Pith. sign in

REVIEW 5 major objections 5 minor 43 references

Sensor Drift Compensation in Electronic-Nose-Based Gas Recognition Using Knowledge Distillation

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Knowledge distillation, a teacher-student training method, compensates for electronic-nose sensor drift better than the previous state-of-the-art DRCA method, with up to 18% higher accuracy and 15% higher F1-score over 30 random test…

desk verdict Useful first application of KD to e-nose drift with a thoughtful evaluation protocol, but the headline claim overreaches: the stats never directly compare KD to DRCA. read the letter →

arxiv 2507.17071 v1 pith:4H3M4FCZ submitted 2025-07-22 cs.LG cs.SYeess.SPeess.SYphysics.ins-det

classification cs.LGcs.SYeess.SPeess.SYphysics.ins-det
keywords driftcompensationelectronicnoseknowledgedistillationsemi-supervisedlearningunsupervisedgasrecognitionsensordomainadaptation
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

Electronic noses—arrays of chemical sensors that identify gases—lose accuracy as their sensor responses drift over time and with changing environments. This paper argues that knowledge distillation (KD) compensates for that drift better than Domain Regularized Component Analysis (DRCA), the previous best method, which aligns source and target feature distributions by projection. In the KD setup, a teacher network trained on clean source readings outputs soft label distributions for unlabeled target readings, and a student network trained on those soft labels generalizes better to the drifted target. Using the standard ten-batch, six-class gas-sensor drift dataset, the authors run two cross-domain tasks—batch 1 predicts all later batches, and all prior batches predict the next batch—with 30 random partitions of each target batch. They report that KD consistently outperforms DRCA and a KD-DRCA hybrid, with up to 18% higher accuracy and 15% higher F1-score, and call this the first application of KD to electronic-nose drift mitigation. If correct, the result gives deployed e-nose systems a label-light way to stay accurate as sensors age.

What carries the argument

The load-bearing mechanism is the teacher-student distillation loop: a supervised teacher $f_S$ trained on labeled source samples produces soft labels $\sigma(T^{-1} f_S(x))$ for every sample in the union of source and unlabeled target data, and a student $f_T$ is trained to match those softened probabilities through the distillation loss. The temperature $T>1$ controls how much inter-class similarity the soft labels preserve, which is what lets the student keep class-related structure while adapting to drifted readings. DRCA, the comparison method, projects both domains with a matrix $P$ obtained from an eigenvalue problem $S_b^{-1}(S_w^S + \alpha S_w^T)P = \theta P$, minimizing between-domain scatter while preserving within-domain scatter; the hybrid method feeds DRCA-projected features into the same KD loop. This teacher-student loop is the central object that carries the argument, and using it for e-nose drift compensation is the new step.

What would settle it

A direct test would rerun both tasks with hyperparameters chosen without any target labels—for example, fixing temperature, subspace dimension, and regularization from source-only validation or unsupervised criteria—and compare KD, DRCA, and KD-DRCA over the same 30 partitions. If KD's advantage over DRCA shrinks to statistical parity, the claim that KD is intrinsically better at drift compensation would be falsified; if the advantage survives, the tuning caveat is minor.

Watch

Extended reading notes

Core claim

The central claim is that knowledge distillation—not domain-alignment projection—is the most effective drift-compensation strategy for electronic-nose gas recognition on the gas-sensor drift benchmark. The teacher model is trained in a supervised way on source-domain readings and emits soft probability vectors for a union of source and unlabeled target readings; the student is trained to match those soft outputs under a temperature-smoothed distillation loss. The paper's experiments compare this KD method with DRCA, a subspace-projection method that minimizes between-domain scatter while preserving within-domain scatter, and with a hybrid that applies DRCA projection before KD. Across two tasks (predicting all later batches from batch 1, and predicting the next batch from all previous batches) and 30 random 50/50 validation-test partitions of each target batch, KD yields the most statistically significant improvements over the no-drift-compensation baseline, and KD alone beats both DRCA and the hybrid on accuracy, precision, recall, and F1-score—up to 18% and 15% gains respectively. The paper maintains that previous DRCA results on this dataset were not statistically validated and that its protocol supplies that missing rigor.

Load-bearing premise

The load-bearing premise is that choosing the method's settings by checking accuracy on a labeled half of each target batch does not inflate the reported test numbers; in a real deployment, no target labels would exist to guide that choice.

Editorial extensions

If this is right

  • If KD's advantage holds, deployed e-nose systems can be recalibrated using only unlabeled readings from the new environment, by letting the old model produce soft labels and retraining a student.
  • Because KD also wins when training on all accumulated prior batches (task 2), it fits continuous, online recalibration workflows.
  • The hybrid's failure to beat KD alone indicates that projection-based domain alignment and model-perspective knowledge transfer do not simply combine, and aggressive alignment can discard class-related variance.
  • The 30-partition protocol with paired significance tests provides a template for evaluating future drift-compensation methods on accuracy, precision, recall, and F1-score rather than a single split.

Reading between the lines

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

  • Our inference: the reported margins are likely sensitive to the target-label-based hyperparameter selection, so a fully unsupervised protocol would be a stricter test of KD's advantage.
  • Our inference: because soft labels preserve inter-class similarities, KD should degrade more gracefully than projection-based alignment as drift increases; task 1's declining accuracy over later batches could be analyzed specifically to test this.
  • Our inference: the teacher-student recipe is not gas-specific, so the same protocol could transfer to other sensor-array drift problems, such as inertial measurement units or surface electromyography, where DRCA has previously been applied.
  • Our inference: adding a reliability filter on the teacher's soft outputs, rather than using all of them, might push performance closer to the roughly 0.98 within-batch accuracy ceiling; the paper does not test this variant.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. This paper studies sensor-drift compensation for electronic-nose gas classification on the UCI Gas Sensor Array Drift Dataset. The authors design two cross-domain tasks (Batch 1 as source for all later batches, and all prior batches as source for the next batch), then compare a knowledge-distillation (KD) method, Domain Regularized Component Analysis (DRCA), a KD-DRCA hybrid, and a no-compensation baseline. For each method they report counts of statistically significant better/equal/worse outcomes relative to the baseline across 72 task-batch-metric cells, computed from 30 random target splits. The abstract claims that KD consistently outperformed DRCA and KD-DRCA, with up to 18% accuracy and 15% F1-score improvement, and that this is the first KD application for e-nose drift mitigation.

Significance. If the empirical claim were sound, the paper would offer a practical and simple alternative to DRCA for unsupervised sensor-drift compensation, and its proposed repeated-split evaluation protocol would be a useful template for the field. The paper has some strengths: it uses a public benchmark, defines two realistic drift scenarios, performs 30 random partitions, reports four classification metrics, and makes an explicit attempt at statistical testing. However, the central comparative claim is not supported by the reported tests, the evaluation protocol uses target-domain labels for model selection, and the quantitative headline numbers are absent from the results. As it stands, the contribution is not established.

major comments (5)
  1. [III.B, Tables 1 and 2] The central claim that KD "consistently outperformed" DRCA and KD-DRCA is not supported by the reported inferential statistics. Tables 1 and 2 only count paired t-tests of each method against the Baseline; no paired test compares KD with DRCA or KD with KD-DRCA. A method can be significantly above baseline in more cells than another method without being significantly better than that method in any direct comparison. Moreover, even against the baseline, KD is significantly better in only 24 of 72 cells, ties in 44, and is significantly worse in 4; this does not support the word "consistently." The abstract's conclusion that "KD significantly outperformed the previously validated DRCA method" therefore rests on an inference that the tables do not make.
  2. [II.E] Hyperparameters are selected using target-domain labels, which biases the reported test results. The text states that the target batch was randomly partitioned into 50% validation and 50% test data, and that a grid search optimized accuracy on the validation set for temperature T, DRCA dimension d, and regularization alpha. Because the validation split is drawn from the same target distribution as the test split, the test performance is conditional on labels from that distribution. This contradicts the paper's framing of unsupervised drift compensation, in which target labels are unavailable during deployment. It also makes the method comparison potentially unfair if the methods differ in the number or flexibility of tuned hyperparameters. The limitations paragraph in Section IV does not acknowledge this selection bias.
  3. [Abstract and Section III] The abstract's quantitative claims of "up to an 18% improvement in accuracy and 15% in F1-score" appear nowhere in the results section. No table or text reports the actual accuracy, F1-score, precision, or recall values for any method on any task, nor their standard deviations or confidence intervals. Figures 3 and 4 are normalized or qualitative and do not allow the reader to verify the claimed effect sizes. Without numeric results, the headline result is unreproducible and cannot be checked against the stated statistical protocol.
  4. [II.B and II.E] The description of the KD method is incomplete and partly inconsistent. The manuscript does not report the optimizer, learning rate, number of epochs, batch size, teacher and student network training details, or how the baseline FCNN is trained for the cross-domain tasks. In addition, Eq. (3) assigns y_i = f_S(x_i) without the temperature-scaled soft labels used in Eq. (4), and Eq. (4) writes the distillation loss as l(σ(T^{-1}f_S(x_i)), σ(f_T(x_i))) without explaining how the same soft targets are generated consistently in the implementation. These omissions prevent reproduction of the method that is the paper's main contribution.
  5. [II.C, Eq. (7)] The DRCA eigenvalue problem as written is not well posed. The between-domain scatter matrix S_b is defined as a rank-one outer product (µ_S - µ_T)(µ_S - µ_T)', so S_b is singular and S_b^{-1} does not exist. Eq. (7) therefore cannot be a standard eigenvalue decomposition unless some regularization or generalized-eigenvalue formulation is intended. Because DRCA is the benchmark method, the text should specify the exact regularized objective actually optimized; otherwise the reported DRCA results may not correspond to the described method.
minor comments (5)
  1. [Abstract and Section I] The phrase "This work was not founded by any specific research grant" should be "funded."
  2. [III.B and Fig. 4 caption] The text says Fig. 4 shows mean test-set performance, while the caption says the values show median performance over 30 parallel experiments; these should be reconciled.
  3. [II.F and Tables 1-2] No multiple-comparison correction is applied across the 72 paired t-tests per method; at α = 0.05, several spurious significant results are expected by chance. Reporting corrected p-values or paired effect sizes would strengthen the analysis.
  4. [II.E] The 30 random target partitions are not described with seeds or provided as code or indices, so the specific results cannot be reproduced by other researchers.
  5. [II.D] The KD-DRCA hybrid is described only at a high level; it is unclear whether the student model in KD-DRCA is trained on the DRCA-projected source and target data with the same temperature schedule and network architecture as the standalone KD method.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the KD claim rests on an empirical benchmark against an external dataset; the missing pairwise statistical test is a correctness issue, not a circular derivation.

full rationale

The paper's derivation chain is not circular. The KD method is defined by Eqs. (1)-(4): a teacher is trained on source-domain labels, then a student is trained on the teacher's soft labels over pooled source and unlabeled target data. No target-domain test label enters the training objective, and the reported accuracy/F1 values come from held-out halves of the UCI target batches after 30 random partitions. DRCA is cited as an external benchmark from [24] and its projection is defined by the eigenvalue problem in Eq. (7); the paper does not redefine DRCA in terms of KD or vice versa. Hyperparameters (T, d, alpha) are tuned on a 50% validation split of each target batch; this is a potential selection-bias or leakage concern when claiming unsupervised deployment performance, but it is not a case where a fitted parameter is renamed as a prediction or where an output is equivalent to an input by construction. The strongest non-circularity concern is statistical: the abstract claims KD 'consistently outperformed' DRCA and KD-DRCA, but Tables 1 and 2 only report counts of each method versus the baseline, not paired KD-versus-DRCA tests. That is an unsupported inference and an overstatement, not a circular-reasoning defect. Self-citations appear (e.g., [25], [40]), but they are not load-bearing for the central KD claim, which relies on the external UCI benchmark and the external KD reference [27]. Therefore, no circular step is present.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities. The free parameters are the hyperparameters tuned on the target-domain validation split, which also creates a data leakage concern. The key domain assumption is that teacher soft labels on target data are reliable, which is not independently verified.

free parameters (3)
  • Temperature T (KD) = grid search over [0.3, 1, 2, 3, 5, 25, 50, 100, 200]
    Tuned on target-domain validation set; controls smoothness of soft labels and directly affects the distillation loss.
  • Subspace dimensionality d (DRCA) = grid search over [50, 100]
    Dimension of the DRCA projection; selected on validation set and affects how much domain-invariant information is retained.
  • Regularization parameter alpha (DRCA) = grid search over [0.001, 0.01, 0.1, 1, 10, 100, 1000]
    Balances within-domain scatter vs between-domain scatter in the DRCA objective; chosen on validation set.
assumptions (3)
  • domain assumption The teacher model trained on source domain provides soft labels on target data that are informative enough for the student to learn a better classifier.
    Central to the KD mechanism; invoked in Eq. (3)-(4). If the teacher's target predictions are noisy or mis-calibrated, the distillation loss could reinforce errors.
  • domain assumption The UCI Gas Sensor Array Drift Dataset's batch structure is a valid proxy for real-world sensor drift, and its 10 batches represent distinct deployment conditions.
    All conclusions are drawn from this single dataset; the authors themselves note lack of generalizability to other datasets as a limitation.
  • standard math The paired t-test over 30 random partitions is valid for detecting differences between methods.
    Paired t-tests assume the differences are normally distributed; with 30 samples this is usually robust, but the paper does not check this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sensor Drift Compensation in Electronic-Nose-Based Gas Recognition Using Knowledge Distillation." pith.science (2026). https://pith.science/paper/4H3M4FCZ

@misc{pith2026250717071,
  author       = {Pith},
  title        = {Pith review of: Sensor Drift Compensation in Electronic-Nose-Based Gas Recognition Using Knowledge Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4H3M4FCZ}},
  note         = {Machine review of arXiv:2507.17071}
}
read the original abstract

Due to environmental changes and sensor aging, sensor drift challenges the performance of electronic nose systems in gas classification during real-world deployment. Previous studies using the UCI Gas Sensor Array Drift Dataset reported promising drift compensation results but lacked robust statistical experimental validation and may overcompensate for sensor drift, losing class-related variance.To address these limitations and improve sensor drift compensation with statistical rigor, we first designed two domain adaptation tasks based on the same electronic nose dataset: using the first batch to predict the remaining batches, simulating a controlled laboratory setting; and predicting the next batch using all prior batches, simulating continuous training data updates for online training. We then systematically tested three methods: our proposed novel Knowledge Distillation (KD) method, the benchmark method Domain Regularized Component Analysis (DRCA), and a hybrid method KD-DRCA, across 30 random test set partitions on the UCI dataset. We showed that KD consistently outperformed both DRCA and KD-DRCA, achieving up to an 18% improvement in accuracy and 15% in F1-score, demonstrating KD's superior effectiveness in drift compensation. This is the first application of KD for electronic nose drift mitigation, significantly outperforming the previous state-of-the-art DRCA method and enhancing the reliability of sensor drift compensation in real-world environments.

Figures

Figures reproduced from arXiv: 2507.17071 by the authors.

Figure 1
Figure 1. Pipeline of KD and DRCA for drift compensation. (A) The process involves training a teacher model on source domain to generate soft labels, which trains the student model. (B) DRCA transforms the source and target domain data into a new feature subspace. The transformed source data is used to train the model to predict the transformed target domain data. as x T i ∈ R D, where i = 1, 2, ..., N(T), and D denotes the d… view at source ↗
Figure 2
Figure 2. The sensor drift visualization in the feature space of the ten batches of data with tSNE. Classes: 0-Ethanol, 1-Ethylene, 2-Ammonia, 3- Acetaldehyde, 4-Acetone, 5-Toluene applications. Compared to the data-perspective methods like DRCA, KD retains original knowledge while adapting to new target data, potentially preserving more stable X → Y relationship information across domains. In contrast, DRCA, which modifies t… view at source ↗
Figure 3
Figure 3. Classification performance with the baseline, KD, DRCA, KD-DRCA methods on the test sets of different batches of data. Task 1: Train on the first batch and test on the other batches; Task 2: Train on all previous batches and test on the next batch. supervised nature appears to be more adept at handling the nuances of the drift problem [42]. Despite overall good results achieved by KD, no drift compensation method is… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Summary radar plots of the mean classification performance of different methods relative to baseline across four metrics for all tasks on the test sets. The radar plots compare four methods: Baseline, KD, DRCA, and KD-DRCA, across two tasks and nine target batches for …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [1]

    Electronic nose for black tea classification and correla- tion of measurements with “tea taster

    N. Bhattacharyya, R. Bandyopadhyay, M. Bhuyan, B. Tudu, D. Ghosh, and A. Jana, “Electronic nose for black tea classification and correla- tion of measurements with “tea taster” marks,” IEEE transactions on instrumentation and measurement , vol. 57, no. 7, pp. 1313–1321, 2008

  2. [2]

    A review of the applica- tion of non-intrusive infrared sensing for gas–liquid flow characteriza- tion,

    K. Sarkodie, A. Fergusson-Rees, and P. Diaz, “A review of the applica- tion of non-intrusive infrared sensing for gas–liquid flow characteriza- tion,” The Journal of Computational Multiphase Flows , vol. 10, no. 1, pp. 43–56, 2018

  3. [3]

    Discrimination of different species of dendrobium with an electronic nose using aggregated conformal predictor,

    Y . Wang, Z. Wang, J. Diao, X. Sun, Z. Luo, and G. Li, “Discrimination of different species of dendrobium with an electronic nose using aggregated conformal predictor,” Sensors, vol. 19, no. 4, p. 964, 2019

  4. [4]

    Discrimination between alternative herbal medicines from different categories with the electronic nose,

    X. Zhan, X. Guan, R. Wu, Z. Wang, Y . Wang, and G. Li, “Discrimination between alternative herbal medicines from different categories with the electronic nose,” Sensors, vol. 18, no. 9, p. 2936, 2018

  5. [5]

    An electronic nose-based assistive diagnostic prototype for lung cancer detection with conformal prediction,

    X. Zhan, Z. Wang, M. Yang, Z. Luo, Y . Wang, and G. Li, “An electronic nose-based assistive diagnostic prototype for lung cancer detection with conformal prediction,” Measurement, vol. 158, p. 107588, 2020

  6. [6]

    Application of a low-cost electronic nose for differentiation between pathogenic oomycetes pythium intermedium and phytophthora plurivora,

    P. Borowik, L. Adamowicz, R. Tarakowski, P. Wacławik, T. Oszako, S. ´Slusarski, and M. Tkaczyk, “Application of a low-cost electronic nose for differentiation between pathogenic oomycetes pythium intermedium and phytophthora plurivora,” Sensors, vol. 21, no. 4, p. 1326, 2021

  7. [7]

    A low-cost interface to high- value resistive sensors varying over a wide range,

    A. Flammini, D. Marioli, and A. Taroni, “A low-cost interface to high- value resistive sensors varying over a wide range,” IEEE Transactions on Instrumentation and Measurement , vol. 53, no. 4, pp. 1052–1056, 2004

  8. [8]

    Developments and applications of electronic nose systems for gas mixtures classification and con- centration estimation,

    C. Pace, L. Fragomeni, and W. Khalaf, “Developments and applications of electronic nose systems for gas mixtures classification and con- centration estimation,” Applications in Electronics Pervading Industry, Environment and Society: APPLEPIES 2014 , pp. 1–7, 2016

Show all 43 references
  1. [9]

    Co, no2 and nox 9 urban pollution monitoring with on-field calibrated electronic nose by automatic bayesian regularization,

    S. De Vito, M. Piga, L. Martinotto, and G. Di Francia, “Co, no2 and nox 9 urban pollution monitoring with on-field calibrated electronic nose by automatic bayesian regularization,” Sensors and Actuators B: Chemical , vol. 143, no. 1, pp. 182–191, 2009

  2. [10]

    Assessment of meat freshness with metal oxide sensor microarray electronic nose: A practical approach,

    V . Y . Musatov, V . Sysoev, M. Sommer, and I. Kiselev, “Assessment of meat freshness with metal oxide sensor microarray electronic nose: A practical approach,” Sensors and Actuators B: Chemical , vol. 144, no. 1, pp. 99–103, 2010

  3. [11]

    Portable electronic nose based on electrochemical sensors for food quality assessment,

    W. Wojnowski, T. Majchrzak, T. Dymerski, J. Gebicki, and J. Namie´snik, “Portable electronic nose based on electrochemical sensors for food quality assessment,” Sensors, vol. 17, no. 12, p. 2715, 2017

  4. [12]

    Electronic noses for food quality: A review,

    A. Loutfi, S. Coradeschi, G. K. Mani, P. Shankar, and J. B. B. Rayappan, “Electronic noses for food quality: A review,” Journal of F ood Engineering, vol. 144, pp. 103–111, 2015

  5. [13]

    Electronic nose for quality control of colombian coffee through the detection of defects in “cup tests

    J. Rodr ´ıguez, C. Dur ´an, and A. Reyes, “Electronic nose for quality control of colombian coffee through the detection of defects in “cup tests”,” Sensors, vol. 10, no. 1, pp. 36–46, 2009

  6. [14]

    Comparison and data fusion of an electronic nose and near-infrared reflectance spectroscopy for the discrimination of ginsengs,

    J. Miao, Z. Luo, Y . Wang, and G. Li, “Comparison and data fusion of an electronic nose and near-infrared reflectance spectroscopy for the discrimination of ginsengs,” Analytical Methods, vol. 8, no. 6, pp. 1265– 1273, 2016

  7. [15]

    Feature engineering in discrimination of herbal medicines from different geo- graphical origins with electronic nose,

    X. Zhan, X. Guan, R. Wu, Z. Wang, Y . Wang, and G. Li, “Feature engineering in discrimination of herbal medicines from different geo- graphical origins with electronic nose,” in 2019 IEEE 7th International Conference on Bioinformatics and Computational Biology (ICBCB) . IEEE, ...

  8. [16]

    Analysis of breath samples for lung cancer survival,

    B. Schmekel, F. Winquist, and A. Vikstr ¨om, “Analysis of breath samples for lung cancer survival,” Analytica chimica acta , vol. 840, pp. 82–86, 2014

  9. [17]

    On the calibration of sensor arrays for pattern recognition using the minimal number of experiments,

    I. Rodriguez-Lujan, J. Fonollosa, A. Vergara, M. Homer, and R. Huerta, “On the calibration of sensor arrays for pattern recognition using the minimal number of experiments,” Chemometrics and Intelligent Laboratory Systems , vol. 130, pp. 123–134, 2014

  10. [18]

    Neural, fuzzy and neuro-fuzzy approach for concentration estimation of volatile organic compounds by surface acoustic wave sensor array,

    S. K. Jha, K. Hayashi, and R. Yadava, “Neural, fuzzy and neuro-fuzzy approach for concentration estimation of volatile organic compounds by surface acoustic wave sensor array,” Measurement, vol. 55, pp. 186–195, 2014

  11. [19]

    Online conformal prediction for classifying different types of herbal medicines with electronic nose,

    X. Zhan, X. Guan, R. Wu, Z. Wang, Y . Wang, Z. Luo, and G. Li, “Online conformal prediction for classifying different types of herbal medicines with electronic nose,” 2018

  12. [20]

    Recognition of coffee using differential electronic nose,

    K. Brudzewski, S. Osowski, and A. Dwulit, “Recognition of coffee using differential electronic nose,” IEEE Transactions on Instrumentation and Measurement, vol. 61, no. 6, pp. 1803–1810, 2012

  13. [21]

    Compensation for the drift- like terms caused by environmental fluctuations in the responses of chemoresistive gas sensors,

    F. Hossein-Babaei and V . Ghafarinia, “Compensation for the drift- like terms caused by environmental fluctuations in the responses of chemoresistive gas sensors,” Sensors and Actuators B: Chemical , vol. 143, no. 2, pp. 641–648, 2010

  14. [22]

    Chaotic time series prediction of e-nose sensor drift in embedded phase space,

    L. Zhang, F. Tian, S. Liu, L. Dang, X. Peng, and X. Yin, “Chaotic time series prediction of e-nose sensor drift in embedded phase space,” Sensors and Actuators B: Chemical , vol. 182, pp. 71–79, 2013

  15. [23]

    Brain deformation estimation with transfer learn- ing for head impact datasets across impact types,

    X. Zhan, Y . Liu, N. J. Cecchi, O. Gevaert, M. M. Zeineh, G. A. Grant, and D. B. Camarillo, “Brain deformation estimation with transfer learn- ing for head impact datasets across impact types,” IEEE Transactions on biomedical Engineering , vol. 71, no. 6, pp. 1853–1863, 2024

  16. [24]

    Anti-drift in e-nose: A subspace projection approach with drift reduction,

    L. Zhang, Y . Liu, Z. He, J. Liu, P. Deng, and X. Zhou, “Anti-drift in e-nose: A subspace projection approach with drift reduction,” Sensors and Actuators B: Chemical , vol. 253, pp. 407–417, 2017

  17. [25]

    Unsupervised cross-user adaptation in taste sensation recognition based on surface electromyography,

    H. Wang, X. Zhan, L. Liu, A. Ullah, H. Li, H. Gao, Y . Wang, R. Hu, and G. Li, “Unsupervised cross-user adaptation in taste sensation recognition based on surface electromyography,” IEEE Transactions on Instrumentation and Measurement , vol. 71, pp. 1–11, 2022

  18. [26]

    Adaptive machine learning head model across different head impact types using unsupervised domain adaptation and generative adversarial networks,

    X. Zhan, J. Sun, Y . Liu, N. J. Cecchi, E. Le Flao, O. Gevaert, M. M. Zeineh, and D. B. Camarillo, “Adaptive machine learning head model across different head impact types using unsupervised domain adaptation and generative adversarial networks,” IEEE Sensors Journal , 2024

  19. [27]

    Knowledge distillation for semi- supervised domain adaptation,

    M. Orbes-Arteainst, J. Cardoso, L. Sørensen, C. Igel, S. Ourselin, M. Modat, M. Nielsen, and A. Pai, “Knowledge distillation for semi- supervised domain adaptation,” in OR 2.0 Context-Aware Operating Theaters and Machine Learning in Clinical Neuroimaging: Second Inter- nationa...

  20. [28]

    Boost ai power: Data augmentation strategies with unlabeled data and conformal prediction, a case in alternative herbal medicine discrimination with electronic nose,

    L. Liu, X. Zhan, R. Wu, X. Guan, Z. Wang, W. Zhang, M. Pilanci, Y . Wang, Z. Luo, and G. Li, “Boost ai power: Data augmentation strategies with unlabeled data and conformal prediction, a case in alternative herbal medicine discrimination with electronic nose,” IEEE Sensors Jou...

  21. [29]

    L. Liu, X. Zhan, X. Yang, X. Guan, R. Wu, Z. Wang, Z. Luo, Y . Wang, and G. Li, “Cpsc: Conformal prediction with shrunken centroids for efficient prediction reliability quantification and data augmentation, a case in alternative herbal medicine classification with electronic n...

  22. [30]

    Reliability-enhanced data cleaning in biomedical machine learning using inductive conformal prediction,

    X. Zhan, Q. Xu, Y . Zheng, G. Lu, and O. Gevaert, “Reliability-enhanced data cleaning in biomedical machine learning using inductive conformal prediction,” PLOS Computational Biology , vol. 21, no. 2, p. e1012803, 2025

  23. [31]

    Gas Sensor Array Drift Dataset,

    A. Vergara, “Gas Sensor Array Drift Dataset,” UCI Machine Learning Repository, 2012, DOI: https://doi.org/10.24432/C5RP6W

  24. [32]

    Chemical gas sensor drift compensation using classifier ensembles,

    A. Vergara, S. Vembu, T. Ayhan, M. A. Ryan, M. L. Homer, and R. Huerta, “Chemical gas sensor drift compensation using classifier ensembles,” Sensors and Actuators B: Chemical , vol. 166, pp. 320–329, 2012

  25. [33]

    A study of drift effect in a popular metal oxide sensor and gas recognition using public gas datasets,

    I.-S. Chang, S.-W. Byun, T.-B. Lim, and G.-M. Park, “A study of drift effect in a popular metal oxide sensor and gas recognition using public gas datasets,” IEEE Access , vol. 11, pp. 26 383–26 392, 2023

  26. [34]

    Drift in a popular metal oxide sensor dataset reveals limitations for gas classification benchmarks,

    N. Dennler, S. Rastogi, J. Fonollosa, A. Van Schaik, and M. Schmuker, “Drift in a popular metal oxide sensor dataset reveals limitations for gas classification benchmarks,” Sensors and Actuators B: Chemical , vol. 361, p. 131668, 2022

  27. [35]

    Multi-classifier tree with transient features for drift compensation in electronic nose,

    A. U. Rehman, S. B. Belhaouari, M. Ijaz, A. Bermak, and M. Hamdi, “Multi-classifier tree with transient features for drift compensation in electronic nose,” IEEE Sensors Journal , vol. 21, no. 5, pp. 6564–6574, 2020

  28. [36]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

  29. [37]

    Knowledge distillation in deep learning and its applications,

    A. Alkhulaifi, F. Alsahli, and I. Ahmad, “Knowledge distillation in deep learning and its applications,” PeerJ Computer Science , vol. 7, p. e474, 2021

  30. [38]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,” International Journal of Computer Vision , vol. 129, no. 6, pp. 1789–1819, 2021

  31. [39]

    Simaug: Learning robust representations from 3d simulation for pedestrian trajectory prediction in unseen cameras,

    J. Liang, L. Jiang, and A. Hauptmann, “Simaug: Learning robust representations from 3d simulation for pedestrian trajectory prediction in unseen cameras,” arXiv preprint arXiv:2004.02022 , vol. 2, 2020

  32. [40]

    Unsupervised cross-user adaptation in taste sensation recognition based on surface electromyography with conformal prediction and domain reg- ularized component analysis,

    H. Wang, X. Zhan, L. Liu, A. Ullah, H. Li, H. Gao, Y . Wang, and G. Li, “Unsupervised cross-user adaptation in taste sensation recognition based on surface electromyography with conformal prediction and domain reg- ularized component analysis,” arXiv preprint arXiv:2110.11339 , 2021

  33. [41]

    Modeling teacher-student techniques in deep neural networks for knowledge distillation,

    S. Abbasi, M. Hajabdollahi, N. Karimi, and S. Samavi, “Modeling teacher-student techniques in deep neural networks for knowledge distillation,” in 2020 International Conference on Machine Vision and Image Processing (MVIP) . IEEE, 2020, pp. 1–6

  34. [42]

    A survey on semi-supervised learning for delayed partially labelled data streams,

    H. M. Gomes, M. Grzenda, R. Mello, J. Read, M. H. Le Nguyen, and A. Bifet, “A survey on semi-supervised learning for delayed partially labelled data streams,” ACM Computing Surveys , vol. 55, no. 4, pp. 1–42, 2022

  35. [43]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2223–2232

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.