Assessing the State of Self-Supervised Human Activity Recognition Using Wearables

May combine two claims.

The test set mean F1-score (macro F1-score) is used as the main metric because the target datasets have substantial class imbalance and a metric resistant to class distribution bias is required.

“We utilize the test set mean F1-score (i.e., macro F1-score) as the main metric to evaluate performance.This is motivated by the substantial class imbalance present in the target datasets (see Fig. 15 for reference), and thus we require a metric that is resistant to such bias in the class distribution [56].”

• cites [56] 56. David MW Powers. 2020. Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation. arXiv preprint arXiv:2010.16061 (2020). sourceDisputed

  The source makes the contrary point about F1: the Introduction (p. 1) says F-measure propagates underlying prevalence and bias, and “ROC and PN Analyses” (p. 5) shows F1 varies with prevalence and bias. Its multiclass simulation caption (p. 9) also describes harmonic-mean F1 as a biased traditional measure. The source does not assess the citing paper’s target datasets. (LLM, conf=0.86)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

May combine two claims.

SelfHAR is not included because it employs self-training with self-supervision, resulting in a separate pre-trained model per target dataset, which limits generalizability and does not allow the self-supervised transfer learning setup used in the study.

“Therefore, each target dataset results in its own pre-trained model, which limits generalizability and thus does not allow for the self-supervised transfer learning setup utilized in our study.”

• cites [71] 71. Chi Ian Tang, Ignacio Perez-Pozuelo, Dimitris Spathis, Soren Brage, Nick Wareham, and Cecilia Mascolo. 2021. SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data. arXiv preprint arXiv:2102.06073 (2021). sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Capture-24 is a wrist-worn dataset with 151 users and free-living activities.

“Capture-24 [4, 18, 81] Wrist 151 -Free living”

• cites [4] 4. S Chan Chang and A Doherty. 2021. Capture-24: Activity tracker dataset for human activity recognition. (2021). sourceNot checked — no full text of the source
• cites [18] 18. Jonathan Gershuny, Teresa Harms, Aiden Doherty, Emma Thomas, Karen Milton, Paul Kelly, and Charlie Foster. 2020. Testing self-report time-use diaries against objective instruments in real time. Sociological Methodology 50, 1 (2020), 318-349. sourcePartly supported

  The Axivity AX3 band accelerometer section (p. 8) says participants wore the device on their wrist during ordinary activities, supporting the wrist-worn, free-living description. But Sample and setting (p. 6) gives an original sample of 148 participants, not 151, with 124 complete diary, camera and accelerometer records. (LLM, conf=0.90)

  Experimental — accuracy not yet measured
• cites [81] 81. Matthew Willetts, Sven Hollowell, Louis Aslett, Chris Holmes, and Aiden Doherty. 2018. Statistical machine learning of sleep and physical activity phenotypes from sensor data in 96,220 UK Biobank participants. Scientific reports 8, 1 (2018), 1-10. sourcePartly supported

  The Methods (Participants and Accelerometer; page unavailable) describe CAPTURE-24 participants wearing wrist-worn accelerometers, and the paper describes their activity data as free-living. The participant count does not match: 143 were recruited, and 132 remained for classifier development, not 151. (LLM, conf=0.91)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Self-supervised learning was first explored for human activity recognition in a multi-task setting.

“Self-supervised learning was first explored for human activity recognition in a multi-task setting [62].”

• cites [62] 62. Aaqib Saeed, Tanir Ozcelebi, and Johan Lukkien. 2019. Multi-task self-supervised learning for human activity detection. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 2 (2019), 1-30. sourcePartly supported

  The paper describes using self-supervised learning for human activity recognition with a multi-task network (p. 1; Conclusions, p. 23). It calls the approach novel, but the cited passages do not establish that it was the first exploration of self-supervised learning for human activity recognition. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

The fourth layer of CPC is the Gated Recurrent Unit, which also learns representations similar to supervised learning.

“The fourth layer of CPC is the Gated Recurrent Unit [8], which also learns representations similar to supervised learning.”

• cites [8] 8. Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014). sourceSupported

• AssumptionRests on an assumption: The gold-standard for activity recognition generally lies in supervised learning, resulting in the best possible performance.

May combine two claims.

CPC has applicability and superior performance on multi-sensor data and in scenarios where transition-style activities are present.

“A strength of this method lies in its applicability and superior performance on multi-sensor data [27], and in scenarios where transition-style activities are present (for example, getting in and out of a car, as in Mobiact).”

• cites [27] 27. Harish Haresamudram, Irfan Essa, and Thomas Plötz. 2021. Contrastive predictive coding for human activity recognition. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 5, 2 (2021), 1-26. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

The choice of augmentations has a large impact on pre-training quality.

“As mentioned in the original paper [6], the choice of augmentations has a large impact on the pre-training quality.”

• cites [6] 6. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning. PMLR, 1597-1607. sourceSupported

Probably not a substantial claim: may not pass S2 (checkable).

Evaluation on a suite of tasks is necessary for a well-rounded understanding of model performance.

“In this work, we follow the philosophy espoused by natural language processing benchmarks such as GLUE [79] and GEM [17], which argues that evaluation on a suite of tasks is necessary for a well-rounded understanding of model performance.”

• cites [79] 79. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461 (2018). sourceSupported
• cites [17] 17. Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal, Pawan Sasanka Ammanamanchi, Aremu Anuoluwapo, Antoine Bosselut, Khyathi Raghavi Chandu, Miruna Clinciu, Dipanjan Das, Kaustubh D Dhole, et al. 2021. The gem benchmark: Natural language generation, its evaluation and metrics. arXiv preprint arXiv:2102.01672 (2021). sourceSupported

May combine two claims.

Best performing self-supervised methods such as SimCLR and Multi-task self-supervision show significant improvements over classifying on raw data for a majority of the datasets, e.g., on PAMAP2, SimCLR shows an increase of approximately 10% F1-score over utilizing raw data.

“Considering the best performing self-supervised methods such as SimCLR and Multi-task self-supervision, we see significant improvements over classifying on raw data for a majority of the datasets (e.g., on PAMAP2, SimCLR shows an increase of approx.10% F1-score over utilizing raw data).”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

Most datasets have class imbalance, with Wetlab, Mobiact, MHEALTH and Myogym having the most severe class compositions.

“Lastly, most of the datasets have class imbalance (see Fig. 15 for reference) with Wetlab, Mobiact, MHEALTH and Myogym having the most severe class compositions.”

May combine two claims.

Pre-training is less effective for Mobiact, with DeepConvLSTM significantly outperforming the best self-supervised methods, increasing by over 10% against CPC and 7% over SimCLR.

“Pre-training is less effective for Mobiact, with DeepConvLSTM significantly outperforming the best self-supervised methods, increasing by over 10% against CPC and 7% over SimCLR.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

Transferring weights from the wrist to the waist results in worsened performance for Mobiact and Motionsense when compared to the best supervised baseline, i.e., the Conv.classifier, or similar performance for USC-HAD.

“The overall trend is that transferring weights from the wrist to the waist results in worsened performance (Mobiact and Motionsense, when compared to the best supervised baseline, i.e., the Conv.classifier) or similar performance (USC-HAD).”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.
• AssumptionRests on an assumption: Sensor data from different datasets are comparable after per-channel normalization, so representations transfer across devices and body locations.

May combine two claims.

The Capture-24 dataset composition is imbalanced, with around 75% of windows corresponding to either sleep or sit-stand, whereas bicycling comprises only 0.78% of the train dataset windows.

“The dataset composition is imbalanced, with around 75% corresponding to either sleep or sit-stand, whereas bicycling comprises of only 0.78% of the train dataset windows, as shown in Fig. 13 in the Appendix.”

May combine two claims.

The LSTM classifier and the MLP classifier show similar performance, with the GRU classifier performing better than both.

“By and large, the LSTM classifier and the MLP classifier show similar performance, with the GRU classifier performing better than both.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

The highest improvements over DeepConvLSTM, LSTM, GRU, and MLP classifiers are observed with PAMAP2, whereas the Conv.classifier performs comparably.

“As mentioned previously, we observe the highest improvements over DeepConvLSTM, LSTM, GRU, and MLP classifiers with PAMAP2, whereas the Conv.classifier performs comparably.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

For linear evaluation, performance is substantially lower relative to all supervised learning baselines across both waist- and leg-based datasets, with only SimCLR obtaining performance comparable to supervised approaches, and Mobiact is an exception.

“Across both the waist-and leg-based datasets, we note that the performance is substantially lower relative to all supervised learning baselines.Only in the case of SimCLR do we obtain performance comparable to the supervised approaches (Mobiact is an exception).”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

Transferring from the wrist to leg results in comparable performance for Daphnet FoG and PAMAP2 against the Conv.classifier, or gains for MHEALTH.

“In comparison, transferring from the wrist to leg results in comparable performance (Daphnet FoG and PAMAP2, against the Conv.classifier), or gains (MHEALTH).”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.
• AssumptionRests on an assumption: Sensor data from different datasets are comparable after per-channel normalization, so representations transfer across devices and body locations.

May combine two claims.

For Mobiact, the difference between the means of the balanced and imbalanced runs is ∼0.9% when ρ=0.1, increasing to ∼2% and ∼3.3% when ρ is 0.01 and 0.001 respectively.

“For Mobiact, the difference between the means of the balanced and imbalanced runs is ∼0.9% when ρ=0.1 (i.e., the rarest class is 10 times smaller than the most frequenct class).This increases to ∼2% and ∼3.3% when ρ is 0.01 and 0.001 respectively.”

• AssumptionRests on an assumption: Wearables-based movement datasets tend to be imbalanced, with more samples of naturally occurring or easily performed activities than strenuous or rare activities.

May combine two claims.

Self-supervision results in significant improvements over DeepConvLSTM for PAMAP2, approximately 9% over the mean, and MHEALTH, approximately 4% over the mean.

“In each case, self-supervision results in significant improvements over DeepConvLSTM for PAMAP2 (∼9% over the mean) and MHEALTH (∼4% over the mean).”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

In HHAR, Multi-task self-supervision increases from 51% to 57.5% with the use of the MLP; CPC and SimCLR show improvements from around 56% to 58%.

“In the case of HHAR, which likely has the most overlap with Capture-24, Multi-task self-supervision increases from 51% to 57.5% with the use of the MLP.Similarly, CPC and SimCLR show improvements from around 56% to 58%.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

Multi-task self-supervision and CPC benefit from a larger number of participants, but there is generally no significant gain between using 25% of the subjects over utilizing the entire training set.

“Multi-task self-supervision and CPC benefit from larger number of participants but there is generally no significant gain between using 25% of the subjects over utilizing the entire training set.”

May combine two claims.

The other datasets were recorded at 100 Hz and the reduction in performance occurs to a smaller extent for some of the self-supervised methods.

“The other datasets were recorded at 100 Hz and the reduction in performance occurs to a smaller extent for some of the self-supervised methods.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.
• AssumptionRests on an assumption: The 50 Hz sampling rate retains sufficient information for the target activities.

May combine two claims.

For SimSiam and BYOL, providing more pre-training windows slightly improves the F1-score for most target datasets, without showing drastic changes in performance.

“For SimSiam and BYOL, providing more pre-training windows slightly improves the F1-score for most target datasets, without showing drastic changes in performance.”

May combine two claims.

Using only 10% of the training windows yields excellent performance, comparable if not better than using more windows.

“We note that excellent performance can be obtained while using 10% of the training windows, and in some cases using as few as 1% of the training windows.”

• Also stated in: § Effect of the number of training windows on activity recognition. · plain

  “As a general trend, we observe that the self-supervised methods obtain comparable if not better performance when utilizing only 10% (or ∼400k) of training windows.”

• Also stated in: § 6.2 Is the Performance of Self-Supervision Getting Hampered by (Too) Simple Encoder Architectures? · plain

  “In our study of the quantity of the unlabeled source dataset windows required for pre-training (see Fig. 4), we observed that using only 10% of the training windows results in comparable if not better downstream activity recognition.”

May combine two claims.

For other datasets, the best performing self-supervised methods still typically outperform the classifier, although improvements may be more modest (e.g., Daphnet FoG and PAMAP2).

“For other datasets, the best performing self-supervised methods still typically outperform the classifier, although the improvements maybe more modest (e.g., in the case of Daphnet FoG dataset and PAMAP2).”

• AssumptionRests on an assumption: Very limited labeled data scenarios are of vital importance to activity recognition and can arise due to cost and privacy concerns, such as in medical applications.
• AssumptionRests on an assumption: Real-world deployment may allow for acquiring a small labeled dataset from users without significant interruptions, such as during device setup.

May combine two claims.

Masked Reconstruction leads to a jump in performance between using 1% and 10% of the windows, while the F1-score steadies beyond that.

“Masked Reconstruction leads to a jump in performance between using 1% and 10% of the windows while the F1-score steadies beyond that.”

May combine two claims.

Using data from just one user does not result in poor performance; rather, the performance is close to utilizing all available training data.

“Surprisingly, using data from just one user does not result in poor performance; rather, the performance is close to utilizing all available training data.”

May combine two claims. The quote only partly states this claim.

The contrastive learning methods show the highest difference in train F1-score between ground truth and randomized labels, thereby learning the least generic features, especially SimCLR, SimSiam, and BYOL.

“The contrastive learning methods show the highest difference in the train F1-score thereby learning the least generic features, especially SimCLR, SimSiam, and BYOL.”

May combine two claims.

In Myogym, Autoencoder increases from 11.4% to 35.5%, and BYOL improves from 17% to 39% with MLP-based classification.

“The impact of the MLP-based classification is more pronounced on the fine-grained gym activities (e.g., curls, raises, and dips) present in Myogym, with Autoencoder increasing from 11.4% to 35.5%, and BYOL improving from 17% to 39%.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

For contrastive methods such as CPC, SimCLR, and BYOL, as well as the Autoencoder, around 95% of the variance can be explained from the first five components, increasing to over 98% when considering 20 components.

“In Fig. 10, we see that for the contrastive methods such as CPC, SimCLR, and BYOL, as well as the Autoencoder, around 95% of the variance can be explained from the first five components itself, which increases to over 98% when considering 20 components.”

May combine two claims.

The Autoencoder follows the general trend of higher similarity in earlier layers, with the first two convolutional layers being very similar.

“the Autoencoder also follows the general trend of higher similarity in the earlier layers with the first two convolutional layers being very similar.”

• AssumptionRests on an assumption: The gold-standard for activity recognition generally lies in supervised learning, resulting in the best possible performance.

May combine two claims.

Multi-task self-supervision generally has a lower difference in performance between ground truth and randomized labels, indicating that it learns more general representations that enable various downstream scenarios.

“We observe that Multi-task self-supervision generally has a lower difference in the performance between ground truth and randomized labels, indicating that it learns more general representations that enable various downstream scenarios.”

May combine two claims.

In general, SimCLR is the most effective self-supervised method for activity recognition using wearables, across both sensor positions and target activities, whereas Multi-task self-supervision also shows high performance.

“In general, SimCLR is the most effective self-supervised method for activity recognition using wearables, across both sensor positions and target activities, whereas Multi-task self-supervision also shows high performance.”

• AssumptionRests on an assumption: Very limited labeled data scenarios are of vital importance to activity recognition and can arise due to cost and privacy concerns, such as in medical applications.
• AssumptionRests on an assumption: Real-world deployment may allow for acquiring a small labeled dataset from users without significant interruptions, such as during device setup.

May combine two claims.

For the lab activities in Wetlab, there is a drop in performance, with a peak difference of approximately 10% between DeepConvLSTM and SimCLR.

“For the lab activities in Wetlab, we observe a drop in performance, with a peak difference of ∼10% between DeepConvLSTM and SimCLR.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.

May combine two claims.

For Myogym, the reduction in performance is ∼0.5%, 1.6%, and 2.4% when ρ is 0.1, 0.01, and 0.001 respectively.

“Myogym follows a similar trend, with reduction ∈ ∼{0.5%, 1.6%, 2.4%} when ρ ∈ {0.1, 0.01, 0.001}.”

• AssumptionRests on an assumption: Wearables-based movement datasets tend to be imbalanced, with more samples of naturally occurring or easily performed activities than strenuous or rare activities.

Probably not a substantial claim: may not pass S3 (load-bearing).

SimSiam's design is simple.

“The main advantages of this method include its simplicity of design, where neither negative pairs nor a momentum encoder or large batch sizes are required for effective learning.”

May combine two claims.

BYOL trains quickly and is not computationally heavy.

“This method also trains quickly and is not computationally heavy, while also keeping the pre-training hyper-parameter space relatively small.”

Classified as probably not a substantial claim.

Self-supervised approaches have strong practical value, exhibiting significant improvements in scenarios where labeled data is at a premium.

“These approaches have strong practical value, exhibiting significant improvements in scenarios where labeled data is at a premium.”

May combine two claims.

Effective learning in SimSiam does not require negative pairs, a momentum encoder, or large batch sizes.

“The main advantages of this method include its simplicity of design, where neither negative pairs nor a momentum encoder or large batch sizes are required for effective learning.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A small yet diverse dataset in terms of participants can result in better recognition than collecting lots of data from a few participants.

“By and large, a small yet diverse dataset (in terms of participants) can result in better recognition than collecting lots of data from a few participants.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The target dataset sampling rate must match the source conditions for optimal performance.

“The target dataset sampling rate must match the source conditions for optimal performance.”

May combine two claims.

Early layers of learned encoders show high similarity to supervised learning, whereas the features diverge deeper into the network.

“Early layers of the learned encoders show high similarity to supervised learning, whereas the features diverge deeper into the network.”

Probably not a substantial claim: may not pass S3 (load-bearing).

CPC requires considerable parameter tuning, including learning rate and weight decay, future prediction horizon, and batch size.

“A downside to this approach is that it requires considerable parameter tuning, including learning rate and weight decay, future prediction horizon, and batch size.”

May combine two claims.

SimSiam has low training times and a small hyperparameter space.

“The model architecture is simple with low training times and a small hyperparameter space.”

May combine two claims.

For Multi-task self-supervision and Autoencoder, activity recognition performance reduces with the addition of another encoder layer.

“For both Multi-task self-supervision and Autoencoder, the performance reduces with the addition of another encoder layer, indicating that such naive increase in encoder depth does not show a corresponding improvement in activity recognition, even if more pre-training data are available.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

CPC training time is considerable and extends particularly with increasing prediction horizons.

“The training time is considerable, and extends particularly with increasing prediction horizons.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Masked reconstruction has the highest implicit dimensionality and thus more efficiently makes use of the learned representation space.

“Masked reconstruction has the highest implicit dimensionality and thus more efficiently makes use of the learned representation space.”

Probably not a substantial claim: may not pass S3 (load-bearing).

BYOL keeps the pre-training hyperparameter space relatively small.

“This method also trains quickly and is not computationally heavy, while also keeping the pre-training hyper-parameter space relatively small.”

Probably not a substantial claim: may not pass S3 (load-bearing).

When training windows are sub-sampled based on a percentage, the class distribution generally follows the composition of the full train dataset, albeit the number of samples in each class are different.

“When the training windows are sub-sampled based on a percentage, the class distribution generally follows the composition of the full train dataset (see Fig. 13) albeit the number of samples in each class are different.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Utilizing the means and variances of the source dataset normalization on the target dataset results in considerable performance gains.

“Utilizing the means and variances of the source dataset normalization on the target dataset results in considerable performance gains.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SimCLR training is relatively fast.

“The training itself is relatively fast, while the hyper-parameter space is not very large, comprising of learning rate, weight decay, and the batch size.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Sleep and sit-stand comprise around 75% of the total Capture-24 train dataset.

“Sleep and sit-stand comprise around 75% of the total dataset”

May combine two claims.

Multi-task self-supervision learns representations applicable to a variety of downstream scenarios, as evidenced by its overall superior performance in Tab. 3 and 4.

“The main advantage of utilizing this approach lies in its effectiveness in learning representations that are applicable for a variety of downstream scenarios, as evidenced by its overall superior performance in Tab. 3 and 4.”

Probably not a substantial claim: may not pass S4 (based).

Self-supervised learning is robust to even extreme imbalances in the source dataset, with minimal reduction in performance relative to balanced data.

“Self-supervised learning is surprisingly robust to even extreme imbalances in the source dataset, with minimal reduction in performance relative to balanced data.”

May combine two claims.

SimCLR is computationally light but necessitates large batch sizes because the negative pairs for contrastive learning consist of all other augmented windows in the batch.

“While it is computationally light as well, the setup necessitates large batch sizes, as the negative pairs for contrastive learning consist of all other augmented windows in the batch.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Bicycling consists of only 0.78% of the Capture-24 train data.

“whereas bicycling only consists of 0.78% of the train data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Layers closer to each other are more similar than those farther apart.

“Also, layers closer to each other are more similar than those farther apart.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SimCLR's hyperparameter space is not very large, comprising learning rate, weight decay, and batch size.

“The training itself is relatively fast, while the hyper-parameter space is not very large, comprising of learning rate, weight decay, and the batch size.”

May combine two claims.

Self-supervised techniques are data efficient and perform comparably while utilizing only a small fraction of available data, both from participant and quantity standpoints, reducing computational burden.

“The techniques are data efficient and perform comparably while only utilizing a small fraction of the available data -both from a participant as well as quantity of data standpoints -thereby reducing the computational burden.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

BYOL has a simple setup and architecture.

“The positives for this technique include its robustness to the choice of augmentations (in direct contrast to SimCLR), simple setup and architecture, and its general effectiveness across a variety of downstream scenarios.”

May combine two claims.

Excellent performance can be obtained while using 10% of the training windows, and in some cases using as few as 1% of the training windows.

“We note that excellent performance can be obtained while using 10% of the training windows, and in some cases using as few as 1% of the training windows.”

May combine two claims.

Autoencoder training time is significantly shorter relative to other methods and is less computationally stressful.

“Given the simplicity of the architecture and the objective, the training time is also significantly shorter relative to other methods, while also being less stressful from a computational perspective.”

May combine two claims.

Multi-task self-supervision succeeds in learning the most generic features of wearable data, and thereby enables a wide variety of downstream tasks.

“Multi-task self-supervision succeeds in learning the most generic features of wearable data, and thereby enables wide variety of downstream tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SimSiam produces surprisingly effective representations with a simple siamese setup.

“SimSiam [7] produces surprisingly effective representations with a simple siamese setup.”

• cites [7] 7. Xinlei Chen and Kaiming He. 2021. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15750-15758. sourceNot checked — no full text of the source

May combine two claims.

Masked reconstruction's Transformer encoder architecture is large, computationally intensive, and has the longest training time by far in this study.

“Additionally, the Transformer [77] encoder architecture is large and computationally intensive, and has the longest training time by far in this study.”

• cites [77] 77. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017).Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Five random seeds and five randomized classifier runs account for the effect of the choice of random seed.

“The best pre-training combination for each dataset is once again trained with five random seeds, and these trained models are utilized for five randomized classifier runs (using the same seeds as pre-training), in order to account for the effect of the choice of random”

Probably not a substantial claim: may not pass S3 (load-bearing).

Myogym contains 31 fine-grained gym exercises.

“whereas Myogym contains 31 fine-grained gym exercises.”

• cites [40] 40. Heli Koskimäki, Pekka Siirtola, and Juha Röning. 2017. Myogym: introducing an open gym data set for activity recognition collected using myo armband. In Proceedings of the 2017 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2017 ACM International Symposium on Wearable Computers. 537-546. sourceNot checked — no full text of the source

May combine two claims.

No filtering or denoising is performed on the datasets because deep networks have shown powerful capabilities towards learning from raw data.

“We perform no filtering or denoising on the datasets, as deep networks have shown powerful capabilities towards learning from raw data itself [42].”

• cites [42] 42. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. nature 521, 7553 (2015), 436-444. sourceNot checked — no full text of the source

May combine two claims.

The typical prior evaluation protocol is limited because it only studies one aspect of model performance, making it challenging to realize the conditions of effective performance and avenues of improvement for these approaches.

“Therefore, this evaluation protocol is limited as it only studies one aspect of model performance, which makes it challenging to realize the conditions of effective performance as well as avenues of improvement for these approaches.”

• Also stated in: § 2.1 Self-Supervised Learning in Human Activity Recognition · plain

  “Such an evaluation protocol only covers one aspect of measuring the model performance -which is under very similar training and testing conditions.”

Classified as probably not a substantial claim. May combine two claims.

Three datasets per location (waist/trousers, wrist, and leg/ankle) were chosen, covering a wide variety of activities and dataset sizes.

“We chose three datasets per location-waist/trousers, wrist, and leg/ankle-covering a wide variety of activities and dataset sizes.”

May combine two claims.

Every participant appears exactly once as part of the test split and pairwise correlations of analysis windows are effectively eliminated through the user-based splits.

“Overall, every participant appears exactly once as part of the test split and pairwise correlations of analysis windows are effectively eliminated through the user-based splits [22].”

• cites [22] 22. Nils Hammerla and Thomas Ploetz. 2015. Let's (not) Stick Together: Pairwise Similarity Biases Cross-Validation in Activity Recognition. In Proc. UbiComp. sourceNot checked — no full text of the source

May combine two claims.

Self-supervised approaches have the potential to alleviate challenges that come with limited annotations, including overfitting and the availability of only less-complex modeling approaches.

“These self-supervised approaches have the potential to alleviate challenges that come with limited annotations (incl.overfitting, and availability of less-complex modeling approaches only), and are therefore early proponents of the 'pretrain-then-finetune' paradigm shift.”

Classified as probably not a substantial claim. May combine two claims.

Five-fold cross validation uses user-based splits where each fold consists of 20% randomly chosen users as test, and the remaining 80% of users are split into train and validation at an 80:20 ratio.

“We setup five-fold cross validation for each target dataset.Each fold consists of 20% of randomly chosen users comprising the test split, whereas the remaining 80% of users are once again separated into the train and validation splits at a 80:20 ratio.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Five-fold cross validation yields an average F1-score that is more robust to the particular choice of participants in the splits.

“For each combination, we perform 5-fold cross validation to obtain the average F1-score, which is more robust to the particular choice of participants in the splits”

May combine two claims.

Accuracy and weighted F1-score are not utilized because they are affected by skewed class distributions, whereas unweighted mean F1-score is a reasonable strategy.

“The accuracy and weighted F1-score are not utilized as they are affected by the skewed class distributions, whereas the unweighted mean F1-score, while not ideal, is a reasonable strategy [53].”

• cites [53] 53. Thomas PlÖtz. 2021. Applying machine learning for sensor data analysis in interactive systems: Common pitfalls of pragmatic use and ways to avoid them. ACM Computing Surveys (CSUR) 54, 6 (2021), 1-25. sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

The nine benchmark datasets were chosen to cover a wide range of application scenarios, including ranges of sensor positions, activities covered, and sampling rates employed for data recording.

“They have been chosen carefully to cover a wide range of application scenarios, including ranges of sensor positions, activities covered, and sampling rates employed for data recording.”

May combine two claims.

In most previously discussed unsupervised and self-supervised HAR approaches, the source and target datasets are typically the same for most experiments, though pre-training does not use label information.

“In most of the previously discussed unsupervised and self-supervised approaches including Autoencoders, Multi-task self-supervision, Masked reconstruction, SimCLR, SimSiam, BYOL and CPC, the source and target datasets are typically the same for most experiments, albeit the pre-training does not use label information.”

May combine two claims.

MHEALTH and PAMAP2 contain activities expected in daily living and are collected at the leg/ankle rather than the wrist.

“MHEALTH and PAMAP2 both have activities that can be expected in daily living, albeit they are collected at the leg/ankle, rather than the wrist.”

• cites [2] 2. Oresti Banos, Rafael Garcia, Juan A Holgado-Terriza, Miguel Damas, Hector Pomares, Ignacio Rojas, Alejandro Saez, and Claudia Villalonga. 2014. mHealthDroid: a novel framework for agile development of mobile health applications. In International workshop on ambient assisted living. Springer, 91-98. sourceNot checked — no full text of the source
• cites [61] 61. Attila Reiss and Didier Stricker. 2012. Introducing a new benchmarked dataset for activity monitoring. In 2012 16th international symposium on wearable computers. IEEE, 108-109. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

In some cases, Mobiact is used for pre-training because it has the largest number of participants.

“In some cases, Mobiact [5] is used for pre-training instead as it is has the largest number of participants.”

• cites [5] 5. Charikleia Chatzaki, Matthew Pediaditis, George Vavoulas, and Manolis Tsiknakis. 2016. Human daily activity and fall recognition using a smartphone's acceleration sensor. In International Conference on Information and Communication Technologies for Ageing Well and e-Health. Springer, 100-118. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The target datasets are collected at waist/trousers or leg/ankle locations, while Capture-24 is collected at the wrist.

“As the Capture-24 dataset is collected at the wrist, we evaluate the learned weights at other on-body locations including the waist/trousers (where people often carry their smartphones), and the leg”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Random search is very effective compared to grid searches while not consuming extensive computational resources.

“We perform a random search over available combinations for Tab. 3 and 4, as it has been shown to be very effective compared to grid searches [3], while also not consuming extensive computational resources.”

• cites [3] 3. James Bergstra and Yoshua Bengio. 2012. Random search for hyper-parameter optimization. Journal of machine learning research 13, 2 (2012). sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The convolutional architecture utilized in many approaches in this study has long range filters of sizes 24, 16, and 8, and is likely more apt for the target datasets because they typically comprise more static activities such as walking, running, sitting, etc.

“It is also interesting to note that the convolutional architecture utilized in many approaches in this study, has long range filters (of sizes 24, 16, and 8), and is likely more apt for the target datasets, as they typically comprise of more static activities such as walking, running, sitting, etc.”

May combine two claims.

In healthcare, collecting large amounts of data may be difficult due to challenges with participant recruitment, compounded by the necessity to enlist healthcare professionals for annotation, which may incur high costs.

“For example, in healthcare, collecting large amounts of data may be difficult, due to challenges with participant recruitment.This is compounded by the necessity to enlist healthcare professionals for annotation, which may incur high costs [31].”

• cites [31] 31. Shruthi K Hiremath and Thomas Plötz. 2020. Deriving effective human activity recognition systems through objective task complexity assessment. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 4, 4 (2020), 1-24. sourceNot checked — no full text of the source

May combine two claims.

For self-supervised HAR methods, model pre-training using unlabeled training data is typically performed either on the same dataset as the target activity recognition system without using label information, or on a larger background dataset such as Mobiact.

“For these methods, model pre-training using unlabeled training data is typically either performed utilizing the same dataset as used for training the target activity recognition system-but without using label information-or with a larger background dataset such as Mobiact [5].”

• cites [5] 5. Charikleia Chatzaki, Matthew Pediaditis, George Vavoulas, and Manolis Tsiknakis. 2016. Human daily activity and fall recognition using a smartphone's acceleration sensor. In International Conference on Information and Communication Technologies for Ageing Well and e-Health. Springer, 100-118. sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Nine target datasets were curated with the goal of having a diverse collection of activities, number of participants, sensor locations, dataset sizes, and sensors utilized.

“In order to facilitate a broad evaluation of these approaches, we curate nine target datasets -with the goal of having a diverse collection of activities, number of participants, sensor locations, dataset sizes, and sensors utilized.”

Probably not a substantial claim: may not pass S4 (based).

Many wearable sensing datasets are inherently imbalanced, containing larger quantities of more easily performed activities such as walking or lying down relative to physically demanding movements such as skipping with a rope.

“many wearable sensing datasets are inherently imbalanced, containing larger quantities of more easily performed activities such as walking or lying down, relative to physically demanding movements such as skipping with a rope.”

Probably not a substantial claim: may not pass S4 (based).

In numerous target scenarios, data collection and subsequent annotation may be difficult due to cost or privacy issues.

“in numerous target scenarios, the data collection and subsequent annotation maybe difficult due to cost or privacy issues.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Wetlab contains laboratory experiments that are typically not performed in day-to-day living.

“Similarly, the Wetlab dataset contains laboratory experiments, which are typically not performed in day-to-day living.”

• cites [66] 66. Philipp M Scholl, Matthias Wille, and Kristof Van Laerhoven. 2015. Wearables in the wet lab: a laboratory system for capturing and guiding experiments. In Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing. 589-599. sourceNot checked — no full text of the source

May combine two claims.

The source dataset need not be artificially balanced for effective performance; the unavailability of coarse labels for balancing is not a crutch for learning useful representations.

“As such, this means that the source dataset need not be artificially balanced for effective performance; rather, the unavailability of even coarse labels (for balancing) is not a crutch for learning useful representations.”

• AssumptionRests on an assumption: Wearables-based movement datasets tend to be imbalanced, with more samples of naturally occurring or easily performed activities than strenuous or rare activities.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Using a smaller training set is advantageous because it can allow for reduced computation costs and more comprehensive hyperparameter searching.

“Using a smaller training is advantageous as it can allow for reduced computation costs and for more comprehensive hyperparameter searching.”

May combine two claims.

It is preferable to match the source dataset sampling rates because different sampling rates result in reduced performance.

“Overall, it is preferable to match the source dataset sampling rates as it different sampling rates result in reduced performance.”

• AssumptionRests on an assumption: The 50 Hz sampling rate retains sufficient information for the target activities.

May combine two claims.

The similarity of the later layers diverges, likely because supervised representations capture more activity-specific characteristics of the data, whereas self-supervised methods optimize different objectives entirely.

“The similarity of the later layers diverges, likely because the supervised representations capture the more activity-specific characterisics of the data, whereas the self-supervised methods optimize different objectives entirely.”

• AssumptionRests on an assumption: The gold-standard for activity recognition generally lies in supervised learning, resulting in the best possible performance.

May combine two claims.

Limited exposure to windows of the rarest class renders the Autoencoder unable to reconstruct them accurately, resulting in less effective representation learning.

“The pre-training is dependent on reconstructing the source data, and limited exposure to windows of the rarest class renders the model unable to reconstruct them accurately, thereby resulting in less effective representation learning.”

• AssumptionRests on an assumption: Wearables-based movement datasets tend to be imbalanced, with more samples of naturally occurring or easily performed activities than strenuous or rare activities.

May combine two claims.

A majority of the self-supervised methods including Multi-task self-supervision, CPC, Autoencoder, SimCLR, SimSiam, and BYOL do not require the entire training set (135 participants) for best performance, and can be trained for increased efficiency with fewer participants without much loss in performance.

“Thus, a majority of the self-supervised methods including Multi-task self-supervision, CPC, Autoencoder, SimCLR, SimSiam, and BYOL do not require the entire training set (135 participants) for best performance, and can be trained for increased efficiency with fewer participants without much loss in performance.”

May combine two claims.

Masked reconstruction struggles the most, likely because it depends on surrounding context to predict missing values.

“Masked reconstruction struggles the most, likely because it depends on surrounding context to predict the missing values.”

May combine two claims.

Self-supervision is more effective when transferring to wrist-based datasets (which match the source dataset location) or leg-based datasets.

“Therefore, self-supervision is more effective when transferring to the wristbased datasets (which match the source dataset location), or the leg.”

• AssumptionRests on an assumption: Very limited labeled data scenarios are of vital importance to activity recognition and can arise due to cost and privacy concerns, such as in medical applications.
• AssumptionRests on an assumption: Real-world deployment may allow for acquiring a small labeled dataset from users without significant interruptions, such as during device setup.

May combine two claims.

Masked Reconstruction is especially affected by using only one subject for pre-training because it uses a large transformer encoder and requires more data.

“This is especially true for Masked Reconstruction, which utilizes a large transformer encoder and thus requires more data.”

May combine two claims.

The comparable performance of all self-supervised methods on Daphnet FoG demonstrates their generalization capacity and their value towards utilization in conditions where labeled data may be difficult to obtain.

“The comparable performance of all self-supervised methods on such rarely accessible medical data demonstrates their generalization capacity and their value towards utilization in conditions where labeled data maybe difficult to obtain.”

• AssumptionRests on an assumption: The target datasets' activity labels are accurate and reliable.
• AssumptionRests on an assumption: The Capture-24 dataset is representative of free-living human activity, so representations learned on it are useful for other wearable HAR tasks.

May combine two claims.

Multi-task self-supervision and Masked reconstruction utilize the latent space more effectively and have higher implicit dimensionality.

“On the other hand, Multi-task self-supervision and Masked reconstruction utilize the latent space more effectively and have higher implicit dimensionality.”

May combine two claims.

The learned representations can generalize to unseen activities and improve performance even when sufficient labeled data is not available.

“This is encouraging as the learned representations can generalize to unseen activities and improve performance even when sufficient labeled data is not available.”

• AssumptionRests on an assumption: Very limited labeled data scenarios are of vital importance to activity recognition and can arise due to cost and privacy concerns, such as in medical applications.
• AssumptionRests on an assumption: Real-world deployment may allow for acquiring a small labeled dataset from users without significant interruptions, such as during device setup.

May combine two claims.

The Conv.classifier's better performance on very limited labels is likely due to its relatively small size and simple architecture, leading to less overfitting.

“We also note that the Conv.classifier typically performs better than the DeepConvLSTM when training on very limited labels likely due to its relatively small size and simple architecture, leading to less overfitting.”

• AssumptionRests on an assumption: Very limited labeled data scenarios are of vital importance to activity recognition and can arise due to cost and privacy concerns, such as in medical applications.
• AssumptionRests on an assumption: Real-world deployment may allow for acquiring a small labeled dataset from users without significant interruptions, such as during device setup.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SimSiam can be thought of as BYOL without the momentum encoder and SimCLR without the negative pairs, acting as the common connection point between several existing methods.

“As per the original paper [7], it can be thought of as 'BYOL without the momentum encoder' and 'SimCLR without the negative pairs', thereby acting as the common connection point between several existing methods.”

• cites [7] 7. Xinlei Chen and Kaiming He. 2021. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15750-15758. sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

For each target dataset fold, labeled windows per class were randomly selected for training while validation and test splits were untouched.

“For each target dataset fold, we randomly select {2, 5, 10, 50, 100} labeled windows per class for training whereas the validation and test splits are untouched.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The test-set F1-score was obtained after five-fold validation and five randomized runs.

“The test-set F1-score obtained after five-fold validation and five randomized runs is visualized in Fig. 11.”

Classified as probably not a substantial claim. May combine two claims.

Three randomized runs of pre-training and classification were performed, with five-fold test set F1-scores.

“Three randomized runs of pre-training (incl.the randomized subsampling) and classification are performed, and the mean and standard deviation of five-fold test set F1-scores are plotted in Fig. 4.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Three randomized runs were performed for choosing training windows, pre-training, and activity recognition.

“We perform three randomized runs for choosing the training windows, performing pre-training and activity recogntition, and plot the mean of the target test set F1-scores along with the standard deviation.”

Classified as probably not a substantial claim. May combine two claims.

Five randomized runs were performed for pre-training, with the most frequent and rarest classes chosen randomly.

“In practice, we perform five randomized runs for the pre-training and in each case, the most frequent and the rarest classes are chosen randomly, thereby making the aggregated performance resistant the choice of the majority class.”

May combine two claims.

The current version of the assessment framework does not take into consideration issues typically arising in multi-sensor datasets, such as variation in data distribution across devices and synchronization, because it only utilizes data from a single accelerometer.

“As we only utilize data from a single accelerometer for the assessment, the current version of the framework does not take into consideration issues typically arise in multi-sensor datasets, such as variation in data distribution across devices, synchronization etc.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SimSiam can collapse to trivial solutions, but the stop gradient operation can generally counteract such scenarios.

“The downside to this approach is that it can collapse to trivial solutions, but the stop gradient operation can generally counteract such scenarios.”

• cites [7] 7. Xinlei Chen and Kaiming He. 2021. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15750-15758. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Five randomized runs were used for sampling training windows and extracting representations.

“We randomly sample 1, 000 training windows in five randomized runs and extract the representations from the last encoder layer.”

Classified as probably not a substantial claim.

PCA was performed and cumulative explained variance was averaged across five randomized runs.

“Principal Component Analysis (PCA) is performed on the extracted features for each target dataset, and the cumulative fraction of the explained variance is averaged across the five randomized runs.”

May combine two claims.

Self-supervised methods have broad applicability in wide-ranging target scenarios and can empower human activity recognition where data collection and annotation are challenging.

“These results and findings are encouraging as they indicate the broad applicability of these self-supervised methods in wide-ranging target scenarios, and affirm the capability of these techniques for empowering human activity recognition in applications where data both collection and annotation may be challenging to accomplish.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Autoencoders may not result in the best possible unsupervised performance.

“The disadvantage of utilizing autoencoders is that it may not result the best possible unsupervised performance.”

Classified as probably not a substantial claim.

The validation users for Capture-24 were untouched.

“The validation users for Capture-24 are untouched.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Five-fold validation was used to obtain test set mean F1-scores.

“We plot the mean and standard deviation of the test set mean F1-score obtained after five-fold validation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SimCLR may be less effective on multi-sensory datasets due to its heavy reliance on augmentations designed specifically for accelerometry.

“One potential downside could be its effectiveness on multi-sensory datasets due to its heavy reliance on augmentations, as the established sensor augmentations were designed specifically for accelerometry (only).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Five randomized runs were performed for pre-training and linear classification.

“In Fig. 9, we perform five randomized runs for both the pre-training and the linear classification, and plot the difference in the mean of the train set F1-scores.”

Classified as probably not a substantial claim.

Three randomized runs were performed for choosing training subjects, pre-training, and classification on target datasets.

“We perform three randomized runs, including choosing training subject(s), pre-training, and classification on the target datasets.”