Universal Physiological Representation Learning with Soft-Disentangled Rateless Autoencoders

Deniz Erdogmus; Mo Han; Ozan Ozdenizci; Toshiaki Koike-Akino; Ye Wang

REVIEW 4 major objections 6 minor 29 references

Universal Physiological Representation Learning with Soft-Disentangled Rateless Autoencoders

T0 review · 4 major / 6 minor · reviewed 2026-08-27 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Soft-disentangled rateless autoencoders improve cross-subject stress classification by up to 11.6 percentage points.

desk verdict Competent incremental method paper with a plausible central claim, but the 11.6% transfer gain is likely inflated by sample-level validation leakage in autocorrelated time series. read the letter →

arxiv 2009.13453 v1 pith:J4NE5UFO submitted 2020-09-28 eess.SP cs.LG

Mo Han , Ozan Ozdenizci , Toshiaki Koike-Akino , Ye Wang , Deniz Erdogmus This is my paper · ORCID

classification eess.SPcs.LG

keywords ratelessautoencoderstochasticbottlenecksoftdisentanglementadversarialtransferlearningphysiologicalbiosignalssubject-invariantrepresentationstressclassificationleave-one-subject-outevaluation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a single feature encoder can produce physiological representations that transfer across new users by decoupling task-relevant from subject-specific information in a soft, probabilistic way. It extends the rateless autoencoder with two adversarial heads, one that hides subject identity and one that preserves it, assigned to each latent unit through complementary dropout probabilities. On a 20-subject, four-class wrist-biosignal stress dataset, using leave-one-subject-out evaluation, the full model (DA-cRAE) outperforms plain and conditional autoencoders, reaching up to 11.6 percentage points higher average accuracy with an MLP classifier. The paper also claims this soft-split design reduces accuracy variance across left-out subjects and across classifiers, and is more robust to latent dimension and training-data size than hard-split or non-adversarial baselines. A sympathetic reader would care because subject-to-subject variability is the main obstacle to practical biosignal-based human-computer interaction.

What carries the argument

The central object is the rateless autoencoder's stochastic bottleneck, used as a soft switch between two adversarial heads. A probabilistic dropout schedule $p_a(d)=((d-1)/(D-1))^\alpha$ with $\alpha=3$ sends latent unit $d$ to the adversary with probability $p_a(d)$ and to the nuisance block with probability $p_n(d)=1-p_a(d)$; the hard split is the boundary case where each $p_a(d)$ is 0 or 1. Because each training pass uses a stochastically reduced effective latent dimension, the model behaves as an ensemble of autoencoders with different task/subject splits, and the complementarity $p_a+p_n=1$ guarantees every latent unit is accounted for by one of the two roles. The objective in Eq. (1) couples decoder reconstruction, nuisance subject-ID likelihood, and adversary subject-ID likelihood; the frozen encoder from that training is the feature extractor handed to the downstream classifier. This ensemble property is what the paper credits for robustness to the latent dimension parameter and for eliminating the need to choose the split ratio by hand.

What would settle it

A direct test would run DA-cRAE and plain AE on a leave-one-subject-out split of a dataset where subject ID and task label are intentionally correlated, for example each subject performs only one stressor or demographic groups are assigned different tasks. If the soft-disentangled encoder's accuracy falls to or below the AE baseline while the nuisance network still decodes subject ID accurately, the independence assumption is carrying the claim. Alternatively, measuring $I(y;s)$ in the training data and showing that DA-cRAE's accuracy gain decreases as $I(y;s)$ grows would settle the mechanism.

Watch

Extended reading notes

Core claim

The paper claims that soft disentanglement, rather than a fixed partition of the latent code, is the mechanism that makes adversarial transfer learning work for physiological signals. In DA-cRAE, the encoder outputs a $D$-dimensional latent $z$; each unit $d$ is routed to the adversary with dropout probability $p_a(d)$ and to the nuisance with probability $p_n(d)=1-p_a(d)$, and both adversarial networks are trained to predict subject ID, while the decoder reconstructs $X$ from $z$ and the subject ID $s$. Training minimizes the reconstruction error plus $\lambda_N$ times the nuisance's subject-ID likelihood, and at the same time weakens the adversary's ability to decode $s$. The resulting frozen encoder is then used with a simple task classifier. On the non-EEG wrist-biosignal stress dataset with 20 subjects and four stress classes, leave-one-subject-out average accuracy reaches 83.8% with an MLP, 11.6 percentage points above the plain AE baseline's 72.2%, and the paper reports accuracy gains across nearest-neighbor, decision tree, LDA, and logistic-regression classifiers as well. It also reports improved worst-case held-out-subject accuracy, which the paper reads as evidence that the stochastic routing keeps task-relevant information while shedding subject-specific variation without needing a pre-chosen split ratio.

Load-bearing premise

The method assumes task label $y$ and subject ID $s$ are independent, so removing subject-specific variation from the task branch cannot remove task-relevant signal; if stress protocols make identity and task correlated, transfer gains could shrink or reverse.

Editorial extensions

If this is right

For any of the five classifiers tested, replacing the plain AE feature encoder with the DA-cRAE encoder raises average leave-one-subject-out stress classification accuracy, with the largest gain being 11.6 percentage points with MLP.
The soft-split models A-cRAE, D-cRAE, and DA-cRAE all outperform their hard-split counterparts, so practitioners no longer need to pre-specify an adversary/nuisance split ratio.
The frozen encoder stays usable when training data are cut to 10% of the full set, with the ordering of models preserved, which matters for biosignal applications with few labeled subjects.
Across latent dimensions from $D=3$ to $D=25$, DA-cRAE's average accuracy is more stable than AE's, so the method is less sensitive to choosing the bottleneck width.
Worst-case held-out-subject accuracy improves, so the model reduces the risk of failing on the hardest users rather than only raising the average.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper fixes the soft-split schedule with $\alpha=3$; an immediate extension is to treat $\alpha$ or the whole $p_a$ curve as a learned parameter, letting the data decide how aggressively to shed subject information.
The same adversary/nuisance pairing need not be limited to subject identity; recording session, sensor location, or demographic group could play the role of $s$, with the same independence caveat.
Because the decoder is conditioned on subject ID while the frozen encoder is not, the architecture implicitly assumes that subject information needed for reconstruction is available at decode time but is not needed for task classification; testing reconstruction fidelity on held-out subjects would directly probe how much subject information remains in $z$.
The evaluation uses one dataset; the strongest testable claim is that the same frozen encoder transfers across datasets or tasks, which the paper does not run.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a solid incremental paper, not a breakthrough. The soft-disentangled rateless autoencoder (DA-cRAE) combines the authors' prior rateless autoencoder and disentangled adversarial autoencoder, with the novelty being a complementary dropout scheme that softly assigns latent units to adversary and nuisance networks. That is a modest but real extension, and the paper is clearly written. The experiments compare against eight baselines across five classifiers on the PhysioNet non-EEG stress dataset, with leave-one-subject-out evaluation. The consistent gains in Table II and Fig. 3 suggest the method has some value, and the central classification-accuracy claim is not circular: it is an independent downstream measurement.

The soft spot is the validation procedure, and it is load-bearing. Section III-B says that for the non-held-out subjects, "90% and 10% of the data ... were randomly split as the training and validation sets." The data are downsampled to 1 Hz from 5-minute tasks, so adjacent samples are strongly autocorrelated. A random sample-level split puts near-duplicates of training samples into the validation set, which means all hyperparameter choices (lambda_A, lambda_N, latent dimension, early stopping) are selected on a validation estimate that rewards fitting subject-specific temporal structure. With a single random split and no repeated-seed variance or statistical tests, the headline 11.6% gap between DA-cRAE and AE cannot be distinguished from selection noise. A trial-level or block-wise split would be the right check; it is not reported. The independence assumption y perp s is also asserted without testing; if subjects have different baseline stress responses, the task classifier could be exploiting subject-specific cues. Table III's low adversary accuracy and high nuisance accuracy are partly by construction, since the objective directly trains the encoder to those ends.

I disagree with any claim that the paper is fundamentally flawed. The method is sensible, the writing is honest, and the empirical question is worth answering properly. But the current evaluation is too weak for the "universal" framing. The authors need to release code, run repeated seeds, and redo validation with trial-level splits. This deserves a serious referee, not a desk reject, because the idea is plausible and the presentation is clear, but it should come back with major revision requests. I would bring it to a reading group as a good case study of temporal leakage in physiological transfer learning, but I would not cite it in my own work until the evaluation is cleaned up.

Referee Report

4 major / 6 minor

Summary. The paper proposes a universal physiological representation learning framework based on rateless autoencoders (RAE) with soft disentanglement via adversarial and nuisance networks. The encoder is trained with a reconstruction loss plus an adversary loss that minimizes subject-ID predictability from task-relevant latent units and a nuisance loss that maximizes subject-ID predictability from subject-specific latent units, with dropout-rate scheduling providing a soft split. The authors evaluate on a public non-EEG stress dataset with leave-one-subject-out cross-validation and multiple downstream classifiers, reporting up to an 11.6% average accuracy improvement over a baseline AE and reduced cross-subject variance.

Significance. If the empirical claims hold, the paper makes a useful contribution: the rateless/soft-split mechanism offers a principled alternative to manually fixing the hard split of latent dimensions, and the framework is evaluated across several classifiers on a public dataset. The LOO protocol is appropriate in spirit, and the comparison against multiple baselines is thorough. However, the evaluation has a load-bearing flaw: the 90/10 sample-level random split of strongly autocorrelated time series makes the validation estimates optimistic, and the absence of repeated seeds or statistical tests undermines the headline quantitative claim as stated. The disentanglement numbers in Table III are partially by construction given the objective in Eq. (1), so they should not be presented as independent evidence. The central methodological idea is sound and the reported results are plausible, but the empirical evidence needs substantial revision before the claims can be accepted.

major comments (4)

[Section III-B, III-C.1, Table II] The 90/10 random sample-level split of the remaining subjects' data is problematic because the 1-Hz physiological signals are strongly autocorrelated within each 5-minute trial (four trials per subject). Randomly assigning adjacent samples to training versus validation does not produce an independent validation set, so model-selection decisions (λA, λN, D, early stopping) are made against a validation estimate that rewards exploitation of subject-specific temporal structure rather than universal, transferable features. This can inflate the reported cross-subject accuracy gains; a block-wise or trial-level split should be used instead, and the analysis should be repeated with multiple random seeds to report variance and statistical significance.
[Section III-C.1] The claim of 'significant improvements in average accuracy up to 11.6%' is not supported by any statistical test, confidence interval, or repeated-seed variance; with a single random split for each of the 20 folds, the difference between DA-cRAE and AE (Table II) cannot be distinguished from selection noise. The authors should report mean and standard deviation over at least 10 random seeds and perform paired significance tests across the 20 held-out subjects.
[Section II-A] The assumption that task label y and subject ID s are marginally independent is asserted but not tested. If the stress protocol induces subject-specific baselines that correlate with y, or if the class distribution varies across subjects, the adversary's removal of subject-specific variation could also remove task-relevant information, limiting transferability. A concrete check is to report the per-subject label distribution and the empirical correlation between s and y in the dataset, or to compare the proposed method against a version without adversarial removal on a subgroup analysis.
[Section III-C.2, Table III] The adversary and nuisance accuracies in Table III are partially by construction, since Eq. (1) directly optimizes the encoder to minimize q_φ(s|z,p_a) and maximize q_ψ(s|z,p_n); these numbers therefore do not provide independent evidence of disentanglement. The downstream task-classification accuracy is the relevant measure, but its reliability is affected by the validation issue raised above.

minor comments (6)

[Fig. 3 caption] The box plots do not indicate the number of random seeds or the exact definition of whiskers and outliers; please clarify so that the reader can assess the variance across the 20 folds.
[Section II-E.3] There is a typo: 'A-cAE resembles to the' should be 'A-cAE resembles the'.
[Section III-C.4] The phrase 'no significant drawback' in the discussion of Fig. 5 uses 'significant' without any supporting statistical test; please either add a test or rephrase.
[Equation (1)] The symbol λAE appears in the equation while the surrounding text defines λA; please unify the notation.
[Section III-B] It is unclear whether the parameter sweeps for each classifier in Table II followed the same sequential procedure described for the MLP classifier; please state the exact protocol used for each classifier.
[Reference [19]] The reference for 'Disentangled adversarial autoencoder for subject-invariant physiological feature extraction' lacks full bibliographic details (volume, pages, or DOI); please complete it.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential reporting in Table III, but no load-bearing circularity: the headline transfer-accuracy gain is a downstream held-out-subject measurement, not forced by Eq. (1).

self definitional [Section III-C.2, Table III; objective in Eq. (1), Section II-C.1]
"With an increasing accuracy of MLP task classiﬁer, stress levels are better discriminated; with a growing accuracy of nuisance network, more person-discriminative features are preserved in the nuisance counterpart; and with a decreasing accuracy of adversary network, more task-speciﬁc information are inherent in the adversary counterpart. ... We observe that the nuisance network produces higher accuracy with increasing λN, where λN = 0.05 particularly results in the better performance on task classiﬁcation."

These 'observations' are direct consequences of Eq. (1): LossRAE adds +λA E[log qφ(s|z,pa)] and −λN E[log qψ(s|z,pn)], so the encoder is explicitly optimized to make adversary accuracy fall as λA grows and nuisance accuracy rise as λN grows. Reporting those expected gradient directions as empirical evidence of the disentangling mechanism is therefore self-referential. The metric moves in the direction the loss was constructed to move. This is ancillary rather than load-bearing: the headline 11.6% accuracy improvement comes from a separate held-out-subject task classifier that Eq. (1) does not directly optimize.

full rationale

The central claim of the paper is the cross-subject task-classification accuracy gain of DA-cRAE over AE, reported in Fig. 3 and Table II with leave-one-subject-out evaluation. That accuracy is measured by a downstream classifier trained on frozen encoder features; it is not mathematically forced by Eq. (1), which only specifies reconstruction, adversary, and nuisance losses. The RAE concept [27] and hard-split disentangled baselines [18], [19] come from prior work by overlapping authors, but they are used transparently as building blocks and comparison methods, not as an unverified uniqueness theorem or as a self-citation chain that forbids alternatives. The by-construction behavior of the adversary and nuisance discriminator accuracies in Table III is a minor self-referential reporting issue, but it does not determine the independent classification result. The reader's concern about random sample-level splits of autocorrelated physiological time series and possible validation leakage is a correctness/experimental-design risk, not a circularity of the derivation chain; it would not change this circularity verdict. Overall, the paper's main empirical derivation is self-contained, with one minor non-load-bearing circular reporting element.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central contribution is empirical, so its main axioms are modeling assumptions and hand-set hyperparameters. The paper does not introduce new physical entities; 'adversary', 'nuisance', and 'rateless' blocks are architecture choices. The y-independence assumption is load-bearing but untested, and D, alpha, lambda_A, lambda_N, and the hard-split ratio are all chosen by hand or by validation sweep rather than derived.

free parameters (5)

latent dimension D = 15
D=15 is used for all main results; Fig. 4 sweeps D and shows stable performance, but the value is a design choice, not derived.
dropout exponent alpha = 3
The soft-split schedule is pa(d)=((d-1)/(D-1))^alpha with alpha=3; chosen by hand in Section II-E.2 with no stated criterion or sensitivity analysis.
adversary weight lambda_A = varies per classifier, e.g., 0.5 for MLP
Selected by a two-stage parameter sweep on a validation split (Section III-B, Table II).
nuisance weight lambda_N = varies per classifier, e.g., 0.05 for MLP
Selected in the same sweep; lambda_N=0.05 is used as the reference for DA-cRAE in Table III.
hard split ratio za:zn = 2:1
Used only for the DA-cAE hard-split baseline; chosen without stated justification in Section II-E.2.

assumptions (4)

domain assumption Task label y and subject ID s are marginally independent, and X is generated as X ~ p(X|y,s).
Stated in Section II-A; the adversarial disentanglement objective relies on this to keep task information while removing subject information.
ad hoc to paper The reconstruction loss plus the adversary and nuisance losses yields latent features that separate subject and task information.
This is the design assumption encoded in Eq. (1); no theoretical guarantee is given that the two adversarial heads achieve clean disentanglement.
ad hoc to paper The stochastic dropout schedule pa(d)=((d-1)/(D-1))^alpha orders latent units by subject versus task relevance.
The paper assumes that varying dropout probabilities across units creates a smooth task-to-subject spectrum; this is not derived from data properties.
domain assumption A random 90/10 split of individual 1 Hz samples provides a valid estimate of cross-subject generalization.
Section III-B randomly splits samples from training subjects; temporal autocorrelation within trials could leak if samples from the same trial appear in both train and validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Universal Physiological Representation Learning with Soft-Disentangled Rateless Autoencoders." pith.science (2026). https://pith.science/paper/J4NE5UFO

@misc{pith2026200913453,
  author       = {Pith},
  title        = {Pith review of: Universal Physiological Representation Learning with Soft-Disentangled Rateless Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4NE5UFO}},
  note         = {Machine review of arXiv:2009.13453}
}

read the original abstract

Human computer interaction (HCI) involves a multidisciplinary fusion of technologies, through which the control of external devices could be achieved by monitoring physiological status of users. However, physiological biosignals often vary across users and recording sessions due to unstable physical/mental conditions and task-irrelevant activities. To deal with this challenge, we propose a method of adversarial feature encoding with the concept of a Rateless Autoencoder (RAE), in order to exploit disentangled, nuisance-robust, and universal representations. We achieve a good trade-off between user-specific and task-relevant features by making use of the stochastic disentanglement of the latent representations by adopting additional adversarial networks. The proposed model is applicable to a wider range of unknown users and tasks as well as different classifiers. Results on cross-subject transfer evaluations show the advantages of the proposed framework, with up to an 11.6% improvement in the average subject-transfer classification accuracy.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

[1]

Physiological signals based human emotion recognition: a review,

S. Jerritta, M. Murugappan, R. Nagarajan, and K. Wan, “Physiological signals based human emotion recognition: a review,” in IEEE 7th International Colloquium on Signal Processing and its Applications . IEEE, 2011, pp. 410–415

work page 2011
[2]

Classiﬁcation of ecg signals using machine learning techniques: A survey,

S. H. Jambukia, V . K. Dabhi, and H. B. Prajapati, “Classiﬁcation of ecg signals using machine learning techniques: A survey,” in 2015 International Conference on Advances in Computer Engineering and Applications. IEEE, 2015, pp. 714–721

work page 2015
[3]

Hands: a multimodal dataset for modeling toward human grasp intent inference in prosthetic hands,

M. Han, S. Y . G ¨unay, G. Schirner, T. Padır, and D. Erdo ˘gmus ¸, “Hands: a multimodal dataset for modeling toward human grasp intent inference in prosthetic hands,” Intelligent Service Robotics , vol. 13, no. 1, pp. 179–185, 2020

work page 2020
[4]

Emotion recognition from EEG using higher order crossings,

P. C. Petrantonakis and L. J. Hadjileontiadis, “Emotion recognition from EEG using higher order crossings,” IEEE Transactions on Information Technology in Biomedicine , vol. 14, no. 2, pp. 186–197, 2009

work page 2009
[5]

A non- EEG biosignals dataset for assessment and visualization of neurological status,

J. Birjandtalab, D. Cogan, M. B. Pouyan, and M. Nourani, “A non- EEG biosignals dataset for assessment and visualization of neurological status,” in IEEE International Workshop on Signal Processing Systems , 2016, pp. 110–114

work page 2016
[6]

Emotional reactivity monitoring using electrodermal activity analysis in individuals with suicidal behaviors,

A. M. Amiri, M. Abtahi, A. Rabasco, M. Armey, and K. Mankodiya, “Emotional reactivity monitoring using electrodermal activity analysis in individuals with suicidal behaviors,” in 10th International Symposium on Medical Information and Communication Technology , 2016, pp. 1–5

work page 2016
[7]

A wrist-worn biosensor system for assessment of neurological status,

D. Cogan, M. B. Pouyan, M. Nourani, and J. Harvey, “A wrist-worn biosensor system for assessment of neurological status,” in 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society , 2014, pp. 5748–5751

work page 2014
[8]

Subject-dependent biosig- nal features for increased accuracy in psychological stress detection,

D. Giakoumis, D. Tzovaras, and G. Hassapis, “Subject-dependent biosig- nal features for increased accuracy in psychological stress detection,” International Journal of Human-Computer Studies , vol. 71, no. 4, pp. 425–439, 2013

work page 2013

Show all 29 references

[9]

Review on psychological stress detection using biosignals,

G. Giannakakis, D. Grigoriadis, K. Giannakaki, O. Simantiraki, A. Roni- otis, and M. Tsiknakis, “Review on psychological stress detection using biosignals,” IEEE Transactions on Affective Computing , 2019

2019
[10]

Time-series prediction of proximal aggression onset in minimally-verbal youth with autism spectrum disorder using physiological biosignals,

O. ¨Ozdenizci et al. , “Time-series prediction of proximal aggression onset in minimally-verbal youth with autism spectrum disorder using physiological biosignals,” in 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society , 2018, pp. 5745– 5748

2018
[11]

Subject-independent mental state classiﬁcation in single trials,

S. Fazli, F. Popescu, M. Dan ´oczy, B. Blankertz, K.-R. M ¨uller, and C. Grozea, “Subject-independent mental state classiﬁcation in single trials,” Neural Networks , vol. 22, no. 9, pp. 1305–1312, 2009

2009
[12]

Learning a common dictionary for subject-transfer decoding with resting calibration,

H. Morioka, A. Kanemura, J.-I. Hirayama, M. Shikauchi, T. Ogawa, S. Ikeda, M. Kawanabe, and S. Ishii, “Learning a common dictionary for subject-transfer decoding with resting calibration,” NeuroImage, vol. 111, pp. 167–178, 2015

2015
[13]

Physiological-signal-based mental workload estimation via transfer dy- namical autoencoders in a deep learning framework,

Z. Yin, M. Zhao, W. Zhang, Y . Wang, Y . Wang, and J. Zhang, “Physiological-signal-based mental workload estimation via transfer dy- namical autoencoders in a deep learning framework,” Neurocomputing, vol. 347, pp. 212–229, 2019

2019
[14]

Cross-subject driver status detection from physiological signals based on hybrid feature selection and transfer learning,

L.-L. Chen, A. Zhang, and X.-G. Lou, “Cross-subject driver status detection from physiological signals based on hybrid feature selection and transfer learning,” Expert Systems with Applications , vol. 137, pp. 266–280, 2019

2019
[15]

Learning invariant representations from EEG via adversarial inference,

O. ¨Ozdenizci, Y . Wang, T. Koike-Akino, and D. Erdo ˘gmus ¸, “Learning invariant representations from EEG via adversarial inference,” IEEE Access, vol. 8, pp. 27 074–27 085, 2020

2020
[16]

Transfer learning in brain-computer interfaces with adversarial variational autoencoders,

——, “Transfer learning in brain-computer interfaces with adversarial variational autoencoders,” in 2019 9th International IEEE/EMBS Con- ference on Neural Engineering (NER) . IEEE, 2019, pp. 207–210

2019
[17]

Adversarial deep learning in EEG biometrics,

——, “Adversarial deep learning in EEG biometrics,” IEEE Signal Processing Letters, vol. 26, no. 5, pp. 710–714, 2019

2019
[18]

Disentangled adversarial transfer learning for physiological biosignals,

M. Han, O. ¨Ozdenizci, Y . Wang, T. Koike-Akino, and D. Erdo ˘gmus ¸, “Disentangled adversarial transfer learning for physiological biosignals,” in 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) , 2020

2020
[19]

Disentangled adversarial autoencoder for subject-invariant phys- iological feature extraction,

——, “Disentangled adversarial autoencoder for subject-invariant phys- iological feature extraction,” in IEEE Signal Processing Letters , 2020

2020
[20]

Censoring representations with an adver- sary,

H. Edwards and A. Storkey, “Censoring representations with an adver- sary,” arXiv preprint arXiv:1511.05897 , 2015

2015 arXiv
[21]

Modality-speciﬁc and shared generative adversarial network for cross-modal retrieval,

F. Wu, X.-Y . Jing, Z. Wu, Y . Ji, X. Dong, X. Luo, Q. Huang, and R. Wang, “Modality-speciﬁc and shared generative adversarial network for cross-modal retrieval,” Pattern Recognition, p. 107335, 2020

2020
[22]

Ad- versarial learning for cross-project semi-supervised defect prediction,

Y . Sun, X.-Y . Jing, F. Wu, J. Li, D. Xing, H. Chen, and Y . Sun, “Ad- versarial learning for cross-project semi-supervised defect prediction,” IEEE Access , vol. 8, pp. 32 674–32 687, 2020

2020
[23]

Fader networks: Manipulating images by sliding at- tributes,

G. Lample, N. Zeghidour, N. Usunier, A. Bordes, L. Denoyer, and M. Ranzato, “Fader networks: Manipulating images by sliding at- tributes,” in Advances in Neural Information Processing Systems , 2017, pp. 5967–5976

2017
[24]

Recent advances in autoencoder-based representation learning,

M. Tschannen, O. Bachem, and M. Lucic, “Recent advances in autoencoder-based representation learning,” arXiv preprint arXiv:1812.05069, 2018

2018 arXiv
[25]

Deep convolution neural network and autoencoders-based unsupervised feature learning of eeg signals,

T. Wen and Z. Zhang, “Deep convolution neural network and autoencoders-based unsupervised feature learning of eeg signals,” IEEE Access, vol. 6, pp. 25 399–25 410, 2018

2018
[26]

Improving the classiﬁcation effectiveness of intrusion detection by using improved conditional vari- ational autoencoder and deep neural network,

Y . Yang, K. Zheng, C. Wu, and Y . Yang, “Improving the classiﬁcation effectiveness of intrusion detection by using improved conditional vari- ational autoencoder and deep neural network,” Sensors, vol. 19, no. 11, p. 2528, 2019

2019
[27]

Stochastic bottleneck: Rateless auto- encoder for ﬂexible dimensionality reduction,

T. Koike-Akino and Y . Wang, “Stochastic bottleneck: Rateless auto- encoder for ﬂexible dimensionality reduction,” in 2020 IEEE Interna- tional Symposium on Information Theory (ISIT) . IEEE, 2020

2020
[28]

Deep learning with convo- lutional neural networks applied to electromyography data: A resource for the classiﬁcation of movements for prosthetic hands,

M. Atzori, M. Cognolato, and H. M ¨uller, “Deep learning with convo- lutional neural networks applied to electromyography data: A resource for the classiﬁcation of movements for prosthetic hands,” Frontiers in Neurorobotics, vol. 10, p. 9, 2016

2016
[29]

Deep learning for healthcare applications based on physiological signals: A review,

O. Faust, Y . Hagiwara, T. J. Hong, O. S. Lih, and U. R. Acharya, “Deep learning for healthcare applications based on physiological signals: A review,”Computer Methods and Programs in Biomedicine , vol. 161, pp. 1–13, 2018

2018

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

Physiological signals based human emotion recognition: a review,

S. Jerritta, M. Murugappan, R. Nagarajan, and K. Wan, “Physiological signals based human emotion recognition: a review,” in IEEE 7th International Colloquium on Signal Processing and its Applications . IEEE, 2011, pp. 410–415

work page 2011

[2] [2]

Classiﬁcation of ecg signals using machine learning techniques: A survey,

S. H. Jambukia, V . K. Dabhi, and H. B. Prajapati, “Classiﬁcation of ecg signals using machine learning techniques: A survey,” in 2015 International Conference on Advances in Computer Engineering and Applications. IEEE, 2015, pp. 714–721

work page 2015

[3] [3]

Hands: a multimodal dataset for modeling toward human grasp intent inference in prosthetic hands,

M. Han, S. Y . G ¨unay, G. Schirner, T. Padır, and D. Erdo ˘gmus ¸, “Hands: a multimodal dataset for modeling toward human grasp intent inference in prosthetic hands,” Intelligent Service Robotics , vol. 13, no. 1, pp. 179–185, 2020

work page 2020

[4] [4]

Emotion recognition from EEG using higher order crossings,

P. C. Petrantonakis and L. J. Hadjileontiadis, “Emotion recognition from EEG using higher order crossings,” IEEE Transactions on Information Technology in Biomedicine , vol. 14, no. 2, pp. 186–197, 2009

work page 2009

[5] [5]

A non- EEG biosignals dataset for assessment and visualization of neurological status,

J. Birjandtalab, D. Cogan, M. B. Pouyan, and M. Nourani, “A non- EEG biosignals dataset for assessment and visualization of neurological status,” in IEEE International Workshop on Signal Processing Systems , 2016, pp. 110–114

work page 2016

[6] [6]

Emotional reactivity monitoring using electrodermal activity analysis in individuals with suicidal behaviors,

A. M. Amiri, M. Abtahi, A. Rabasco, M. Armey, and K. Mankodiya, “Emotional reactivity monitoring using electrodermal activity analysis in individuals with suicidal behaviors,” in 10th International Symposium on Medical Information and Communication Technology , 2016, pp. 1–5

work page 2016

[7] [7]

A wrist-worn biosensor system for assessment of neurological status,

D. Cogan, M. B. Pouyan, M. Nourani, and J. Harvey, “A wrist-worn biosensor system for assessment of neurological status,” in 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society , 2014, pp. 5748–5751

work page 2014

[8] [8]

Subject-dependent biosig- nal features for increased accuracy in psychological stress detection,

D. Giakoumis, D. Tzovaras, and G. Hassapis, “Subject-dependent biosig- nal features for increased accuracy in psychological stress detection,” International Journal of Human-Computer Studies , vol. 71, no. 4, pp. 425–439, 2013

work page 2013