acceptodds
Under review as a conference paper at ICLR 2027

WearableBridge: Task-Guided Latent Transport for Free-Living Activity Recognition

Abstract

Wrist-worn accelerometers make activity monitoring feasible at scale, but incidental arm motion, device orientation, and shifting context blur the signal that separates one activity from another. Calibration studies supply higher-quality reference measurements such as concurrent thigh recordings or supervised laboratory wrist sessions, but they are expensive, cover few participants, and are rarely available at deployment. How can limited reference data improve recognition for participants who wear only a wrist device? We propose WearableBridge, an algorithm that learns a task-guided transport map from the wrist-feature distribution toward the reference-feature distribution. WearableBridge holds the reference encoder and classifier fixed and learns a map, the bridge, that carries each wrist feature into the region the classifier associates with its activity. Reference features or the reference classifier's predictions indicate the destination, and a classification loss penalizes wrist features that land in the region of another activity. We instantiate WearableBridge with three bridges that differ in how the reference guides the map: Mapper regresses onto reference features, Distillation matches the reference classifier's class probabilities, and MeanFlow learns a one-step flow from wrist toward reference features. On 40 held-out free-living PAAWS participants, all three bridges improve macro-F1 in every training run, and the strongest, MeanFlow, raises mean macro-F1 from 35.1% to 56.0% with concurrent thigh references (Thigh2Wrist), and from 54.0% to 63.5% with group-matched laboratory wrist references (Lab2Freeliving). Fitted only on PAAWS, MeanFlow also raises macro-F1 on the CAPTURE-24 cohort from 42.8% to 54.7% without using its labels. The gains replicate on SELFback, hold across four reference encoders, and grow further with temporal context. Reference recordings from a calibration subset can therefore improve wrist-only recognition for new participants while the reference model stays unchanged.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
transfer learning, domain adaptation, wearable sensors, human activity recognition, free-living monitoring, flow matching, knowledge distillation, learning with privileged information

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CNjcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper focus on cross-domain human activity recognition using accelerometer sensors, which is a reasonable and realistic problem. To address the domain gap between training and testing data, the paper presents WearableBridge. This framework learns a feature mapping from free-living wrist data to the reference feature space with calibration data. The paper investigates three mapping approaches, including Mapper, Distillation, and MeanFlow, and reports that MeanFlow achieves the strongest performance. Experiments on PAAWS, SELFback, and CAPTURE-24 shows that the proposed framework improves cross-domain human activity recognition performance over the unadapted-reference baseline.

Critical strengths

[+] The paper studies a realistic and important cross-domain human activity recognition problem.

[+] The writing is overall clear, and the paper is easy to follow.

Critical weaknesses

[-] There is a noticeable gap between the paper’s motivation and its implementation setting. A major motivation of this work is the limited availability of high-quality reference data due to high data collection cost (sec 1, para 2). However, the main PAAWS experiments rely on a relatively large calibration dataset of 170 training participants and 43 validation participants. The bridge is trained using 200K+ paired samples (that's hundreds of hours of calibration data!). Given that these calibration dataset requires concurrent thigh recordings or laboratory reference recordings, collecting calibration data at this scale is very difficult in practical free-living deployments.

[-] There are critical concerns regarding the experimental evaluation. The most evident issue is the lack of comparisons with representative external baselines and prior SOTA methods. The evaluation mainly compares WearableBridge against internal or relatively simple training baselines. As a result, it does not establish whether WearableBridge provides a meaningful advantage over existing methods that address similar cross-domain HAR settings.

Questions for authors
  1. Please clarify the reasons for limited baselines selection. The paper discusses domain adaptation, cross-position HAR, cross-modal distillation, and privileged-information learning as closely related directions, but the evaluation does not include representative external baselines from these areas. Are there technical reasons why these methods cannot be fairly adapted to the proposed setting? If not, how does WearableBridge compare against them under matched data and backbone settings?

  2. Please clarify the “limited reference data” setting. A key motivation of the paper is that high-quality reference data are limited and expensive to collect. However, the main PAAWS experiments use 170 training participants and 43 validation participants, with 224,697 Thigh2Wrist training pairs or 137,299 Lab2Freeliving training pairs. Could the authors clarify what regime they consider to be “limited reference data,” and provide evidence that the method remains effective when the number of calibration participants or reference pairs is substantially reduced?

4 comments
AuthorsOpenReview ↗

Response to Reviewer CNjc (1/2)

We thank the reviewer for the careful review. We aim to provide a useful method for practical applications, motivated by our own work in wearable health monitoring. The reviewer's comments prompted new experiments, and their results answer each concern directly. We would be grateful if the reviewer would consider raising the score in light of these results, which we will add to the manuscript.

1. Does WearableBridge remain strongest against external methods under matched settings? (Weakness 2, Question 1)

Yes. Under the same frozen encoder and adapter architecture, Full MeanFlow outperforms generalized distillation and Deep CORAL, representative methods from distillation, privileged-information learning, and domain adaptation, in both designs, with all four CIs above zero.

Following the reviewer's suggestion, we adapted both methods to our setting with the same data and backbone. Generalized distillation unifies knowledge distillation and privileged-information learning, and Deep CORAL is a domain adaptation method. Both follow their published objectives (generalized distillation, Deep CORAL) with trainable classifiers and loss weights selected on development data, and each has a control without its extra loss. Cross-position HAR methods adapt across body positions by aligning feature distributions (Chang et al., 2020, and Chen et al., 2019), so Deep CORAL represents this direction as well.

Method (three seeds each)Thigh2Wrist macro-F1 (%)Lab2Freeliving macro-F1 (%)
Distillation without the teacher-prediction loss (control)53.55 ± 0.5461.62 ± 0.43
Generalized distillation53.52 ± 0.2762.43 ± 1.46
CORAL without the covariance-matching loss (control)53.55 ± 1.0856.96 ± 0.32
Deep CORAL53.38 ± 0.1657.16 ± 0.24
Full MeanFlow55.98 ± 0.5163.54 ± 1.23
Full MeanFlow minus generalized distillation, pp [95% CI]+2.46 [1.62, 3.98]+1.11 [0.25, 2.33]
Full MeanFlow minus Deep CORAL, pp [95% CI]+2.60 [1.81, 3.91]+6.38 [0.87, 8.80]

Against generalized distillation, Full MeanFlow improves 40 of 40 participants in Thigh2Wrist and 36 of 40 in Lab2Freeliving. The Lab2Freeliving gain over Deep CORAL comes mainly from cycling (47.21% versus 20.59% F1). Per participant, Full MeanFlow improves 17 of 40 with a mean gain of 0.025 pp, and we will report this alongside the pooled result in the revised manuscript.

WearableBridge also works on top of any frozen encoder and its classifier, foundation models included, and improves recognition with every encoder we tested. With HARNet10, self-supervised on UK Biobank wrist data, MeanFlow fitted on PAAWS raises CAPTURE-24 macro-F1 from 42.75% to 54.73% without target labels. With DeepConvLSTM, InceptionTime, and HART encoders, every bridge improves mean macro-F1 by 9.62 to 20.15 pp in Thigh2Wrist and 8.13 to 12.76 pp in Lab2Freeliving. WearableBridge therefore remains the strongest method under matched settings and adds value on top of the field's established encoders.

AuthorsOpenReview ↗

Response to Reviewer CNjc (2/2)

2. Does the method remain effective with substantially less calibration data? (Weakness 1, Question 2)

Yes. With 40 calibration participants in total, about one fifth of the original pairs, both bridge versions beat the direct wrist classifier in all three runs of both designs.

“Limited reference data” means that reference recordings exist for a calibration subset of participants, and a calibration cohort of about 200 participants is realistic for clinical studies in our experience. The full PAAWS calibration set has 170 training and 43 validation participants, with 224,697 Thigh2Wrist or 137,299 Lab2Freeliving training pairs. The Thigh2Wrist pairs come from 113,091 ten-second thigh windows, nearly all paired with both wrists, which is about 314 hours in total and under two hours of concurrent recording per training participant. To test smaller cohorts, we ran two budgets that count validation participants and retrained every component learned from PAAWS within each. Each budget reduces the reference recordings and the wrist labels together, as a smaller calibration study would. With 32 training and 8 validation participants, each run trains on 41,662 to 43,670 Thigh2Wrist pairs or 25,868 to 26,838 Lab2Freeliving pairs. With 8 training and 2 validation participants, the counts fall to 10,192 to 11,722 and 6,018 to 7,174.

DesignTrain + validationDirect wrist classifierMeanFlow, classifier fixedMeanFlow, classifier trainable
Thigh2Wrist8 + 252.14 ± 2.1052.63 ± 3.3651.73 ± 3.80
Thigh2Wrist32 + 853.81 ± 0.7656.34 ± 1.5256.47 ± 1.46
Lab2Freeliving8 + 253.64 ± 1.5855.90 ± 2.9856.60 ± 2.64
Lab2Freeliving32 + 857.68 ± 0.5860.37 ± 0.7760.99 ± 0.83

With 40 participants and a trainable classifier, the gains over the direct wrist classifier are 2.66 pp [1.00, 3.67] in Thigh2Wrist and 3.30 pp [0.39, 4.39] in Lab2Freeliving, and with the classifier fixed, the Thigh2Wrist mean (56.34%) is close to the full-cohort mean (55.98%). With only 10 participants, Lab2Freeliving keeps a gain of 2.96 pp [0.58, 4.27], and in Thigh2Wrist the three methods score within one point of each other. The bridge therefore keeps its advantage with one fifth of the calibration data, and in Lab2Freeliving with one twentieth.

Metrics. Macro-F1 (%) is the mean ± SD over three runs, evaluated on the 40 held-out free-living PAAWS participants of each design with the encoder frozen. In the smaller-cohort experiments, each run also draws its own calibration participants. Gains are in percentage points (pp). The 95% CIs use 5,000 paired bootstrap resamples of the test participants, with the fitted models held fixed.

Thank you for your rebuttal. The additional results address some of my concerns, but my overall assessment remains largely unchanged. In my view, appropriate comparisons with external baselines are fundamental for an ICLR-level paper and should ideally have been included in the original submission. Meanwhile, I find the technical contribution and presentation to be below the expected standard for an ICLR paper. Overall, I will still maintain my rating.

AuthorsOpenReview ↗

Thank you for your thoughtful discussion

Thank you for reviewing the additional results. We agree that appropriate external comparisons are essential. Your comments motivated the additional matched experiments (results pasted below for convenience), which provide direct evidence for the method’s benefit. We will promptly incorporate the matched comparisons and clarify the method’s contribution in the revised manuscript.

Full MeanFlow minus comparator, macro-F1 pp [95% CI]Thigh2WristLab2Freeliving
Generalized distillation+2.46 [1.62, 3.98]+1.11 [0.25, 2.33]
Deep CORAL+2.60 [1.81, 3.91]+6.38 [0.87, 8.80]
Activity-only bridge+2.768 [1.777, 4.486]+1.694 [0.769, 3.043]

Our technical contribution is a task-guided latent transport formulation that learns from richer reference recordings. Reference features provide useful activity cues alongside sensor-specific details that the wrist may not capture. We therefore combine reference guidance with an activity-prediction loss, training the wrist-feature map both to use the reference and to recognize activities. Once trained, the bridge requires wrist data alone. Full MeanFlow achieves higher macro-F1 than the activity-only bridge, generalized distillation, and Deep CORAL in both designs.

Our experience with public and clinical datasets has repeatedly shown us how difficult wrist-only free-living recognition remains. This difficulty, along with the intuition above, motivated our effort to use this novel approach to explain the problem through representations and develop methods with measurable benefits. Together, our results establish a concrete contribution: using thigh or lab recordings during calibration to improve free-living recognition for new participants monitored only at the wrist.

We also believe this opens a promising direction for further research: learning more effectively from complementary reference recordings during training while keeping deployment simple and wrist-only.

Reviewer_ha7HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses how limited higher-quality reference data, available only for a calibration subset of participants, can improve wrist-only activity recognition for new participants who have no reference recording. It considers two calibration designs, Thigh2Wrist and Lab2Freeliving. The proposed method, WearableBridge, keeps the reference encoder and classifier fixed and learns only a small map, the bridge, that transforms wrist features within the reference feature space. Its objective combines a cross-entropy loss through the frozen classifier with a reference-guidance term.

Critical strengths
  1. The paper addresses a realistic setting in which only a calibration subset of participants has reference recordings, covering both instance-level (Thigh2Wrist) and group-level (Lab2Freeliving) references.
  2. The evaluation uses free-living recordings from 40 held-out participants with a participant-level bootstrap, and generalization is examined with additional reference encoders and external datasets (SELFback and CAPTURE-24).
Critical weaknesses
  1. The headline improvement (35.1% to 56.0%) is measured against the unadapted reference, which uses no wrist labels, whereas the activity-only bridge trained on the same labels already reaches 54.10%. The additional gain of MeanFlow is observed only in a single matched run on PAAWS, is reversed on SELFback, and cannot be assessed in Lab2Freeliving, where this comparison is not reported.
  2. Although the paper is motivated by limited reference data, no results are reported for varying amounts of reference data or wrist labels, so any advantage in sample efficiency remains unclear.
Questions for authors
  1. In Lab2Freeliving, how can the improvement from 54.00% to 63.54% be attributed separately to adaptation with labeled SimFL wrist windows and to the reference guidance?
  2. What practical advantage does keeping the reference classifier fixed offer over retraining it, for example in generalization, sample efficiency, or deployment? Are there results from a comparison with matched model capacity?
  3. Does “limited reference data” refer to the setting in which only a calibration subset of participants has reference recordings, or does it also imply that the method remains effective with a small amount of calibration data?
3 comments
AuthorsOpenReview ↗

Response to Reviewer ha7H (1/2)

We thank the reviewer for the careful review. We aim to provide a useful method for practical applications, motivated by our own work in wearable health monitoring. The reviewer's comments prompted new experiments, and their results answer each concern directly. We would be grateful if the reviewer would consider raising the score in light of these results, which we will add to the manuscript.

1. Does reference guidance add accuracy beyond wrist labels, in both designs? (Weakness 1, Question 1)

Yes. Full MeanFlow, which adds the reference-flow loss to a bridge trained on the same wrist labels, exceeds the activity-only bridge by 2.77 pp in Thigh2Wrist and 1.69 pp in Lab2Freeliving, in all three seeds.

Activity labels are not always reliable in our setting, which is why the bridge does not rely on them alone. The reference-flow loss adds a second training signal by training a path from each wrist feature toward its paired reference feature. The activity-only bridge and Full MeanFlow below share the architecture, labels, data, and training schedule.

Method (three seeds each)Thigh2Wrist macro-F1 (%)Lab2Freeliving macro-F1 (%)
Reference classifier without a bridge35.07 ± 0.2154.00 ± 1.29
Activity-only bridge53.21 ± 0.9061.85 ± 0.63
Full MeanFlow, with reference-flow loss55.98 ± 0.5163.54 ± 1.23
Full MeanFlow minus activity-only bridge, pp [95% CI]+2.768 [1.777, 4.486]+1.694 [0.769, 3.043]

Both CIs lie above zero. In Lab2Freeliving, labeled SimFL adaptation accounts for 7.85 pp of the gain from 54.00% to 63.54%, and reference guidance adds 1.69 pp. The single-seed activity-only result in the submission (54.10%) is 53.21 ± 0.90% over three seeds, and we will revise the manuscript to report the gain from reference guidance separately from the total gain. On selfBACK (seven test participants), the activity-only bridge has the higher mean (67.76 ± 1.34% versus 66.74 ± 0.38%). There the measured thigh (64.42%) scores below every bridge, so pulling wrist features toward thigh features also carries the thigh's errors, and scripted sessions give reliable labels for the activity loss alone. Reference guidance therefore makes a consistent contribution beyond wrist labels in both PAAWS designs.

2. Does the method remain effective with less calibration data? (Weakness 2, Question 3)

Yes. With 40 calibration participants in total, about one fifth of the original pairs, both bridge versions beat the direct wrist classifier in all three runs of both designs.

“Limited reference data” means that reference recordings exist for a calibration subset of participants, and a calibration cohort of about 200 participants is realistic for clinical studies in our experience. The full PAAWS calibration set has 170 training and 43 validation participants, with 224,697 Thigh2Wrist or 137,299 Lab2Freeliving training pairs. To test smaller cohorts, we ran two budgets that count validation participants and retrained every component learned from PAAWS within each. Each budget reduces the reference recordings and the wrist labels together, as a smaller calibration study would. With 32 training and 8 validation participants, each run trains on 41,662 to 43,670 Thigh2Wrist pairs or 25,868 to 26,838 Lab2Freeliving pairs. With 8 training and 2 validation participants, the counts fall to 10,192 to 11,722 and 6,018 to 7,174.

DesignTrain + validationDirect wrist classifierMeanFlow, classifier fixedMeanFlow, classifier trainable
Thigh2Wrist8 + 252.14 ± 2.1052.63 ± 3.3651.73 ± 3.80
Thigh2Wrist32 + 853.81 ± 0.7656.34 ± 1.5256.47 ± 1.46
Lab2Freeliving8 + 253.64 ± 1.5855.90 ± 2.9856.60 ± 2.64
Lab2Freeliving32 + 857.68 ± 0.5860.37 ± 0.7760.99 ± 0.83

With 40 participants and a trainable classifier, the gains over the direct wrist classifier are 2.66 pp [1.00, 3.67] in Thigh2Wrist and 3.30 pp [0.39, 4.39] in Lab2Freeliving, and the Thigh2Wrist mean (56.47%) is close to the full-cohort mean (56.31%). With only 10 participants, Lab2Freeliving keeps a gain of 2.96 pp [0.58, 4.27], and in Thigh2Wrist the three methods score within one point of each other. The bridge therefore keeps its advantage with one fifth of the calibration data, and in Lab2Freeliving with one twentieth.

AuthorsOpenReview ↗

Response to Reviewer ha7H (2/2)

3. Does keeping the classifier fixed offer practical value at matched capacity? (Question 2)

Yes. Reference guidance adds 2.76 pp in Thigh2Wrist and 2.05 pp in Lab2Freeliving even with a trainable classifier, and the fixed classifier trails the trainable one by at most 0.7 pp in every setting while one reference model serves all later participants.

Freezing the classifier is optional, so a study can choose either. The first two rows below train the same components, so the reference-flow loss is their only difference.

MeanFlow with the full calibration cohort, encoder frozenThigh2Wrist macro-F1 (%)Lab2Freeliving macro-F1 (%)
Classifier trainable, without reference-flow loss53.55 ± 0.5461.62 ± 0.43
Classifier trainable, with reference-flow loss56.31 ± 0.7863.67 ± 0.79
Classifier fixed, with reference-flow loss55.98 ± 0.5163.54 ± 1.23
Trainable classifier, with minus without reference-flow loss, pp [95% CI]+2.76 [1.71, 4.43]+2.05 [1.15, 3.33]
  • Generalization. In our experience, models trained on cleaner data, such as thigh or laboratory recordings, often outperform models trained or pretrained on noisier data, even when the noisier data are more plentiful. This holds especially when the noise comes from structured artifacts (non-white noise). With six activities, including the rarer lying and stairs, Full MeanFlow with the fixed thigh classifier exceeds the direct wrist classifier by 3.49 pp [2.58, 4.34] and improves 39 of 40 participants.
  • Sample efficiency. With 8 training and 2 validation participants in Thigh2Wrist, the fixed classifier is 0.90 pp ahead [0.55, 1.29], and training the classifier gives slightly higher means in the other settings.
  • Deployment. The reference classifier is trained once, on thigh or laboratory recordings from the calibration sample. Because it stays fixed, every later participant, with or without reference recordings, is classified by the same activity boundaries, so activity estimates stay comparable across participants and studies. A new wrist device then needs only a small bridge fitted on its own calibration data, and the reference classifier stays unchanged. A strong reference model thus becomes a foundation for future use.
  • Flexibility. A study can still train a new classifier when needed, as in the trainable rows of the table. The main text (“Why keep the classifier frozen?”) shows that a sufficiently expressive bridge followed by the fixed classifier can reproduce any classifier trained on wrist features, so a good bridge already contains the capacity of a new classifier. Keeping the classifier fixed costs little accuracy, at most 0.7 pp in every setting.

Metrics. Macro-F1 (%) is the mean ± SD over three runs, evaluated on the 40 held-out free-living PAAWS participants of each design with the encoder frozen. In the smaller-cohort experiments, each run also draws its own calibration participants. Gains are in percentage points (pp). The 95% CIs use 5,000 paired bootstrap resamples of the test participants, with the fitted models held fixed.

Thank you for the additional experiments and the detailed response. The new comparisons clarify the contribution of reference guidance in both PAAWS settings. The finding that the proposed method outperforms the direct wrist classifier even with 40 calibration participants also partly addresses my concern about sample efficiency. However, the 40-participant experiments do not include a direct comparison with the activity-only bridge, so it remains unclear how much of the gain in that setting comes from reference guidance. As noted in my original review, the activity-only bridge also achieved a higher mean score on selfBACK. Given the remaining uncertainty about when reference guidance provides an advantage, I am maintaining my rating.

Reviewer_TKw1OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WearableBridge that improves wrist-only activity recognition by learning a small bridge that remaps wrist features so that a frozen classifier trained on higher-quality reference recordings can read them. The paper evaluates three bridge objectives, i.e., feature matching, prediction distillation, and one-step MeanFlow, using either concurrent thigh recordings or group-matched laboratory wrist recordings during calibration, while requiring only a wrist sensor at deployment. The evaluation on PAAWS, SELFback and CAPTURE-24 show improved macro-F1 over applying the reference classifier directly to wrist features, and the gains replicate on SELFback and transfer to CAPTURE-24. However, because the bridge is trained with wrist activity labels and an activity-only bridge captures much of the improvement, the method may operate primarily as a supervised wrist classifier expressed through a frozen reference classifier, with reference guidance serving as a regularizer.

Critical strengths
  1. The paper targets a realistic scenario where reference sensors or lab sessions are available only for a calibration subset, while deployment uses a wrist device alone.
  2. The method is evaluated extensively across multiple datasets, reference modalities, encoder architectures, and held-out participants, including transfer to CAPTURE-24 without using its labels. I appreciate that the authors also report comparators that could weaken their own claims (activity-only bridge, direct wrist retraining, joint fine-tuning).
Critical weaknesses
  1. Because the bridge is trained using activity labels from the wrist data, it may simply learn to classify wrist signals rather than meaningfully learn from the higher-quality reference sensor. Therefore, it is unclear whether the reference data itself is responsible for the improvement. The ablations do not adequately isolate the contribution of reference guidance from bridge architecture and ordinary supervised learning. The activity-only bridge captures most of the improvement and even performs best on SELFback, while the direct wrist classifier remains competitive.
  2. The method assumes wrist and reference features are directly comparable, but never tests this. Thigh and wrist sensors measure different body segments. Thigh posture, for example, has no direct wrist counterpart. Yet the method treats their features from a single frozen encoder as points in a common space. This assumption seems to be the basis for MeanFlow's straight-line path between wrist and thigh features, the term keeping outputs near the wrist input, and the paper's interpretation of the bridge as transport toward more informative reference features. The low unadapted baseline suggests the two sensors' features are poorly aligned but alignment isn't examined directly. Without such evidence, the transport interpretation is unsupported. It is also somewhat weakened by the paper's own result that a bridge can emulate any wrist classifier.
Questions for authors
  1. Can you report multi-seed results for the activity-only bridge and the direct wrist classifier on PAAWS? Can you also add a bridge trained with only the reference term and no activity labels? This would show how much of the gain comes from the reference data itself.
  2. Have you measured whether wrist and thigh features from the same moment or activity are closer to each other than to features of other activities in the frozen encoder space? Did MoCA's pretraining involve cross-placement reconstruction that would justify treating these features as comparable?
5 comments
AuthorsOpenReview ↗

Response to Reviewer TKw1 (1/3)

We thank the reviewer for the careful review, and for recognizing the importance of transparently reporting all relevant results, including those that offer additional perspectives on our findings (below we have provided detailed interpretations of these results to deepen our understanding of the underlying phenomena). We aim to provide a useful method for practical applications, motivated by our own work in wearable health monitoring. The reviewer's comments prompted new experiments, and their results answer each concern directly. We would be grateful if the reviewer would consider raising the score in light of these results, which we will add to the manuscript.

1. Do the reference recordings add accuracy beyond wrist-label supervision? (Weakness 1, Question 1)

Yes. With the wrist labels, bridge architecture, and training settings held fixed, Full MeanFlow, which adds the reference-flow loss, exceeds the activity-only bridge by 2.77 pp in Thigh2Wrist and 1.69 pp in Lab2Freeliving, in every seed, and also exceeds the direct wrist classifier.

Activity labels are not always reliable in our setting, which is why the bridge does not rely on them alone. The reference-flow loss adds a second training signal by training a path from each wrist feature toward its paired reference feature. The activity-only bridge and Full MeanFlow below share the architecture, labels, data, and training schedule.

Method (three seeds each)Thigh2Wrist macro-F1 (%)Lab2Freeliving macro-F1 (%)
Reference classifier without a bridge35.07 ± 0.2154.00 ± 1.29
Activity-only bridge53.21 ± 0.9061.85 ± 0.63
Full MeanFlow, with reference-flow loss55.98 ± 0.5163.54 ± 1.23
Direct wrist classifier, encoder frozen54.60 ± 0.2961.29 ± 0.70
Full MeanFlow minus activity-only bridge, pp [95% CI]+2.768 [1.777, 4.486]+1.694 [0.769, 3.043]
Full MeanFlow minus direct wrist classifier, pp [95% CI]+1.38 [0.59, 2.08]+2.25 [0.39, 3.32]

All four CIs lie above zero. The reference recordings therefore add accuracy that wrist labels alone do not provide. They also let a study reuse encoders and classifiers trained on high-quality thigh or laboratory recordings, which are common in this field.

On selfBACK (seven test participants), the activity-only bridge has the higher mean (67.76 ± 1.34% versus 66.74 ± 0.38%). Reference guidance pulls each wrist feature toward its concurrent thigh feature, so it helps when thigh features are classified more accurately than bridged wrist features. On PAAWS, the measured thigh reaches 82.50%, far above every bridge, and the pull points to the correct activity region. On selfBACK, the measured thigh reaches only 64.42%, below every bridge, so the pull also carries the thigh's errors into the wrist features. selfBACK activities are also scripted, so their labels are reliable and the activity loss alone suffices. Reference guidance adds the most when the reference is more informative than the wrist and labels are less reliable, as in free-living PAAWS, which tells a study when reference recordings are worth collecting.

2. Can the reference data help without any activity labels? (Question 1)

Yes, when the recordings themselves provide the pairing. In Thigh2Wrist, where wrist and thigh are recorded simultaneously, a bridge trained without any wrist labels improves macro-F1 by 2.40 pp over the reference classifier used without a bridge.

We ran two versions. The first removes only the activity-label loss, so wrist labels still enter through sample weights and, in Lab2Freeliving, through pairing and checkpoint selection. The second removes wrist labels from every bridge-training and selection step.

Bridge trainingThigh2Wrist macro-F1 (%)Lab2Freeliving macro-F1 (%)
Reference classifier without a bridge (comparator)35.07 ± 0.2154.00 ± 1.29
Remove only the activity-label loss35.96 ± 1.2358.65 ± 2.18
Remove wrist labels from all bridge-training and selection steps37.47 ± 1.9153.09 ± 0.85
Bridge trained without wrist labels minus no bridge, pp [95% CI]+2.40 [2.05, 2.78]−0.92 [−1.48, −0.44]

The Thigh2Wrist gain comes from the reference recordings alone. In Lab2Freeliving, the laboratory references are recorded in separate sessions, and activity labels are what pair them with wrist windows. Without labels, pairs join windows from the same participant and wrist side but often of different activities, so the pairing loses its structure, and lower accuracy is the natural result. Combining the cross-entropy term with the reference-flow loss therefore keeps the method flexible, so a study can train the bridge with labels, paired references, or both, whichever its calibration design provides.

AuthorsOpenReview ↗

Response to Reviewer TKw1 (2/3)

3. Are wrist and thigh features comparable? (Weakness 2, Question 2)

Yes. One frozen encoder maps wrist and thigh windows into one feature space, and there a wrist window is closer to thigh windows of the same activity than to those of other activities, and closest to the thigh window recorded at the same moment.

Two reasons led us to expect this. Wrist and thigh move under one coordinated body motion, so their signals during an activity are not independent. The bridge then does the remaining work: it moves each wrist feature toward its reference and aims to remove the artifacts, such as incidental arm motion and device orientation, that separate the two. We tested this on 6,029 wrist queries from 8,192 wrist-thigh pairs across 40 participants. Comparisons stay within the same participant and day, and windows from a different time are at least 60 seconds from the query.

Distance from a wrist window to thigh windowsDifference in standardized MSE [95% CI]
Different activity minus same activity, both from a different time+0.2500 [0.2145, 0.2853]
Same activity at a different time minus the simultaneous window+0.0842 [0.0732, 0.0948]

Both differences are positive, and their CIs exclude zero, so a wrist feature lies closer to thigh features of the same activity and closest to the one recorded at the same moment. Cosine distance gives the same result. The next table shows how each bridge changes the distance to the simultaneous thigh feature.

Wrist representationMean standardized MSE to the simultaneous thigh feature
Unadapted wrist features2.158
Activity-only bridge output3.511
Full MeanFlow output1.877

Full MeanFlow moves each wrist feature closer to its simultaneous thigh feature (standardized MSE from 2.158 to 1.877), and the activity-only bridge moves it farther away (3.511). This closeness is specific to each wrist window and its own thigh recording. When activity is read off the nearest thigh window of another participant, the activity-only bridge scores higher (56.91% versus 52.31% accuracy).

The added results also address two further concerns in Weakness 2.

  • “Thigh posture, for example, has no direct wrist counterpart.” This is the information the thigh adds. The bridge needs only that wrist windows of each activity lie closer to thigh windows of that activity than to those of other activities, which the distance table above shows, and it learns to place wrist windows where the thigh places their activity. With Full MeanFlow, recall rises from 56.3% to 78.7% for sitting and from 32.9% to 56.0% for standing, two postures the thigh separates and the wrist confuses.
  • “The low unadapted baseline suggests the two sensors' features are poorly aligned.” Alignment and classification measure different things. Alignment, in the distance table above, uses all feature dimensions: on average, a wrist window lies closer to thigh windows of its own activity than to those of other activities (+0.2500 [0.2145, 0.2853] in standardized MSE). Classification uses only a few: with four activities, the linear thigh classifier decides along at most three directions of the feature space. A shift of wrist features along these directions can move many windows into another activity's region while changing the overall distances little. The unadapted classifier, for example, labels 56.6% of wrist walking windows as cycling, which is why the baseline is low (35.07%). The bridge corrects this shift: Full MeanFlow lowers the share to 1.4% and raises macro-F1 to 55.98%.

These results also address the three parts of the method that the reviewer ties to this assumption. The straight-line path is well defined because the shared encoder and standardization place wrist and thigh features in common coordinates, and the measurements above show that these coordinates correspond. The term that keeps outputs near the wrist input regularizes the size of each change. The result that a bridge can emulate any wrist classifier states which predictions are representable, and the measurements supply the evidence of alignment. Wrist and thigh features are therefore comparable, and Full MeanFlow makes them more so.

AuthorsOpenReview ↗

Response to Reviewer TKw1 (3/3)

4. Did MoCA pretraining include cross-placement reconstruction? (Question 2, clarification)

No. MoCA was pretrained on waist-worn accelerometer and gyroscope channels from UCI-HAR, with no wrist-thigh reconstruction. We use its public checkpoint, frozen throughout, because it has proven useful and accurate across a variety of datasets in our experience. Its pretraining builds in no correspondence between wrist and thigh, so our case for comparing them rests on direct evidence. The same body motion drives both sensors, and point 3 measures the result in the frozen encoder space: wrist features lie closer to thigh features of the same activity and closest to those recorded at the same moment.

(Metrics used in this response. Macro-F1 (%) is the mean ± SD over three training seeds, evaluated on the 40 held-out free-living PAAWS participants of each design. Gains are in percentage points (pp). The 95% CIs use 5,000 paired bootstrap resamples of the test participants, with the fitted models held fixed.)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.