Explainable Deep Anomaly Detection with Sequential Hypothesis Testing for Robotic Sewer Inspection

Alex George; Lyudmila Mihaylova; Sean R. Anderson; Simon Tait; Will Shepherd

REVIEW 4 major objections 4 minor 15 references

Explainable Deep Anomaly Detection with Sequential Hypothesis Testing for Robotic Sewer Inspection

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A two-stage system that combines explainable deep anomaly detection with sequential hypothesis testing can localise sewer faults in CCTV frames while cutting false alarms through temporal evidence aggregation.

desk verdict Competent application of FCDD+SPRT to sewer inspection, but the headline SPRT gain rests on a non-comparable frame set and needs a fair evaluation before the claim lands. read the letter →

arxiv 2507.22546 v1 pith:JUNNW5D5 submitted 2025-07-30 cs.RO

Alex George , Will Shepherd , Simon Tait , Lyudmila Mihaylova , Sean R. Anderson This is my paper

classification cs.RO

keywords explainableanomalydetectionFCDDsequentialprobabilityratiotesttemporalevidenceaggregationsewerinspectionroboticCCTVone-classdeeplearningdeposit

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

This paper aims to make automated sewer inspection more accurate and easier for a human to trust by combining two established components: Fully Convolutional Data Description (FCDD), a one-class deep anomaly detector that produces a per-image heatmap localising the fault, and the Sequential Probability Ratio Test (SPRT), which accumulates the detector's per-frame scores until enough evidence has built up to declare a normal or anomalous stretch of video. The authors report that on a robotic sewer-inspection video dataset, the combined system raises the F1 score from 84.07% (single-frame thresholding) to 94.36% and cuts the false-positive rate from 11.30% to 0.96%, while still showing where the anomaly is. On a large single-image dataset, the FCDD detector alone also outperformed support-vector-machine baselines (F1 88.03% versus 44.94% and 71.72%). The point of the combination is that frame-level noise, such as motion blur and lighting changes, is suppressed by requiring consistent evidence across time rather than a single threshold crossing.

What carries the argument

Two mechanisms carry the argument. The first is FCDD: a fully convolutional network whose heatmap $A(X)=\sqrt{\phi(X;W)^2+1}-1$ is upsampled to the input resolution, with the scalar anomaly score $z=\frac{1}{uv}\|A\|_1$; it supplies spatial localisation and a per-frame score without pixel-level supervision. The second is SPRT: a sequential hypothesis test that accumulates $\Lambda_t=\Lambda_{t-1}+\log\frac{p(z_t|H_1)}{p(z_t|H_0)}$ across frames, with $p(\cdot|H_0)$ a gamma distribution, $p(\cdot|H_1)$ a Gaussian mixture, and stopping bounds derived from user-specified false-positive and false-negative probabilities. The test's job is to convert noisy single-frame scores into decisions that require sustained evidence, and its thresholds set the explicit trade-off between missing faults and raising false alarms.

What would settle it

On a held-out sewer video with known ground truth, compute the empirical false-alarm rate of the SPRT system while resampling the frame sequence in blocks (preserving frame-to-frame correlation) rather than as independent frames; if the empirical rate is materially higher than 0.96%, or if a goodness-of-fit test rejects the gamma and Gaussian-mixture models on a larger calibration set, the robustness claim as stated does not transfer to real correlated footage.

Watch

Extended reading notes

Core claim

The central claim is that temporal evidence aggregation and explainable spatial localisation can be combined in one sewer-inspection pipeline: FCDD provides a heatmap and scalar anomaly score for every image, and SPRT models the score distributions under normal and anomalous hypotheses--gamma for normal, Gaussian mixture for anomalous--then accumulates the log-likelihood ratio across frames until one hypothesis is accepted. The authors claim this materially improves detection: on the video dataset, the F1 score rose from 84.07% with the single-frame threshold to 94.36% with SPRT, precision from 79.69% to 97.85%, and the false-positive rate dropped from 11.30% to 0.96%. They also report that the heatmap localises the deposit even though training used only image-level labels, and that the threshold-based system misses one anomaly at around frame 5600, illustrating the deliberate trade-off between decision robustness and sensitivity to short-lived faults.

Load-bearing premise

The load-bearing premise is that the per-frame anomaly scores really follow the fitted distributions (gamma for normal frames, Gaussian mixture for anomalous frames), that those fits are accurate enough when learned from a small calibration set of 75 anomaly frames, and that consecutive video frames can be treated as independent; if any of these fails, the claimed 0.96% false-alarm rate is optimistic.

Editorial extensions

If this is right

Operators can see not just an alert but a heatmap of where the anomaly is in the pipe, which helps confirm and repair faults.
A single-class training setup, without pixel-level labels, can catch a fault class (deposits here) and can be retrained for other defect types with similar data.
SPRT's alpha and beta parameters let a utility tune the system toward fewer false alarms or fewer missed faults, matching maintenance priorities.
On the large single-image benchmark, the explainable one-class detector beat feature-based SVM classifiers by a wide margin, suggesting deep features generalise better for sewer defect detection.
Frame-level noise such as blur and illumination changes is suppressed by evidence accumulation, so alerts are more likely to correspond to persistent faults.

Reading between the lines

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

The error bounds quoted for the sequential test assume independent frames, but consecutive CCTV frames from a moving robot are strongly correlated; in deployment the true false-alarm rate could exceed the reported 0.96% unless the score process is decorrelated or the bounds are adjusted.
Because the score distributions are fitted on a small calibration set (75 anomaly frames), any new pipe material, camera, or lighting regime will shift the scores; a non-parametric or continuously updated version of the test would be more robust across sites.
The paper reports decision quality but not decision latency; a useful extension would report the expected number of frames to reach a decision, or an operating-characteristic curve, to quantify how long a robot must stare at a fault before it is flagged.
The one-class structure could be extended to a multi-hypothesis sequential test over several fault classes, such as cracks, roots, and blockages, while keeping the same heatmap interface.

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.

Desk Editor's Note

read the letter

Quick take: the paper is a competent application of existing methods, but the headline SPRT improvement is not actually demonstrated by the numbers as reported. The problem is real, the authors are transparent, and the flaws are fixable, so it deserves a serious referee, not a desk reject.

What's new: the first FCDD application to sewer pipe CCTV, and the FCDD+SPRT combination for temporal aggregation. That's application-level novelty, not a new method, but it is legitimate for the venue. The WRc experiment is clean: FCDD beats two SVM baselines on a decent-sized image dataset, and the spatial localization in Fig. 2 is genuinely informative without pixel-level training.

Soft spots, in order. First, the iCAIR F1 comparison is apples-to-oranges. The SPRT F1 of 94.36% is computed only on frames that terminated the test; undecided frames are dropped. The single-frame FCDD baseline is computed on all frames. So the two numbers cannot be compared directly, and the abstract's 'demonstrated improved anomaly detection performance' overstates the evidence. The paper does acknowledge this in Section 3, which I give the authors credit for, but the abstract and conclusion should carry the same caveat. A fair protocol would include undecided frames as errors, report decision rates per video, or apply the same frame exclusion to the baseline. Second, the iCAIR training set has 25 anomaly frames and the calibration set 75, so the distribution parameters for the SPRT are fit to very little data. No error bars or multiple runs are reported, so the 94.36% might not be stable. Third, the SPRT assumes the fitted gamma/GMM distributions and frame independence; indoor CCTV frames are autocorrelated, so Wald's error bounds are approximate at best. That is a modeling concern, not a fatal one, but the 0.96% FPR should be read as conditional on the model.

Overall: this is an honest, clearly written systems paper with public code and data. The central idea—temporal aggregation of anomaly scores—is sensible, and the single-frame results are solid. The SPRT benefit needs a fair evaluation before it is claimed.

For a referee: I would take it. The application is timely and the flaws are addressable. Ask for a corrected comparison metric, error bars or at least a sensitivity analysis on the SPRT thresholds, and a discussion of the i.i.d. assumption. If those are fixed, it is a reasonable paper for an applied robotics or infrastructure conference.

Referee Report

4 major / 4 minor

Summary. The paper proposes a two-stage system for automated sewer inspection: a Fully Convolutional Data Description (FCDD) anomaly detector provides per-frame anomaly scores and spatial heatmaps, and a Sequential Probability Ratio Test (SPRT) aggregates these scores across video frames to reach a temporal decision. The method is evaluated on two datasets: the WRc single-image dataset, where FCDD is compared with SVM baselines, and the iCAIR video dataset, where FCDD+SPRT is compared with single-frame FCDD thresholding. The authors report that FCDD+SPRT achieves an F1-score of 94.36% versus 84.07% for single-frame thresholding on iCAIR, and that FCDD outperforms SVMs on WRc.

Significance. If the reported improvement is real, the work would be a useful practical contribution: it combines pixel-level explainability with temporal evidence accumulation, and the authors state that code and model are publicly available, which aids reproducibility. The idea of applying SPRT to anomaly scores from a deep one-class classifier is sensible and the paper is clearly written at a high level. However, the headline quantitative claim suffers from a comparison on different frame populations, and the small size of the iCAIR training set plus the i.i.d. assumption in the SPRT derivation limit the strength of the conclusions. The paper is a plausible workshop-length contribution, but the evaluation needs substantial tightening before the performance claims can be accepted.

major comments (4)

[Section 3, Table 1] The headline F1 comparison is not apples-to-apples. The FCDD+SPRT F1 of 94.36% is computed on the frames on which SPRT reached a decision, while the FCDD baseline of 84.07% is computed on all frames. The paper itself states in Section 3 that 'the total frame count used in SPRT's metric calculations was reduced due to undecided frames between threshold boundaries.' Because undecided frames are excluded only from the SPRT metric, the improvement could reflect the removal of hard cases rather than better detection. The authors should either include undecided frames as errors (for example, counting an undecided anomalous frame as a false negative), report per-video decision rates and the fraction of frames decided, or apply the same frame-exclusion rule to the single-frame baseline. Without this, Eq. (8) error bounds and the reported 0.96% FPR are conditional on termination and do not support the stated improvement.
[Section 2.2, Eq. (4)] Wald's SPRT error bounds in Eq. (8) assume that the observations z_t are independent and identically distributed under each hypothesis. Consecutive video frames of a sewer inspection are strongly temporally correlated, and the authors do not address this. If the gamma and Gaussian mixture PDFs fitted on the iCAIR calibration set (75 anomaly frames) are misspecified, or if the temporal correlation invalidates the i.i.d. assumption, the claimed false positive rate of 0.96% is optimistic. The authors should either provide evidence that the PDFs adequately describe the score distributions, apply a sequential test that accounts for dependence, or explicitly qualify the FPR as an idealized bound.
[Section 2.3, iCAIR training setup] The iCAIR model is trained on only 25 anomaly frames and calibrated on 75 anomaly frames. The SPRT PDFs, especially the Gaussian mixture for p(z_t|H1), are fitted on 75 anomaly score values, which is a very small sample for estimating multiple mixture components. No uncertainty or sensitivity analysis is reported, so it is unclear whether the favorable SPRT results are robust to the choice of PDF model or to random variation in the calibration set. The authors should report confidence intervals, repeat calibration/splits, or use a simpler model for p(z_t|H1) that is better matched to the sample size.
[Table 1] The comparison of FCDD with SVM baselines on the WRc dataset is performed on a different dataset and task setting than the iCAIR video evaluation. The paper uses WRc to claim that 'FCDD improved anomaly detection over an SVM method,' but the SVM comparison is not used in the iCAIR evaluation, and the iCAIR model is trained on only 25 anomaly frames. The reader cannot infer from Table 1 that FCDD+SPRT would outperform SVM-based temporal fusion; the two halves of the table should be clearly separated, and the claims in the Abstract and Section 4 should be scoped accordingly.

minor comments (4)

[Eq. (1)] The equation for the anomaly heatmap A(X) is typeset ambiguously: it appears to contain a square root and the term '+ 1 - 1' is unclear. Please re-typeset as A(X) = sqrt(phi(X;W)^2 + 1) - 1 or the intended formula, and define phi explicitly.
[Section 3, Fig. 3] The caption 'Threshold Hits SPRT Hits Ground Truth' is not fully descriptive; it should state which color corresponds to which method and what the y-axis or legend entries (Normal, Undecided, Anomaly) represent.
[Section 2.3] The sentence describing the iCAIR dataset as 'publicly available' should cite the exact repository rather than only reference [2], which is a previous paper that may not contain a direct download link.
[Section 2.2] The notation a and b in Eq. (8) uses approximations with inequalities; it would be clearer to state that the decision limits are a = log(beta/(1-alpha)) and b = log((1-beta)/alpha), and that these are approximations due to the discrete overshoot of the random walk.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the SPRT and FCDD components are evaluated on held-out data; the only self-citation (iCAIR dataset) is a data source, not a load-bearing argument.

full rationale

The paper's derivation chain is not circular. FCDD is trained with a one-class objective (Eq. 3) on independent training splits (WRc training set; iCAIR training split), the threshold tau is tuned on calibration data ('Anomaly score predictions using a calibration dataset were used to tune a threshold parameter tau'), and the SPRT densities p(zt|H1) and p(zt|H0) are fitted to calibration anomaly scores (Section 2.2). The final F1=94.36% vs 84.07% is computed on a held-out iCAIR evaluation split (5475 frames), so the headline result is not an identity with the fitted inputs. The iCAIR dataset is cited to the authors' prior work [2], but it is a publicly available data collection described externally and used as ground truth; this is a data self-citation, not a self-citation that supplies the paper's conclusions. The acknowledged reduction in frame count for SPRT metrics ('the total frame count used in SPRT's metric calculations was reduced due to undecided frames') is a legitimate statistical comparability concern, but it is an evaluation-flaw issue, not circularity: SPRT decisions are still generated from held-out frames. No uniqueness theorem, ansatz, or renamed empirical pattern is imported from the authors' prior work. Therefore no circular step is established by the paper's own equations.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim depends on user-specified SPRT error bounds, thresholds and PDF parameters fitted on calibration data, and on distributional and i.i.d. assumptions for the anomaly scores. No new entities are introduced.

free parameters (5)

FCDD threshold tau (WRc) = 1.09
Tuned on WRc calibration data using Youden's index to separate normal and anomalous samples.
FCDD threshold tau (iCAIR) = 0.64
Tuned on iCAIR calibration data, used as the single-frame baseline threshold.
SPRT Type I error alpha = 1e-6
User-specified false positive probability chosen to demonstrate the effect of reducing false positives.
SPRT Type II error beta = 0.01
User-specified false negative probability used in the reported configuration.
SPRT PDF parameters (gamma and Gaussian mixture) = not reported
Parameters of the gamma (H0) and Gaussian mixture (H1) score distributions are fit to calibration scores and directly determine the log-likelihood ratio in Eq. (4).

assumptions (3)

domain assumption Anomaly scores under H0 follow a gamma distribution and under H1 follow a Gaussian mixture model.
Invoked in Section 2.2; the adequacy of these distributional families is not validated against the data.
domain assumption Video frames are independent and identically distributed draws from the anomaly score distributions.
Implicit in applying SPRT, which assumes i.i.d. observations. Consecutive video frames are temporally correlated, so this assumption is likely violated.
domain assumption The FCDD loss in Eq. (3) trains a valid one-class model and the Inception-v3 backbone features transfer to sewer imagery.
Background assumption borrowed from FCDD literature [8] and not explicitly tested on the sewer datasets beyond the reported results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Explainable Deep Anomaly Detection with Sequential Hypothesis Testing for Robotic Sewer Inspection." pith.science (2026). https://pith.science/paper/JUNNW5D5

@misc{pith2026250722546,
  author       = {Pith},
  title        = {Pith review of: Explainable Deep Anomaly Detection with Sequential Hypothesis Testing for Robotic Sewer Inspection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUNNW5D5}},
  note         = {Machine review of arXiv:2507.22546}
}

read the original abstract

Sewer pipe faults, such as leaks and blockages, can lead to severe consequences including groundwater contamination, property damage, and service disruption. Traditional inspection methods rely heavily on the manual review of CCTV footage collected by mobile robots, which is inefficient and susceptible to human error. To automate this process, we propose a novel system incorporating explainable deep learning anomaly detection combined with sequential probability ratio testing (SPRT). The anomaly detector processes single image frames, providing interpretable spatial localisation of anomalies, whilst the SPRT introduces temporal evidence aggregation, enhancing robustness against noise over sequences of image frames. Experimental results demonstrate improved anomaly detection performance, highlighting the benefits of the combined spatiotemporal analysis system for reliable and robust sewer inspection.

Figures

Figures reproduced from arXiv: 2507.22546 by the authors.

**Figure 2.** (a) Sewer inspection robot. (b) Example of CCTV image with anomaly. (c) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Anomaly predictions for FCDD on the iCAIR video frames from thresh [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 14 canonical work pages

[1]

Drinkwater, A., Moy, F.: Wipes in sewers blockage study (2017), https://www.water.org.uk/news-views-publications/publications/wipes-sewers- blockage-study, last accessed 2025/03/24

work page 2017
[2]

Frontiers in Robotics and AI10, 1150508 (2023)

Edwards, S., Zhang, R., Worley, R., Mihaylova, L., Aitken, J., Anderson, S.R.: A robust method for approximate visual robot localization in feature-sparse sewer pipes. Frontiers in Robotics and AI10, 1150508 (2023)

work page 2023
[3]

IEEE Access 8, 39574–39586 (2020)

Fang, X., Guo, W., Li, Q., Zhu, J., Chen, Z., Yu, J., Zhou, B., Yang, H.: Sewer pipelinefault identificationusing anomaly detection algorithms onvideo sequences. IEEE Access 8, 39574–39586 (2020)

work page 2020
[4]

George, A.: Explainable-Anomaly-Detection-SPRT (2025), https://github.com/ alexgeorge13/Explainable-Anomaly-Detection-SPRT, last accessed 2025/07/24

work page 2025
[5]

In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

Haurum, J.B., Moeslund, T.B.: Sewer-ml: A multi-label sewer defect classification dataset and benchmark. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 13456–13467 (June 2021)

work page 2021
[6]

Journal of Computing in Civil Engineering34(1), 04019047 (2020)

Kumar, S.S., Wang, M., Abraham, D.M., Jahanshahi, M.R., Iseley, T., Cheng, J.C.: Deep learning–based automated detection of sewer defects in CCTV videos. Journal of Computing in Civil Engineering34(1), 04019047 (2020)

work page 2020
[7]

Sensors22(7), 2722 (2022)

Li,Y.,Wang,H.,Dang,L.M.,Song,H.K.,Moon,H.:Vision-baseddefectinspection and condition assessment for sewer pipes: A comprehensive survey. Sensors22(7), 2722 (2022)

work page 2022
[8]

In: International Conference on Learning Representations (2021)

Liznerski, P., Ruff, L., Vandermeulen, R.A., Franks, B.J., Kloft, M., Muller, K.R.: Explainable deep one-class classification. In: International Conference on Learning Representations (2021)

work page 2021

Show all 15 references

[9]

Journal of Pipeline Systems Engineering and Practice 11(4), 03120002 (2020)

Malek Mohammadi, M., Najafi, M., Kermanshachi, S., Kaushal, V., Sera- jiantehrani, R.: Factors influencing the condition of sewer pipes: State-of-the-art review. Journal of Pipeline Systems Engineering and Practice 11(4), 03120002 (2020)

2020
[10]

ACM computing surveys (CSUR)51(5), 1–36 (2018)

Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, Y., Reyes, M.P., Shyu, M.L., Chen, S.C., Iyengar, S.S.: A survey on deep learning: Algorithms, techniques, and applications. ACM computing surveys (CSUR)51(5), 1–36 (2018)

2018
[11]

Rayhana, R., Jiao, Y., Zaji, A., Liu, Z.: Automated vision systems for condition assessmentofsewerandwaterpipelines.IEEETransactionsonAutomationScience and Engineering 18(4), 1861–1878 (2020)

2020
[12]

In: Proceedings of the IEEE International Conference on Computer Vision

Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- CAM: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 618– 626 (2017)

2017
[13]

https://spring-innovation.co.uk, last accessed 2025/07/16

Spring: Spring Innovation - Accelerating Water Sector Transformation. https://spring-innovation.co.uk, last accessed 2025/07/16

2025
[14]

In: Breakthroughs in statistics: Foundations and basic theory, pp

Wald, A.: Sequential tests of statistical hypotheses. In: Breakthroughs in statistics: Foundations and basic theory, pp. 256–298. Springer (1992)

1992
[15]

In: 2022 26th International Conference on Pattern Recog- nition (ICPR)

Zhao, M., Liu, Y., Liu, J., Zeng, X.: Exploiting spatial-temporal correlations for video anomaly detection. In: 2022 26th International Conference on Pattern Recog- nition (ICPR). pp. 1727–1733. IEEE (2022)

2022

Pith tools

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

[1] [1]

Drinkwater, A., Moy, F.: Wipes in sewers blockage study (2017), https://www.water.org.uk/news-views-publications/publications/wipes-sewers- blockage-study, last accessed 2025/03/24

work page 2017

[2] [2]

Frontiers in Robotics and AI10, 1150508 (2023)

Edwards, S., Zhang, R., Worley, R., Mihaylova, L., Aitken, J., Anderson, S.R.: A robust method for approximate visual robot localization in feature-sparse sewer pipes. Frontiers in Robotics and AI10, 1150508 (2023)

work page 2023

[3] [3]

IEEE Access 8, 39574–39586 (2020)

Fang, X., Guo, W., Li, Q., Zhu, J., Chen, Z., Yu, J., Zhou, B., Yang, H.: Sewer pipelinefault identificationusing anomaly detection algorithms onvideo sequences. IEEE Access 8, 39574–39586 (2020)

work page 2020

[4] [4]

George, A.: Explainable-Anomaly-Detection-SPRT (2025), https://github.com/ alexgeorge13/Explainable-Anomaly-Detection-SPRT, last accessed 2025/07/24

work page 2025

[5] [5]

In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

Haurum, J.B., Moeslund, T.B.: Sewer-ml: A multi-label sewer defect classification dataset and benchmark. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 13456–13467 (June 2021)

work page 2021

[6] [6]

Journal of Computing in Civil Engineering34(1), 04019047 (2020)

Kumar, S.S., Wang, M., Abraham, D.M., Jahanshahi, M.R., Iseley, T., Cheng, J.C.: Deep learning–based automated detection of sewer defects in CCTV videos. Journal of Computing in Civil Engineering34(1), 04019047 (2020)

work page 2020

[7] [7]

Sensors22(7), 2722 (2022)

Li,Y.,Wang,H.,Dang,L.M.,Song,H.K.,Moon,H.:Vision-baseddefectinspection and condition assessment for sewer pipes: A comprehensive survey. Sensors22(7), 2722 (2022)

work page 2022

[8] [8]

In: International Conference on Learning Representations (2021)

Liznerski, P., Ruff, L., Vandermeulen, R.A., Franks, B.J., Kloft, M., Muller, K.R.: Explainable deep one-class classification. In: International Conference on Learning Representations (2021)

work page 2021