Pith. sign in

REVIEW 3 major objections 5 minor 24 references

UniTac-NV: A Unified Tactile Representation For Non-Vision-Based Tactile Sensors

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that joint autoencoder training on matched tactile contacts creates a shared latent representation across non-vision-based tactile sensors, enabling cross-sensor data transfer and downstream contact geometry estimation…

desk verdict A credible proof-of-concept for cross-sensor transfer in non-vision tactile sensing, honest about its limits, but the matched-contact premise needs closer validation. read the letter →

arxiv 2506.19699 v1 pith:FCYKCC7Q submitted 2025-06-24 cs.RO

Jian Hou , Xin Zhou , Qihan Yang , Adam J. Spiers This is my paper · ORCID
classification cs.RO
keywords tactilesensingcross-sensortransferlatentspacealignmentautoencodernon-vision-basedsensorscontactgeometryestimationsample-matchedtrainingrepresentationlearning
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

UniTac-NV claims that two very different non-vision tactile sensors can be made to share one latent representation of touch. The paper builds an autoencoder with one encoder per sensor and a single shared decoder, and trains it on matched contacts—same object, pressing angle, and force recorded by both sensors. After training, a raw reading from either sensor can be encoded into the common space and decoded into the other sensor's format, and a downstream contact-geometry model trained on one sensor's latent codes works on the other's without retraining. If true, this addresses a practical gap: algorithms built for one tactile sensor would not become obsolete with that sensor, and data collected across sensors could be pooled.

What carries the argument

The central mechanism is the sample-matched multi-encoder autoencoder with a shared decoder: two sensor-specific encoders map raw readings to a common 16-dimensional latent vector, and a single decoder reconstructs both sensors' outputs from any latent code. The loss is the sum of self- and cross-reconstruction mean absolute errors over paired samples, which implicitly enforces latent alignment because the decoder must produce consistent reconstructions from both encoders. The named architecture is UniTac-NV, and the training data are force-controlled presses against matched objects at matched angles and force levels.

What would settle it

Press the two sensors, with a known small offset between their contact-surface centers, against a flat plate with matched force and angle, and measure cross-reconstruction error. If the error rises sharply as the offset increases, the latent alignment is a property of the matched dataset, not of the physical contact, and the claimed sensor-agnostic space would not survive realistic mounting variation.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that implicit alignment is enough: no contrastive or alignment loss is needed, only joint reconstruction through a shared decoder on sample-matched inputs. The shared 16-dimensional latent space becomes sensor-agnostic because the decoder must reconstruct both the 4×6×3 uSkin readings and the 3×3×3 PapillArray readings from either encoder's code. On six seen objects, cross-sensor translations keep structural similarity above 0.95. On an unseen irregular object, same-sensor contact geometry estimation errors are roughly 0.35–0.40 mm while cross-sensor errors rise to 0.64–0.67 mm, which the paper attributes to sensor morphology differences and interprets as only shared information transferring.

Load-bearing premise

The load-bearing assumption is that the two sensors truly feel the same contacts—same object, same angle, same force, and same contact-surface center—so that paired training aligns physical contact properties rather than dataset-specific coincidences.

Editorial extensions

If this is right

  • A downstream contact-geometry estimator trained on latent codes from one sensor can be applied to the other sensor's latent codes with no retraining; the paper reports an average estimation error of 0.513 mm across all scenarios.
  • Raw tactile readings can be translated between sensor formats: on the six seen objects, cross-sensor translations keep structural similarity above 0.95, with larger errors when upsampling the smaller PapillArray grid to the uSkin grid.
  • Because the latent space separates object material and shape even though no classification loss is used, the representation preserves physical contact information beyond what is needed for reconstruction.
  • The approach extends to unseen objects only partially: same-sensor reconstructions remain near individually trained baselines on the irregular test object, but the PapillArray-to-uSkin transfer degrades sharply (SSIM 0.581), attributed to sensor morphology and limited shared information.
  • The framework requires matched contacts only for training; once trained, single-encoder inference lets either sensor's raw data enter the shared latent space directly.

Reading between the lines

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

  • If the implicit alignment generalizes beyond the two tested sensors, adding a new sensor would only require a new encoder and matched-contact data; the shared decoder could then translate between all pairs without pairwise training.
  • The matched-contact collection procedure is the real bottleneck for deployment: real robot hands cannot usually reproduce identical contact poses across different sensors, so a practical version would need relaxed alignment such as simulated contacts, force-conditioned alignment, or self-supervised pairing.
  • The reported error jump from roughly 0.35–0.40 mm to 0.64–0.67 mm on cross-sensor geometry estimation suggests an information-theoretic bound: transfer quality should track the spatial resolution and taxel count shared by the two sensors, which could be tested by adding a third sensor with intermediate resolution.
  • A strong test of the 'common language' claim would be compositional transfer: train UniTac-NV on pairs (A,B) and (B,C) separately, then check whether encoder A's codes decode correctly through the B–C decoder into C's format without ever seeing A–C pairs.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes UniTac-NV, an encoder-decoder architecture with two sensor-specific encoders and a single shared decoder, trained on paired tactile samples from two non-vision-based sensors (Xela uSkin and Contactile PapillArray) collected under a designed pressing protocol. The central claim is that joint training of this autoencoder on matched physical contacts yields a sensor-agnostic latent space that supports (i) cross-sensor reconstruction/translation of raw tactile readings and (ii) downstream reuse of a contact-geometry estimator trained on one sensor's latent codes and applied to the other. Experiments on six seen objects (three shapes × two materials) show strong self- and cross-reconstruction (average SSIM > 0.95), while results on an unseen irregular object degrade substantially, especially for PapillArray-to-uSkin transfer (NMAE 0.318, SSIM 0.581). A downstream contact geometry estimation task on the unseen object demonstrates that cross-sensor latent transfer is possible but with increased estimation error (from ~0.35-0.40 mm to ~0.64-0.67 mm).

Significance. If the matched-contact premise holds, this is a meaningful step toward cross-sensor generalization for non-optical tactile sensors. The architecture is simple, uses only reconstruction losses without explicit alignment terms, and includes a reproducible data collection procedure. The paper also benchmarks against individually trained autoencoders, giving a useful near-upper bound. The main significance is as a proof of concept: it shows that implicit latent alignment can be induced from paired data and that the resulting representation carries enough information for a downstream task. However, the strength of the conclusion is limited by the lack of direct evidence for physical contact equivalence and by the large performance drop on the unseen object. The authors themselves acknowledge that only information common between sensors can be transferred, which appropriately frames the contribution.

major comments (3)
  1. [Sec. II-D and III-A] The 'matched contacts' are matched only in commanded task parameters (same object, approach angle, and tooltip z-force thresholds), not in terms of measured physical contact equivalence. The two sensors have different surface geometries (flat elastic sheet over 4×6 taxels versus exposed nubs with a raised center on a 3×3 grid) and different compliance, so at the same commanded 10 N normal force the actual pressure distributions, contact areas, and deformation histories are likely to differ. This is a load-bearing assumption: if the paired samples are not physically equivalent, the shared decoder may align latent codes according to the experimental protocol (object, angle, force) rather than to a common physical contact state. The authors should provide evidence of contact equivalence (e.g., comparison of measured force distributions or repeatability tests) or, failing that, explicitly discuss how violations of this premise would affect the interpretation of the latent space. The poor transfer on the unseen irregular object (Table I, PapillArray→uSkin SSIM 0.581) is consistent with this concern.
  2. [Sec. III-C and Fig. 5] The quantitative evidence for latent space alignment is limited. The text mentions a Manhattan distance metric and shows a decreasing curve in Fig. 5(A) and t-SNE plots in Fig. 5(B), but no numeric values for the alignment distance are reported in the text. Given that the paper's central claim is that the latent space is 'sensor-agnostic,' the authors should report the actual mean and standard deviation of the Manhattan distances between matched latent codes, and ideally add a quantitative analysis (e.g., linear probing or a classification/regression task on the latent codes) to show that the latent space encodes physical contact properties rather than merely the experimental labels. Without such evidence, the t-SNE visualization alone does not fully support the sensor-agnostic claim.
  3. [Sec. IV-C and Abstract] The abstract's claim of 'cross-sensor data transfer with low errors' is too strong in light of Table I, where the unseen irregular object yields SSIM 0.581 for PapillArray-to-uSkin transfer, and the downstream cross-sensor errors increase by nearly a factor of two compared to same-sensor evaluation (0.637/0.666 mm vs 0.353/0.397 mm). The authors should qualify the 'low errors' claim to the seen-object conditions or to cases where sensor morphologies are more similar, and should provide error bars or confidence intervals for the downstream results, which are reported as point estimates only.
minor comments (5)
  1. [Sec. III-B] The normalized mean absolute error (NMAE) is not defined precisely; please specify the normalization factor (e.g., range of sensor readings or maximum absolute value).
  2. [Table I] Consider reporting standard deviations or confidence intervals for the NMAE and SSIM values, especially for the row 'Irregular (Unseen)', where the spread is likely large.
  3. [Fig. 5] In Fig. 5(A), the 'latent space alignment error' is not defined in the caption; please state explicitly that it is the Manhattan distance in the 16-dimensional latent space, and add axis labels and units.
  4. [Sec. IV-A] The split of the unseen-object data into 920 test and 8,280 training samples sums to 9,200, but the preceding description suggests 4 rotations × 91 presses × 25 force levels = 9,100 samples; please reconcile these numbers.
  5. [Sec. V] The conclusion reports an 'average estimation error of 0.513 mm' but does not list the four individual errors (0.353, 0.397, 0.637, 0.666 mm); please state the individual values for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the cross-sensor transfer and downstream geometry estimation are evaluated on held-out and unseen data, and the latent alignment is an explicit training objective rather than an assumed or self-referential result.

full rationale

The paper's central claim is that joint training of sensor-specific encoders with a shared decoder on matched contacts creates a sensor-agnostic latent space. This is not circular. The latent alignment is induced by an explicit training objective, the sum of self- and cross-reconstruction errors in Section II-A.2, and the paper explicitly notes that "neither latent alignment nor separation are utilized as training losses." The claim is then tested on data that were not used to fit the model: a stratified held-out test set containing 10% of approaching angles for each seen object plus all data from the irregular unseen object (Section III-A), and a downstream contact-geometry MLP trained on one sensor's latent codes and evaluated on the other sensor's latent codes (Section IV). This downstream transfer experiment is an external validation, not a renaming of the training objective. The matched-contact data collection procedure is an empirical assumption about physical equivalence, not a circular definition: even if the contacts were not physically matched, that would be a data-quality limitation, and the paper itself acknowledges the consequence in the poor PapillArray-to-uSkin transfer on the irregular object (NMAE 0.318, SSIM 0.581 in Table I) and in the statement that "only information common between sensors can be transferred." The references to the authors' own prior work ([3], [5], [7]) are contextual background on tactile sensing and are not load-bearing for the proposed architecture or its evaluation. There is no self-citation chain, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. The paper's limitations about dataset scale and sensor diversity are honest boundary conditions, not evidence of circular reasoning.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

No mathematical derivation is attempted; the paper is an empirical system. The parameters above are the hand-chosen design choices and learned weights that the results depend on. The axioms are the domain assumptions that the experimental setup and architecture rely on.

free parameters (5)
  • Encoder/decoder network weights = Learned from data, exact values not reported
    The central claim of latent alignment depends on these trained weights, but this is standard supervised learning rather than a physical free parameter.
  • Latent space dimension = 16
    Hand-chosen capacity with no ablation; determines how much contact information is preserved and affects cross-sensor transfer fidelity.
  • Autoencoder architecture sizes = Encoders 64/48, decoder 64/96
    Hand-chosen layer widths, no sensitivity analysis reported.
  • Training hyperparameters for autoencoder = 1000 epochs, lr 5e-4, dropout 0.007, batch 64
    Chosen without reported tuning; these affect convergence and the degree of latent alignment.
  • Downstream MLP hyperparameters = Dropout 0.2 (uSkin) and 0.3 (PapillArray), lr 5e-5, weight decay 1e-3, 80 epochs
    Hand-chosen for the geometry estimator; no sensitivity analysis, and the differing dropout values are set per sensor.
assumptions (4)
  • domain assumption Matched contact interactions can be accurately reproduced across sensors via the robot procedure.
    Sec. II-D pairs data by object, approach angle, force threshold, and closest time frame; if matching is inaccurate, the alignment target is meaningless.
  • domain assumption A shared 16-dimensional latent space can capture transferable contact properties across the two sensors.
    This is central to the architecture; there is no theoretical guarantee, only empirical evidence.
  • domain assumption The two sensors' contact surfaces are in the same pose relative to the tool flange.
    Sec. II-B states custom adapters ensure this alignment, but it is not directly measured or verified after mounting.
  • domain assumption The data splitting prevents leakage between training and test sets.
    Sec. III-A describes a stratified angle-based split; the assumption is that this prevents nearly identical samples from appearing in both sets.
invented entities (1)
  • UniTac-NV shared latent space
    purpose: Sensor-agnostic representation for non-vision tactile data
    The latent space is an algorithmic construct, not a physical entity. Its existence is evidenced only through the paper's own reconstruction and downstream experiments; there is no independent dataset or external benchmark verifying it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniTac-NV: A Unified Tactile Representation For Non-Vision-Based Tactile Sensors." pith.science (2026). https://pith.science/paper/FCYKCC7Q

@misc{pith2026250619699,
  author       = {Pith},
  title        = {Pith review of: UniTac-NV: A Unified Tactile Representation For Non-Vision-Based Tactile Sensors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCYKCC7Q}},
  note         = {Machine review of arXiv:2506.19699}
}
read the original abstract

Generalizable algorithms for tactile sensing remain underexplored, primarily due to the diversity of sensor modalities. Recently, many methods for cross-sensor transfer between optical (vision-based) tactile sensors have been investigated, yet little work focus on non-optical tactile sensors. To address this gap, we propose an encoder-decoder architecture to unify tactile data across non-vision-based sensors. By leveraging sensor-specific encoders, the framework creates a latent space that is sensor-agnostic, enabling cross-sensor data transfer with low errors and direct use in downstream applications. We leverage this network to unify tactile data from two commercial tactile sensors: the Xela uSkin uSPa 46 and the Contactile PapillArray. Both were mounted on a UR5e robotic arm, performing force-controlled pressing sequences against distinct object shapes (circular, square, and hexagonal prisms) and two materials (rigid PLA and flexible TPU). Another more complex unseen object was also included to investigate the model's generalization capabilities. We show that alignment in latent space can be implicitly learned from joint autoencoder training with matching contacts collected via different sensors. We further demonstrate the practical utility of our approach through contact geometry estimation, where downstream models trained on one sensor's latent representation can be directly applied to another without retraining.

Figures

Figures reproduced from arXiv: 2506.19699 by the authors.

Figure 1
Figure 1. The UniTac-NV autoencoder architecture, which unifies tactile data [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Data collection procedure. Step 1: Mount the object on the optical [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Hardware setup for tactile contact data collection. A: The UR5e [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of the data splitting strategy. Matched between sensors, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: A: Testing samples’ reconstruction errors (Normalized MAEs) and [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Examples of tactile data reconstructions for both seen and unseen objects. Each cell represents a sensing unit, with color indicating the normal force [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: Evaluation of contact shape (local geometry) estimation. Top: [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 14 canonical work pages

  1. [1]

    Rotating without seeing: Towards in-hand dexterity through touch,

    Z.-H. Yin, B. Huang, Y . Qin, Q. Chen, and X. Wang, “Rotating without seeing: Towards in-hand dexterity through touch,”arXiv preprint arXiv:2303.10880, 2023

    arXiv 2023

  2. [2]

    Design of a multimodal fingertip sensor for dynamic manipulation,

    A. SaLoutos, E. Stanger-Jones, M. Guo, H. Kim, and S. Kim, “Design of a multimodal fingertip sensor for dynamic manipulation,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 8017–8024

    work page 2023

  3. [3]

    Location and orientation super- resolution sensing with a cost-efficient and repairable barometric tactile sensor,

    J. Hou, X. Zhou, and A. J. Spiers, “Location and orientation super- resolution sensing with a cost-efficient and repairable barometric tactile sensor,”IEEE Transactions on Robotics, 2024

    work page 2024

  4. [4]

    Object recognition for robotics from tactile time series data utilising different neural network architectures,

    W. Bottcher, P. Machado, N. Lama, and T. M. McGinnity, “Object recognition for robotics from tactile time series data utilising different neural network architectures,” in2021 International Joint Conference on Neural Networks (IJCNN). IEEE, 2021, pp. 1–8

    work page 2021

  5. [5]

    Tactile identification of object shapes via in- hand manipulation with a minimalistic barometric tactile sensor array,

    X. Zhou and A. J. Spiers, “Tactile identification of object shapes via in- hand manipulation with a minimalistic barometric tactile sensor array,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 12 590–12 596

    work page 2023

  6. [6]

    Single- grasp object classification and feature extraction with simple robot hands and tactile sensors,

    A. J. Spiers, M. V . Liarokapis, B. Calli, and A. M. Dollar, “Single- grasp object classification and feature extraction with simple robot hands and tactile sensors,”IEEE transactions on haptics, vol. 9, no. 2, pp. 207–220, 2016

    work page 2016

  7. [7]

    E-troll: Tactile sensing and classification via a simple robotic gripper for extended rolling manipulations,

    X. Zhou and A. J. Spiers, “E-troll: Tactile sensing and classification via a simple robotic gripper for extended rolling manipulations,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 11 826–11 833

    work page 2022

  8. [8]

    To- wards robust autonomous grasping with reflexes using high-bandwidth sensing and actuation,

    A. SaLoutos, H. Kim, E. Stanger-Jones, M. Guo, and S. Kim, “To- wards robust autonomous grasping with reflexes using high-bandwidth sensing and actuation,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 10 254–10 260

    work page 2023

Show all 24 references
  1. [9]

    Learning to detect slip with barometric tactile sensors and a temporal convolutional neural network,

    A. Grover, P. Nadeau, C. Grebe, and J. Kelly, “Learning to detect slip with barometric tactile sensors and a temporal convolutional neural network,” in2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 570–576

    2022

  2. [10]

    Papillarray: An incipient slip sensor for dexterous robotic or prosthetic manipulation–design and prototype validation,

    H. Khamis, R. I. Albero, M. Salerno, A. S. Idil, A. Loizou, and S. J. Redmond, “Papillarray: An incipient slip sensor for dexterous robotic or prosthetic manipulation–design and prototype validation,”Sensors and Actuators A: Physical, vol. 270, pp. 195–204, 2018

    2018

  3. [11]

    Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation,

    M. Lambeta, P.-W. Chou, S. Tian, B. Yang, B. Maloon, V . R. Most, D. Stroud, R. Santos, A. Byagowi, G. Kammereret al., “Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation,”IEEE Robotics and Automation Letters, v...

    2020

  4. [12]

    Sensing tactile microvibrations with the biotac—comparison with human sensitivity,

    J. A. Fishel and G. E. Loeb, “Sensing tactile microvibrations with the biotac—comparison with human sensitivity,” in2012 4th IEEE RAS & EMBS international conference on biomedical robotics and biomechatronics (BioRob). IEEE, 2012, pp. 1122–1127

    2012

  5. [13]

    Takktile sensors,

    Soft Robotics Toolkit, “Takktile sensors,” accessed: Feb 22, 2025. [Online]. Available: https://softroboticstoolkit.com/book/takktile- sensors

    2025

  6. [14]

    Gelsight: High-resolution robot tactile sensors for estimating geometry and force,

    W. Yuan, S. Dong, and E. H. Adelson, “Gelsight: High-resolution robot tactile sensors for estimating geometry and force,”Sensors, vol. 17, no. 12, p. 2762, 2017

    2017

  7. [15]

    The tactip family: Soft optical tactile sensors with 3d-printed biomimetic morphologies,

    B. Ward-Cherrier, N. Pestell, L. Cramphorn, B. Winstone, M. E. Giannaccini, J. Rossiter, and N. F. Lepora, “The tactip family: Soft optical tactile sensors with 3d-printed biomimetic morphologies,”Soft robotics, vol. 5, no. 2, pp. 216–227, 2018

    2018

  8. [16]

    High-speed tactile braille reading via biomimetic sliding interactions,

    P. Potdar, D. Hardman, E. Almanzor, and F. Iida, “High-speed tactile braille reading via biomimetic sliding interactions,”IEEE Robotics and Automation Letters, 2024

    2024

  9. [17]

    Sparsh: Self-supervised touch representations for vision-based tactile sensing,

    C. Higuera, A. Sharma, C. K. Bodduluri, T. Fan, P. Lancaster, M. Kalakrishnan, M. Kaess, B. Boots, M. Lambeta, T. Wuet al., “Sparsh: Self-supervised touch representations for vision-based tactile sensing,”arXiv preprint arXiv:2410.24090, 2024

    arXiv

  10. [18]

    Contrastive touch-to-touch pretraining,

    S. Rodriguez, Y . Dou, W. v. d. Bogert, M. Oller, K. So, A. Owens, and N. Fazeli, “Contrastive touch-to-touch pretraining,”arXiv preprint arXiv:2410.11834, 2024

    2024 arXiv

  11. [19]

    Transferable tactile transformers for representation learning across diverse sensors and tasks,

    J. Zhao, Y . Ma, L. Wang, and E. H. Adelson, “Transferable tactile transformers for representation learning across diverse sensors and tasks,”arXiv preprint arXiv:2406.13640, 2024

    2024 arXiv

  12. [20]

    On the design and development of vision-based tactile sensors,

    U. H. Shah, R. Muthusamy, D. Gan, Y . Zweiri, and L. Seneviratne, “On the design and development of vision-based tactile sensors,”Journal of Intelligent & Robotic Systems, vol. 102, pp. 1–27, 2021

    2021

  13. [21]

    Interaction force estimation for tactile sensor arrays: Toward tactile-based interaction control for robotic fingers,

    E. Chelly, A. Cherubini, P. Fraisse, F. B. Amar, and M. Kho- ramshahi, “Interaction force estimation for tactile sensor arrays: Toward tactile-based interaction control for robotic fingers,”arXiv preprint arXiv:2411.13335, 2024

    2024 arXiv

  14. [22]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,”arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  15. [23]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  16. [24]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research, vol. 9, no. 11, 2008

    2008

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.