4Human Activity Recognition Self-Supervised Time Series Representation Learning
College of Information and Computer Sciences, University of Massachusetts, Amherst, United States · Stevens Institute of Technology, Hoboken, United States · Google Health, Seattle, United States +1
W-JEPA: Wavelet Joint-Embedding Predictive Architecture for Wearable Activity Recognition
Oct 14, 2026Yogesh Prabhu, Loki Natarajan, Jingjing ZouHuman Activity Recognition Representation Learning
Department of Computer Science and Engineering, University of California San Diego (UCSD). · Herbert Wertheim School of Public Health and Human Longevity Science, University of California San Diego (UCSD). · UCSD Shiley-Marcos Alzheimer’s Disease Research Center. +1

Jul 7, 2026J/K move · Enter open · S save

Inertia-1: An Open Exploration of Wearable Motion Foundation Models

Zongzhe Xu, Aakarsh Anand, Sarah Jiang, Chuntung Zhuang+3

University of California, Los Angeles · Duke University · John Hopkins University

Human Activity Recognition Wearable Sensing Motion Representation Learning Time-Series Foundation Models

Open paper AskPDF arXiv

Wearable motion sensing provides a continuous and scalable window into human behavior and health, making it a natural fit for foundation models, yet its pretraining and scaling principles remain poorly understood. Prior work studies isolated design choices, such as sensor placement or sampling frequency, often under fixed settings and narrow downstream tasks that fail to capture real-world sensing diversity. We introduce Inertia-1, a fully open exploration of wearable motion foundation models. Using massive corpora of accelerometer data from global sources spanning more than 18.2M hours, we build a controlled framework for studying the full lifecycle of wearable motion foundation models, covering data choices such as sensor modality, device placement, sampling rate, window length; model choices such as architectures and model size; and training choices such as pretraining objective and data scale. Extensive evaluations across 15 datasets spanning human activity recognition, freezing-of-gait detection, and disease prediction reveal intriguing findings for building motion foundation models that generalize across tasks and sensing conditions. Collectively, Inertia-1 not only presents state-of-the-art recipes for diverse downstream tasks, but also serves as a comprehensive, practical, and open cookbook for wearable motion representation learning.

Abstract

Explore similar work

Inertia-1: An Open Exploration of Wearable Motion Foundation Models

Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

W-JEPA: Wavelet Joint-Embedding Predictive Architecture for Wearable Activity Recognition