Variable Input Masked Autoencoders for Wrist-Worn Accelerometry

1 INTRODUCTION

• Wearable activity tasks require different window durations and sampling rates, while existing accelerometer foundation models typically fix both during pretraining and evaluate in similar configurations.Examples range from one-second, 100 Hz gesture windows to ten-second, 30 Hz daily-activity windows.Source: §1 · p. 1
• VarMAE trains on long, high-frequency and short, lower-frequency views of the same signal, reconstructing masked regions within each view and across resolutions.The short view is cropped from the long view; visible long-view regions condition predictions of masked short-view tokens.Source: §1 · p. 2
• Cropping removes surrounding context while downsampling removes fine-grained temporal information, and cross-resolution reconstruction bridges these complementary information bottlenecks.The setup exposes one encoder to multiple sampling rates and window durations during pretraining.Source: §1 · p. 2
• VarMAE retains a Transformer encoder-decoder while variable-rate tokenization maps patches to fixed-dimension tokens and a variable-rate head reconstructs at the target sampling rate.Patches are defined over fixed durations, so their sample counts vary with sampling rate.Source: §1 · p. 2
• With 3.82M parameters, VarMAE matches or outperforms specialist foundation models on all eight datasets, while the 0.7M version outperforms general-purpose time-series models 30–550× larger.The evaluation spans human activity recognition datasets with varied activities, subject counts, recording conditions, and input configurations.Source: §1 · p. 2

2 RELATED WORK

• Wearable foundation models target applications such as gait analysis, fall detection, and activity recognition, while self-supervised learning can reduce reliance on annotations and produce representations robust to class imbalance.The passage describes reconstruction and contrastive learning as dominant self-supervised paradigms.Source: §2 · p. 2
• Accelerometer self-supervision includes augmentation prediction, contrastive learning, and reconstruction, including masked reconstruction in masked autoencoders.Contrastive approaches use varied positive-pair strategies, while reconstruction methods recover inputs or intentionally masked portions.Source: §2 · p. 3
• Prior accelerometer foundation models largely fix window size and sampling rate during pretraining and address configuration differences with configuration-specific pretraining; VarMAE trains one model across both.The paper states that VarMAE operates directly on heterogeneous configurations without adaptations or setting-specific pretraining.Source: §2 · p. 3

3 METHODOLOGY

• VarMAE’s core components are a variable-rate masked autoencoder and cross-resolution reconstruction of a short, downsampled view using context from a longer, higher-frequency view.The views are encoded independently by the same encoder.Source: §3 · p. 3, §3.2 · p. 5

3.1 VARIABLE-RATE MASKED AUTOENCODER

• Defining patches in seconds rather than samples keeps their temporal span fixed across rates, while a variable-rate tokenizer and reconstruction head let the shared encoder-decoder process and reconstruct differing sample counts.Each patch spans 0.2 seconds; the passage gives examples of 2 samples at 10 Hz and 10 samples at 50 Hz.Source: §3.1 · pp. 3–4
• The tokenizer uses cross-attention to map rate-dependent patch samples to a fixed-size token, and the decoder’s variable-rate head reconstructs the number of samples required at the target rate.Positions are specified in physical time rather than sample indices, and the encoder processes visible tokens.Source: §3.1 · p. 4

3.2 CROSS-RESOLUTION MASKED RECONSTRUCTION

• Training pairs a fixed 30-second, 50 Hz long view with randomly cropped short views of 1, 2, 5, or 10 seconds at 10–50 Hz, using cropping and downsampling as complementary information bottlenecks.Cropping removes surrounding context, while downsampling removes fine-grained temporal information.Source: §3.2 · p. 5
• Short-view masked patches are co-masked in the long view to prevent the cross-decoder from seeing its reconstruction targets directly, while retaining surrounding long-view context.The short view uses 75% masking in contiguous spans; additional long-view patches are masked outside the short-view interval to reach 50%.Source: §3.2 · p. 5
• The summed objective reconstructs masked samples within each view and reconstructs masked short-view tokens again using visible long-view tokens as context.The cross-resolution loss conditions its predictions on visible long-view tokens through the cross-decoder.Source: §3.2 · p. 5

4 EXPERIMENTAL SETTINGS

• Pretraining uses approximately 16k NHANES users, with about one week of wrist-accelerometer data per user, recorded at 80Hz and downsampled to 50Hz.Source: §4 · p. 6
• Evaluation spans eight HAR datasets covering gestures, locomotion, and daily activities, with variation in sensor location, participants, and classes.Source: §4 · p. 6

4.1 COMPARATIVE BASELINES

• Baselines include controlled pretraining on NHANES and public checkpoints, including Harnet-Oxford, BioPM, and general-purpose time-series FMs.Source: §4.1 · p. 6
• Controlled comparisons include AugPred, SimCLR, and Vanilla MAE; Vanilla MAE shares VarMAE’s encoder-decoder and patch duration, making it the most direct comparison.Source: §4.1 · p. 6
• Subject-disjoint evaluation uses five folds, with each user appearing in the test set once; performance is average macro-F1 across folds.Source: §4.1 · p. 6
• Linear evaluation freezes the pretrained encoder and trains one linear activity classifier, measuring representations without varying classifier capacity.Source: §4.1 · p. 6

5.1 ACTIVITY RECOGNITION PERFORMANCE

• At 5s, VarMAE-Small outperforms Harnet-Oxford by >8% despite using 0.7M rather than 10M parameters; VarMAE variants also outperform BioPM on SimFL and Capture-24.On PAAWS-SimFL, Harnet-Oxford scores 0.782 versus VarMAE-Small’s 0.762, while VarMAE-Medium and Large score 0.793 and 0.800.Source: §5.1 · p. 7
• VarMAE-Medium matches or outperforms all controlled baselines across datasets, while all VarMAE variants outperform general-purpose time-series FMs.With a frozen encoder and linear classifier, VarMAE-Small also outperforms end-to-end models on four datasets, rising to six datasets for larger variants.Source: §5.1 · p. 8

5.2 FEW-SHOT LEARNING EVALUATION

• VarMAE is competitive with or outperforms specialist FMs in few-shot evaluation, despite using one shared model while specialists match each dataset’s duration and sampling rate.Performance improves with more labeled data, especially above 25 windows per activity; AugPred and SimCLR often trail by more than 10% F1-score.Source: §5.2 · p. 8, Figure 2 · p. 8

5.3 ABLATIONS

• Removing LShort causes performance drops exceeding 25% F1-score at 5Hz; removing Lcross also consistently reduces performance, while LLong has modest impact.At 50Hz, removing any single loss has little effect; the impact grows as sampling rates decrease.Source: §5.3 · p. 9, Figure 3 · p. 9
• At 5Hz, lowering the short-view masking ratio from 0.75 to 0.5 or 0.25 drops performance by 14.5% or more; long-view ratio changes have substantially smaller effects.At 50, 25, and 10Hz, performance is largely insensitive to masking ratios in either view.Source: §5.3 · p. 9, Table 2 · p. 9
• High short-view masking ratios help retain effectiveness when evaluation provides less temporal resolution, with short-view choices mattering more after downsampling.Source: §5.3 · p. 9

6 CONCLUSION

• One VarMAE representation remains effective across substantially different sensing configurations, reducing the need for bespoke training at specific configurations.The evaluation spans eight activity-recognition datasets with varied activities, sensing locations, window durations, and subject counts.Source: §6 · p. 10
• VarMAE jointly reconstructs long, high-frequency and short, downsampled views, using the longer view as context for cross-resolution reconstruction.The model uses variable-rate tokenization and reconstruction with a shared Transformer encoder-decoder.Source: §6 · p. 10
• Generative AI tools assisted with code, grammar, and typo correction, but not idea development, method or experiment design, or result interpretation.The authors reviewed the AI-assisted work and take responsibility for the paper's contents.Source: AI USE STATEMENT · p. 10
• The authors identify privacy and potential health disparities as ethical concerns, and report that this study used publicly available data without personally identifiable information.They state that physiological-data collection, storage, and use should be subject to strict governance.Source: ETHICS STATEMENT · p. 10

A DATASETS

• Table 3 groups evaluated datasets by window length and reports tri-axial signals downsampled to 50Hz when their original frequency is higher.Window overlap is 50% for 2s and 5s datasets, 20% for 1s datasets, and 0 for Capture-24 and SimFL.Source: Table 3 · p. 15
• Many downstream datasets have substantial class imbalance, as shown by their class proportions.Source: Figure 4 · p. 15
• The datasets use 1s windows for short wrist gestures, 2s for smartphone locomotion, 5s for longer locomotion segments, and 10s for heterogeneous daily activities.The 10s datasets, Capture-24 and PAAWS-SimFL, also have the largest participant counts.Source: A DATASETS · pp. 15–16

B PRE-TRAINING AND EVALUATION SETTINGS

• Training uses AdamW with a 10−3 learning rate, no weight decay, up to 250k steps, and a 1024-window total batch across two GPUs.The schedule warms up for 20k steps and then uses cosine decay; training uses bf16 mixed precision.Source: B PRE-TRAINING AND EVALUATION SETTINGS · p. 16
• Training batches sample 64 users per GPU, then one day per user and eight 30s windows per day, favoring windows with higher standard deviation.This yields 512 windows per GPU; the sampling is intended to sample stationary periods less often.Source: B PRE-TRAINING AND EVALUATION SETTINGS · p. 16
• Specialist foundation models use 1, 2, 5, or 10-second crops from 30-second segments at 50Hz.Source: B PRE-TRAINING AND EVALUATION SETTINGS · p. 16
• VarMAE pairs a short view of randomly selected duration and input rate with its corresponding full 30s window sampled at 50 Hz.Short-view durations are selected from 1, 2, 5, and 10 seconds; rates are selected from 10, 20, 30, 40, and 50 Hz.Source: B PRE-TRAINING AND EVALUATION SETTINGS · p. 16
• Table 4 summarizes input configurations for accelerometer foundation models.Source: Table 4 · p. 16

B.1 BASELINE IMPLEMENTATION DETAILS

• Controlled-evaluation baselines use identical preprocessing and training settings within the experimental setup.Source: §B.1 · p. 16
• AugPred uses random permutation, time flipping, and time warping; SimCLR samples pairs of transforms including noise, scaling, rotations, negation, flipping, warping, and channel shuffling.SimCLR uses an InfoNCE temperature of 0.1.Source: §B.1 · pp. 16–17
• Vanilla MAE is the closest match to VarMAE, differing by lacking variable-rate tokenization and reconstruction heads.Its masking ratio is 0.5.Source: §B.1 · p. 17
• Public-checkpoint evaluation uses Harnet-Oxford and BioPM preprocessing and representation-extraction procedures, with BioPM's filtering and linear movement-element extraction pipeline also applied.Linear evaluation is then performed within the shared evaluation setup.Source: §B.1 · p. 17

• All size variants use 0.2 s patches (10 samples at 50 Hz) and a single-query tokenizer with 4 heads.Source: Table 5 · p. 17

• Table 6 evaluates the impact of each loss component on performance, with bold marking the best result within each dataset and rate.Rates marked † were not used in pretraining.Source: Table 6 · p. 18
• Table 7 evaluates the impact of the long- and short-view masking ratios, with bold marking the best result within each dataset and rate.Rates marked ∗ were not used in pretraining.Source: Table 7 · p. 18