acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

A major challenge in wearable device activity recognition is the heterogeneity across datasets in sensor placement, channel count, sampling rate and subjects. Moreover, the motion patterns that distinguish activities often appear as subtle, scale-specific variations within otherwise similar waveforms. As a result, raw signals are often ambiguous, and models trained on them tend to rely on cues specific to a particular sensor setup or participant, which limits both cross-dataset transfer and generalization to unseen participants. Addressing this requires representations that preserve such activity-relevant, scale-specific variation. We introduce W-JEPA, a self-supervised framework that exploits the multiscale structure of inertial motion and leverages a joint-embedding predictive architecture, by organizing windows into a fixed grid of wavelet tokens indexed by scale, sensor axis, and time. These wavelet tokens help expose activity-specific detail within broadly similar raw signals, and serve as the units of masked latent prediction. W-JEPA predicts masked token representations supplied by an exponential-moving-average (EMA) target encoder. Because recovering a masked token requires relating motion at one scale and axis to motion at other scales and axes, the model learns the cross-scale, cross-axis structure of human movement. Across six wearable activity benchmarks, W-JEPA outperforms state-of-the-art methods under random-window splits and remains the strongest when subjects are held out, improving average macro-F1 by 8.7 percentage points over the strongest published baseline. These results indicate that physically grounded multiscale tokenization, combined together with latent predictive learning, provides a promising foundation for transferable representations of heterogeneous wearable inertial data.

Primary area:
learning on time series and dynamical systems
Keywords:
Human activity recognition (HAR), Self-supervised learning, Wearable sensing, Wavelets, Joint-embedding predictive architecture (JEPA), Time-series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AFAfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

W-JEPA proposes wavelet-based tokens for self-supervised wearable activity recognition. The paper uses wavelet mean/RMS statistics as tokens for JEPA pretraining, then fine-tunes the model for wearable activity recognition. It evaluates on six datasets, with the target dataset excluded from pretraining and participants separated in the main downstream evaluation.

Critical strengths

I like the evaluation setting. Leaving the target dataset out of pretraining and testing on unseen participants makes the results more relevant to actual use. The gains are also quite good.

The paper is easy to follow. Rerunning the baselines with common preprocessing, splits and seeds is a good effort. The raw-patch ablation is useful too, since token count and embedding MLP depth/width are matched. The restricted-sensor experiments provide additional evidence that the approach works beyond the full-sensor setting.

Critical weaknesses
  1. The novelty seems mainly in how the tokens are constructed and masked. Using wavelet statistics as prediction units is a reasonable idea, but wavelet features for time-series are already established (e.g., Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization, Learning filter widths of spectral decompositions with wavelets), and the JEPA framework itself is largely standard. This also makes the missing training from scratch comparison important. Since all downstream results use end-to-end fine-tuning, could the same wavelet model already get similar results without pretraining?

  2. The wavelet tokens contain mean and RMS, while each raw token contains 50 ordered samples. So this comparison changes both the frequency decomposition and how much information is summarized. It shows that the complete tokenizer is useful, but doesn’t tell us which part matters most. I’d like to see a simple temporal mean/RMS tokenizer without wavelets.

  3. do the five seeds change the participant split, or just training randomness? Three datasets use only two participants for finetuning, so the choice of participants could matter quite a bit. These details need clarification before I can be confident about the comparisons. I’m not saying the baselines were treated unfairly, bur the paper just doesn’t give enough information to check this

Questions for authors

see above

1 comment

Thank you for the thoughtful review. We ran the component analyses you requested and documented the baseline setup, and every result confirms that each design choice in W-JEPA is justified and that the full model is strongest. All of these results will be added to the revised paper. Based on the strong results we present, we kindly ask you to  consider raising your score.

1. Is JEPA pretraining necessary? Yes. W-JEPA is built to unify heterogeneous wearable data across sensors, body locations, and populations, and JEPA’s latent prediction is designed exactly for this: it learns structure shared across sources rather than sensor-specific detail. The results confirm it. On the same wavelet tokens and participant splits, JEPA pretraining beats training from scratch on all six datasets:

Subject-disjoint macro-F1 (%), mean ± SD

DatasetNo pretrainingWavelet MAEW-JEPA
PAMAP280.76 ± 0.8784.73 ± 0.8087.19 ± 0.61
DSADS76.72 ± 1.9980.65 ± 0.9782.76 ± 1.76
MHealth78.98 ± 2.0083.70 ± 1.7788.89 ± 1.24
RealWorld71.90 ± 0.5878.53 ± 0.3880.53 ± 1.10
UCI-HAR81.74 ± 0.4583.73 ± 2.7383.97 ± 2.08
USC-HAD56.86 ± 0.5957.44 ± 0.5960.35 ± 1.40
Average74.4978.1380.62

JEPA also beats a matched Wavelet MAE on all six datasets (see table). Prior work uses wavelets differently: WaveToken quantizes coefficients for forecasting (Masserano et al., 2025), and Khan and Yener (2018) learn wavelet filter widths for classification. Our contribution is using wavelet tokens indexed by scale, sensor axis, and time bin as units of masked latent prediction for wearable signals. Motivation and evidence together confirm that JEPA pretraining is necessary.

2. Does each part of the tokenizer contribute? Yes. We follow your suggestion and separate frequency decomposition from pooling. With the same mean/RMS statistics, wavelet tokens beat temporal summaries by 5.81 points and score higher on all six datasets, so frequency decomposition matters most. Raw DWT coefficients already improve over raw temporal patches by 6.20 points, and mean/RMS pooling adds a further gain on five of six datasets. The full tokenizer remains strongest.

Subject-disjoint macro-F1 (%) for W-JEPA variants, mean ± SD

DatasetRaw temporal patchesTemporal mean/RMSRaw DWT coefficientsDWT + mean/RMS
PAMAP281.64 ± 1.2879.90 ± 1.0585.83 ± 1.0387.19 ± 0.61
DSADS69.98 ± 0.8872.51 ± 2.1282.71 ± 0.5682.76 ± 1.76
MHealth73.68 ± 1.8582.72 ± 2.7489.69 ± 0.7488.89 ± 1.24
RealWorld71.48 ± 1.3671.76 ± 1.6579.63 ± 2.2680.53 ± 1.10
UCI-HAR82.00 ± 1.0381.62 ± 1.0182.19 ± 0.4883.97 ± 2.08
USC-HAD59.52 ± 1.7360.33 ± 1.3155.47 ± 3.1860.35 ± 1.40

3. Participant partitions and seeds. For each dataset, every method is evaluated on the identical participant partition, so no method benefits from a favorable split; the five seeds measure training variability on each partition. The target dataset is excluded from pretraining, so the held-out participants are unseen in both stages. We will state the this protocol explicitly in §4.1 .

Reviewer_vKWYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes W-JEPA, a new self-supervised learning appraoch for wearable activity recognition. Essentially, it is a combination of DWT and JEPA, where a model is inputted summaries (the RMS and mean) of the coefficients of DWT at multiple scales, and is trained to predict the latent embeddings from a teacher EMA model. The paper argues that using DWT coefficients, the model is able to capture subtle patterns in motion signals in activities that are difficult to distinguish from one another across participants. Through evaluations across 6 datasets, W-JEPA outperforms 6 or so state-of-the-art baselines under different settings.

Critical strengths
  • The paper is very clearly written. It was one of the most easy to follow papers I have read recently.
  • The proposed method shows strong and consistent performance across 6 datasets.
Critical weaknesses
  • The novelty of the proposed method isn’t explored enough. The main components of the approach are DWT, summarization (RMS, mean), JEPA, and EMA. While the combination of existing methods isn’t to be criticized, I feel like a deeper analysis of the components and their combination is needed to clearly ground the contribution. For example, how does the number of bands affect the downstream performance? And what if the RMS/mean summary isn’t used and the raw DWT coefficients are used instead? Why these design choices were made, or what made them necessary, isn’t established clearly. While it’s good to know that these combinations produce great results, it feels like a technical report of the best performing design through a trial of many different variants.
  • The one major critical weakness I find for this paper is the lack of information on how the baselines are implemented. Firstly, how do the model sizes compare? And more importantly, do the baselines also use the same number of input channels for the downstream tasks? Section 4.1 says “downstream fine-tuning uses all configured target channels”, so I’m assuming for PAMAP2, W-JEPA uses all 27 channels. So that means the pretrained context encoder is kept for finetuning, the learned 6-channel embedding (from pretraining) is replaced by a 27-channel embedding to be trained (finetuned) from scratch, as well as a linear classifier. Then how about the baselines? How are they pretrained on 6 channels and then finetuned on a different number? If the baselines use only 6 channels for downstream tasks, I wouldn’t say the comaprison is fair.
Questions for authors
  • I was surprised by the results of the “relative attention change across noise doses” experiment. I would expect the noise to somewhat contaminate the higher frequency information and thus make the model attend more to the bands that contain lower frequencies (A4, D4). What explains this shift toward higher-frequency bands as more noise is added?
  • I find it interesting that the performance drops quite a bit, specifically for the ~120 bin size (Figure 5). It seems like having either very few bins, or very many bins, is better (and not much different). Do you have some hypothesis for why this could be?
  • I personally think Table 2 is wasting valuable space that could be used by other ablations or analyses. I think most people in this field would know that random-splitting (thus having data from the same participant in both the training and testing set) is not a feasible method at all, and that subject-disjoint splitting is the norm. So alhtough Table 2 shows that W-JEPA has a smaller difference between random-window and subject-discjoint, I don’t think the full results for all comparison baselines are necessary. This is just a minor suggestion.
3 comments

Thank you for the thoughtful review. We ran the component analyses you requested and documented the baseline setup, and every result confirms that each design choice in W-JEPA is justified and that the full model is strongest. All of these results will be added to the revised paper. Based on the strong results we present, we kindly ask you to  consider raising your score.

1. Is each component of W-JEPA necessary? Yes. Waveforms are the primary format of wearable data, yet they differ across sensors, placements, and populations. Wavelets remove this heterogeneity while preserving the structure needed for latent prediction, and since activity signals are highly repetitive, mean/RMS summaries keep critical scale-specific information while discarding redundancy. JEPA then learns structure shared across sources rather than sensor-specific detail. We test each component: mean/RMS beats raw DWT coefficients on five of six datasets, four levels (five bands) give the best macro-F1 among depths 2–6, and replacing latent prediction with masked reconstruction (Wavelet MAE) lowers macro-F1 on all six datasets

Subject-disjoint macro-F1 (%)

DatasetWavelet MAEW-JEPA, raw coefficientsW-JEPA, mean + RMS
PAMAP284.73 ± 0.8085.83 ± 1.0387.19 ± 0.61
DSADS80.65 ± 0.9782.71 ± 0.5682.76 ± 1.76
MHealth83.70 ± 1.7789.69 ± 0.7488.89 ± 1.24
RealWorld78.53 ± 0.3879.63 ± 2.2680.53 ± 1.10
UCI-HAR83.73 ± 2.7382.19 ± 0.4883.97 ± 2.08
USC-HAD57.44 ± 0.5955.47 ± 3.1860.35 ± 1.40
Average78.1379.2580.62
Wavelet typeMean subject-disjoint macro-F1 (%)
Daubechies 4 (db4)80.62
Haar80.14
Daubechies 279.05
Coiflet 378.60
Adaptive db478.52
Symlet 677.92
Symlet 477.79
Biorthogonal 4.477.25
db4 DepthMean subject-disjoint macro-F1 (%)
275.60
377.78
480.62
579.61
680.27

3. Are the baselines compared fairly in channels and model size? Yes. All methods are pretrained on the same six-channel views (one triaxial accelerometer and gyroscope per source) and fine-tuned on the same target channels: 27 (PAMAP2), 30 (DSADS), 15 (MHealth), 42 (RealWorld), and 6 (UCI-HAR, USC-HAD). Your reading of W-JEPA is correct: we transfer the pretrained encoder, initialize a new channel embedding and classifier for the target channels, and fine-tune end to end. Each baseline likewise replaces its input layer for the target channels. W-JEPA is also competitive at a smaller size:

Encoder blocksParamsMean accuracyMean macro-F1
23.7M81.92%80.93%
47.2M81.62%80.91%
8 (default)14.4M81.43%80.62%

The 2-block model is smaller than HAR-JEPA (6.3M) and CALANet (5.2M) yet far exceeds HAR-JEPA’s macro-F1 (71.91%). The 8-block model keeps a slight edge on the harder UCI-HAR and USC-HAD. W-JEPA’s advantage therefore does not come from model size.

4. Discussion: why does attention shift toward higher-frequency bands under added noise? First, W-JEPA remains robust: accuracy and macro-F1 change by less than 1% at the 30% noise dose. The attention shift is an interesting phenomenon rather than a failure: at 30% noise, attention share changes by −6.1% for A₄, +11.1% for D₂, and +9.4% for D₁. One heuristic is that strong additive noise resembles high-frequency vibration, so the model inspects the detail bands where it appears in order to discount it. Fine-scale cues are also informative on clean data; D₁ already receives the largest attention share for running and jumping. Predictions stay stable throughout, and we will add this discussion to the paper.

5. Discussion: the dip near ~120 bins in Fig. 5. We thank the reviewer for noticing this interesting pattern. Each point in Fig. 5 comes from a single run with one seed, so the dip likely reflects run-to-run randomness rather than a property of the bin size. Re-running this setting with a different seed removes some of the dip, and performance stays high across the full range of bin sizes. We will report multi-seed results and clarify the axis label in the caption.

6. The paper will be strengthened by repositioning the random split results table. We will move the full random-window comparison to the appendix and use the space for the component analyses above.

Thank you for the response and additional experiments. I appreciate the effort to address my concerns.

Regarding my first weakness, although the new ablations do show that each component contributes to the performance, my broader concern still remains. Since the proposed method is largely a combination of existing techniques (DWT, summarization, JEPA), I think it's especially important to explain what we can learn from it beyond showing that the combination works well. The new experiments show that the proposed design outperforms several alternatives, but I still find it difficult to identify a clear takeaway useful for future work. What can we learn from this approach that wasn't already established by prior work? I think the paper needs to explain this more clearly.

Regarding my second weakness, I now understand that the baselines use the same number of channels. However, the baseline implementations are still not clearly reported. Only two baselines' model sizes are given even though the paper compares against many more. And it's unclear whether the authors used official implementations, how the different architectures were adapted for pretraining, and whether the training budgets were comparable. Because the paper's major contribution is in its strong empirical results, these details are necessary to properly evaluate the fairness of the comparisons.

Overall, the additional experiments have strengthened the paper and I appreciate the authors' effort. However, I don't think my main concerns have been sufficiently addressed yet, so I will maintain my current rating.

Thank you for engaging with our response and for recognizing that the new experiments strengthen the paper. We address both remaining concerns below.

1. What does W-JEPA teach beyond showing that the combination works? Across all ablations, a single principle emerges for self-supervised learning on wearables: the prediction target should keep the energy at each scale and coarse timing, and discard waveform detail that varies across sensors and people. This holds at both levels of the design:

  • Token level. The choice of prediction unit matters most. Moving from raw temporal patches to wavelet coefficients gives +6.20 macro-F1, and with identical mean/RMS statistics, wavelet tokens beat temporal tokens by +5.81 on all six datasets. Summarizing coefficients into band energy and sign then beats predicting raw coefficients on five of six datasets. The useful information for activity recognition sits in band energy at coarse timing, not in exact coefficient values.
  • Objective level. The same principle explains why latent prediction beats reconstruction on identical tokens (+2.49 over Wavelet MAE on all six datasets). Reconstruction forces the model to reproduce detail that the token-level results show is not useful for the task.

This gives future work concrete guidance. When pooling heterogeneous wearable sources, the choice of what to predict matters more than model capacity: a 2-block, 3.7M-parameter encoder matches the 14.4M default, and the gains come from the prediction target rather than model size. These results also help explain why the gap between random-window and subject-disjoint evaluation is smaller for W-JEPA: scale-indexed energy features vary less across people than raw waveforms do. We will add a dedicated paragraph to the paper stating this principle and the evidence for it.

2. Are all baseline implementations reported in full? Yes. Every baseline follows the official codebase released with its original paper. All self-supervised methods are pretrained on the same six-channel views from the same five source datasets for 50 epochs. CALANet is purely supervised, so it has no pretraining stage. Every method is then fully fine-tuned end to end on the same target channels for 70 epochs (for CALANet, 70 epochs of supervised training), with shared preprocessing and participant splits. Training budgets are therefore identical across methods. Encoder sizes for all methods:

MethodParams
TSLANet0.55M
BYOL1.64M
ModCL1.90M
MoPFormer2.52M
CALANet5.21M
HAR-JEPA6.33M
W-JEPA (2 blocks)3.70M
W-JEPA (default, 8 blocks)14.38M

The 2-block W-JEPA is smaller than CALANet and HAR-JEPA and matches the default model's performance (80.93% vs. 80.62% macro-F1), so its advantage does not come from size. We will add these details, with per-baseline hyperparameters, to the appendix.

Reviewer_i5JcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper aims to tackle the problem of cross-dataset transfer (model pre-trained on activity datasets) to unseen participants (pre-trained models fine-tuned with target datasets evaluated under a held-out participant eval protocol). Using recent work in JEPA-type approaches, the paper proposes W-JEPA, a self-supervised framework for wearable human activity recognition that combines wavelet-based tokenization with a joint-embedding predictive architecture. The method decomposes inertial signals into multiple frequency bands, summarizes them using temporal tokens, and applies masked latent prediction to learn representations across sensor channels and timescales. The paper evaluates W-JEPA on six activity recognition datasets under random-window and subject-disjoint protocols, reporting improvements over several existing approaches.

Critical strengths
  • The paper has a clear motivation and is generally well written. The proposed architecture is clearly illustrated, and the methodology is understandable with macro-F1 eval, treating each class equally.
  • Good empirical comparison: The authors evaluate six datasets against several existing methods, reporting accuracy and macro-F1 across five random seeds.
  • Promising performance: W-JEPA achieves improvements over the evaluated baselines, particularly on PAMAP2, DSADS, and RealWorld. The comparison against W-JEPA without wavelets also provides evidence that the proposed tokenization benefits the architecture.
Critical weaknesses

Although the results are promising and the paper is generally well organized, I have concerns regarding the experimental validation and justification of the proposed methodology. In particular, several design choices are not sufficiently investigated, and the experiments do not adequately isolate the contribution of JEPA from the proposed wavelet representation and masking strategy.

  • The motivation for exploiting multiscale information in wearable activity recognition is not new (see for example: "Ambulatory system for human motion analysis using a kinematic sensor: monitoring of daily physical activity in the elderly", 2003.)

  • Introduction and methodological motivation. The necessity of JEPA is not sufficiently established. The introduction motivates the proposed approach through the limitations of raw temporal patches, arguing that conventional temporal representations mix information from different timescales and may obscure subtle activity-specific patterns. However, this motivation does not inherently require a joint-embedding predictive architecture. A conventional masked Transformer operating on wavelet-based tokens could potentially capture similar cross-scale and cross-axis dependencies. Although the authors compare against methods such as MoPFormer and HAR-JEPA, these comparisons involve different tokenization strategies, architectures, and learning objectives. Consequently, they do not isolate the advantage of JEPA when using the proposed wavelet representation. The comparison between W-JEPA and W-JEPA without wavelets establishes the usefulness of the tokenizer within the proposed framework but does not determine whether latent prediction contributes beyond wavelet tokenization. A controlled comparison against wavelet-based masked reconstruction, using the same architecture and tokenization, would significantly strengthen the paper. A supervised Transformer operating on the same wavelet representations would also help establish whether self-supervised pretraining provides an additional benefit.

  • Section 3.1: Insufficient justification of the wavelet representation.

    • The authors employ a four-level Daubechies-4 decomposition, with two temporal bins per frequency band, summarized using mean and RMS statistics. While this is a reasonable representation, several important design choices are insufficiently justified. Although the authors reference the wavelet decomposition appropriately, the motivation for selecting Daubechies-4 with four decomposition levels remains unclear. Given that wavelet tokenization is the central methodological contribution, an analysis of different decomposition depths or wavelet families would help establish whether the reported improvements are robust to these design choices.
    • Second, representing each frequency band using only two temporal bins and simple summary statistics substantially compresses the original temporal information. Given that the paper motivates its approach through fine-grained temporal differences between activities, it is unclear how much of this information is preserved by the proposed tokenization. This also raises the possibility that the observed improvements originate primarily from discriminative wavelet statistics rather than the proposed predictive-learning mechanism.
  • Section 4.2: The participant-diversity experiment does not isolate participant diversity.

    • Figure 3 evaluates performance as additional participants are included in the fine-tuning set. However, increasing the number of participants simultaneously increases the amount of training data and potentially the diversity of temporal patterns. The experiment does not control for training-set size or temporal variability, making it difficult to attribute the observed behavior specifically to participant diversity.
    • Moreover, the figure largely reinforces the existing observation that W-JEPA outperforms its no-wavelets counterpart, without providing much additional insight into the underlying mechanism. A more informative experiment would control the total number of training windows while varying the number of participants and, if possible, characterize the associated temporal or spectral variability.
  • Section 4.3: Restricted sensor evaluation is not a methodological ablation. Section 4.3 is presented as an ablation, but primarily evaluates performance under a restricted sensor configuration. While this is a reasonable practical evaluation, it does not isolate the contribution of any specific component of W-JEPA. In particular, it remains unclear how the asynchronous paired-token masking, cross-scale information, and latent prediction objective individually contribute to performance.

  • Hyperparameter selection and experimental sensitivity.

    • Some hyperparameters are introduced without sufficient justification. For example, Equation (3) uses a variance regularization coefficient of 0.0625, but the paper does not explain how this value was selected. Although Appendix C.1 compares alternative regularizers, it does not establish whether the regularization term is necessary or how sensitive the method is to its weighting.
    • Likewise, the selected wavelet family and decomposition depth are not investigated.
    • The authors do provide sensitivity analyses for masking ratio, temporal bins, and window length, which is appreciated. However, these experiments are limited to one dataset and one run per configuration.
  • Experimental protocol and presentation

    • There are six channels, but each dataset consists of multiple sensors, in multiple locations and different numbers of sensors. How this is dealt with is not in the paper form what I can gather.
    • Unclear how sensor orientation issues were addressed or where they ignored? Not all datasets have the same number of sensors, types not locations.
    • Do not consider, models such as Tri-HAR, or other models that aim to build models invariant to sensor placement/orientation as relevant baselines
Questions for authors
  1. What is the specific benefit of JEPA compared with conventional masked reconstruction using exactly the same wavelet tokens and masking strategy? Is there evidence that latent prediction provides additional benefits beyond the wavelet representation?
  2. Have the authors evaluated a supervised Transformer using the proposed wavelet tokenization without JEPA pretraining? This would help isolate the contribution of the representation itself.
  3. What motivates selecting Daubechies-4 with four decomposition levels? How sensitive is performance to these choices?
  4. Given that each wavelet band is compressed into two temporal bins using mean and RMS statistics, how do the authors establish that the proposed tokens preserve the fine-grained temporal differences motivating the method?
  5. In Figure 3, how can the improvements be attributed to participant diversity rather than increased training data or temporal variability?
  6. How was the variance regularization coefficient of 0.0625 selected? What happens if this regularizer is removed entirely?
  7. Are the five reported seeds associated with different participant partitions, or are participant selections fixed across runs?
1 comment

Thank you for the careful detailed review. We ran every controlled experiment you requested, and each confirms that W-JEPA’s gains come from both the wavelet tokenizer and latent prediction. All of these results will be added to the revised paper. Based on the strong results we present, we kindly ask you to consider raising your score.

1. Do JEPA and the wavelet summaries each add value? Yes. W-JEPA is built to unify heterogeneous wearable data across sensors, body locations, and populations. Latent prediction is designed exactly for this: it learns structure shared across sources rather than sensor-specific detail. With the tokenizer, encoder, asynchronous paired masking, and training budget all matched, W-JEPA beats Wavelet MAE on all six datasets and beats a supervised wavelet Transformer trained from scratch on all six. Mean/RMS keep each band’s signed average and energy, the cues that separate similar activities, and beat raw DWT coefficients on five of six datasets. Both JEPA and the wavelet summaries are necessary.

Subject-disjoint macro-F1 (%)

DatasetWavelet Transformer (from scratch)Wavelet MAEW-JEPA, raw coefficientsW-JEPA, mean + RMS
PAMAP280.76 ± 0.8784.73 ± 0.8085.83 ± 1.0387.19 ± 0.61
DSADS76.72 ± 1.9980.65 ± 0.9782.71 ± 0.5682.76 ± 1.76
MHealth78.98 ± 2.0083.70 ± 1.7789.69 ± 0.7488.89 ± 1.24
RealWorld71.90 ± 0.5878.53 ± 0.3879.63 ± 2.2680.53 ± 1.10
UCI-HAR81.74 ± 0.4583.73 ± 2.7382.19 ± 0.4883.97 ± 2.08
USC-HAD56.86 ± 0.5957.44 ± 0.5955.47 ± 3.1860.35 ± 1.40
2. Are db4 and 4 levels the best design choices? Yes. Four levels give four detail bands plus a coarse approximation, covering both fast and slow motion. Db4 is best among 8 families, and depth 4 is best among depths 2–6. We will cite Najafi et al. (2003); our contribution is using scale, axis, and time-indexed wavelet tokens as units of masked latent prediction.
Wavelet typeMean subject-disjoint macro-F1 (%)
Daubechies 4 (db4)80.62
Haar80.14
Daubechies 279.05
Coiflet 378.60
Adaptive (adaptive db4)78.52
Symlet 677.92
Symlet 477.79
Biorthogonal 4.477.25
db4 DepthMean subject-disjoint macro-F1 (%)
275.60
377.78
480.62
579.61
680.27

3. Participant diversity and seeds. For each dataset, every method is evaluated on the identical participant partition, so no method benefits from a favorable split; the five seeds measure training variability on each partition. For Fig. 3, we will revise the text to state that adding participants also adds windows. The comparisons are fair, and W-JEPA leads consistently.

4. Is the multi-sensor setup handled consistently across datasets? Yes. For each held-out target, we pretrain on the other five datasets using six channels from one body location per source, ordered as accelerometer then gyroscope: hand (PAMAP2), right arm (DSADS, MHealth), waist (RealWorld, UCI-HAR), and front right hip (USC-HAD). This standardizes sensor type and axis order while letting placement vary, which is exactly the heterogeneity W-JEPA is designed to unify. Fine-tuning uses all configured channels of the target. We use recorded sensor axes without orientation alignment; orientation-invariant methods such as Tri-HAR are complementary, and we will discuss them. We will document this setup in §4.1 and relabel §4.3 as a restricted-sensor evaluation.

5. Is 0.0625 the best regularizer weight, and is the regularizer needed? Yes. Across weights 0, 0.015625, 0.03125, 0.0625, 0.125, and 0.25, mean macro-F1 is 79.19, 79.24, 78.78, 80.62, 79.47, and 79.17%; 0.0625 also gives the highest accuracy (81.43%). The chosen weight is best, and the regularizer is necessary.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.