Pith. sign in

REVIEW 6 minor 33 references

Synthetic and Derived Training Images for Campus Waste Detection: A Multi-Seed Evaluation with YOLOv8n

T0 review · 0 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read Derived training images never beat real-only waste detection.

desk verdict A careful, honest negative-result study in a narrow subfield; the central claim is supported on its own scoped terms, and the paper is worth refereeing despite the single-split limitation. read the letter →

arxiv 2607.19535 v2 pith:CWGSWNTC submitted 2026-07-21 cs.CV

classification cs.CV
keywords wastedetectionobjectYOLOv8nsynthetictrainingdatadomainshiftmulti-seedevaluationbootstrapintervalssmartbin
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 asks whether synthetic and derived training images can improve a compact waste detector for a campus bin camera that watches a person's hand as they discard an item. The answer it defends is no: on a fixed 31-photo real test set, none of the twelve evaluated mixtures of derived images surpassed a model trained only on 86 real photographs. The real-only baseline reached 0.691 mAP@0.5; background-replacement images dropped to 0.560, and the full augmentation pool to 0.487, while nearly size-matched isolated-object images stayed close to baseline at 0.680. The authors use four matched seeds and bootstrap intervals to show that seed-to-seed noise alone cannot explain the main drops, while cautioning that the intervals do not cover split-to-split or test-sample uncertainty. The practical upshot, if the claim holds, is that adding more transformed copies of the same roughly 100 base objects does not add deployment-relevant diversity, so scarce data budgets are better spent on new real object instances.

What carries the argument

The controlled comparison is the machinery: architecture (YOLOv8n), optimizer, epoch budget, validation criterion, and the 31-photo test set are held fixed while the training images vary. Seven principal settings are repeated with four matched seeds, and 95% percentile bootstrap intervals are computed over the four runs, with paired differences resampled seed-wise. The added-image pool consists of 695 derived images from roughly 100 base objects, which allows the paper to separate amount of data from source, and includes an approximately size-matched contrast between isolated-object and background-replacement images. A provenance-corrected hand-composite intervention isolates one deployment-

What would settle it

Collect or construct a group-disjoint test set in which no physical object appears in both training and test, then re-run the same configurations; if the real-only advantage over the 25%-derived mix (or over background replacement) does not reproduce with an independent 31+ photo test set, the paper's central ranking is split-dependent. A simpler check: run the same seven configurations on multiple new splits and see whether the 0.691 real-only reference changes by more than the reported intervals.

Watch

Extended reading notes

Core claim

The central claim is stated in the conclusion: no evaluated synthetic/derived configuration surpassed real-only training on the fixed 31-photo test set. In the multi-seed comparison, the real-only model had mean mAP@0.5 of 0.691 [0.665, 0.722]; adding 25% derived images reduced it to 0.633 [0.608, 0.646] with a paired difference of -0.059 [-0.114, -0.020]; the full augmentation pipeline fell to 0.487 [0.438, 0.537]; the approximately size-matched isolated-object set gave 0.680 [0.644, 0.724], overlapping the baseline. The corrected hand-composite experiment, rebuilt after two test-split cutouts leaked into the initial set, showed a paired difference of +0.034 [-0.063, 0.199], so no reliable

Load-bearing premise

The load-bearing assumption is that the single class-stratified split of 148 photos into 86/31/31, which was not checked for repeated physical objects within source batches, fairly represents the deployment problem; all rankings and intervals rest on that one split, and the bootstrap intervals cover only four training seeds, not split-to-split or test-sample uncertainty.

Editorial extensions

If this is right

  • Adding more transformations of the same base objects does not create deployment-relevant diversity; image count is not evidence of object diversity.
  • Background replacement, the most aggressive contextual change, can hurt mean mAP by about 0.13 relative to real-only training and is less stable across seeds.
  • Nominal training set size is not a reliable lever: larger derived-image pools produced the worst detector, and doubling added data did not restore performance.
  • Hand-composite augmentation built from a tiny training-only library does not reliably restore real-scene behavior; the corrected run does not support a benefit.
  • Model size alone is not a deployability criterion: the INT8 export was slower and produced no class scores, so quantization must be validated on actual outputs.

Reading between the lines

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

  • [Inference] The ranking would likely not survive a group-disjoint split that keeps the same physical object out of both training and test; the paper itself notes the split was not checked for repeated objects, so the -0.059 effect at 25% added data could shrink or reverse.
  • [Inference] A strong testable extension is to repeat the size-matched contrast with several group-disjoint splits and paired class-level metrics; the metal class result (0.360 to 0.586 AP with isolated-object images) hints that aggregate mAP may hide per-class wins.
  • [Inference] If the fixed test set is treated as a sample, the bootstrap intervals understate uncertainty; a reader betting on deployment performance should collect a larger held-out set before concluding synthetic data is useless for this view.
  • [Inference] The result transfers a caution about 'more images' to the smart-bin setting: before investing in generative pipelines, a practitioner should verify that new synthetic instances add physical-object diversity rather than re-rendering the same instances.
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

0 major / 6 minor

Summary. The paper evaluates whether synthetic and derived training images improve a YOLOv8n detector for a campus bin-mounted camera. The authors merge 148 real photographs (86 train / 31 validation / 31 test), create a 695-image derived pool from ~100 base objects, and compare 12 joint-training configurations against a real-only baseline on a fixed 31-photo test set. Seven principal settings are repeated with four matched seeds, and bootstrap percentile intervals over seeds are reported. The central finding is that no evaluated configuration exceeded the real-only baseline (mean mAP@0.5 = 0.691 [0.665, 0.722]). The paper also documents a corrected hand-composite experiment after discovering test-split leakage, exploratory transfer experiments, per-class and difficult-condition diagnostics, and a four-thread CPU benchmark.

Significance. The paper's main value is its careful, honestly scoped negative result. It explicitly distinguishes seed variation from test-split uncertainty, reports a leaked experiment and its correction, and avoids overclaiming: intervals are described as descriptive only, and class-level results are labeled tentative. The approximately size-matched comparison between isolated-object and background-replacement sources is a clean test of source effects. If the result withstands routine scrutiny, it provides a useful counterexample to the assumption that simply adding derived images from the same base objects improves small-data detection. The authors also provide seed-level data and code on request, and their limitation section (§6) is unusually thorough. The single-split concern is real but acknowledged; the central claim is explicitly scoped to the fixed 31-photo test set, so the absence of group-disjoint splits does not contradict the paper's stated conclusions.

minor comments (6)
  1. [§5.4] The text switches between four-seed means and seed-2026 values without explicit flags; e.g., 'isolated-object result from 0.683 to 0.618' uses seed 2026, while §5.2 reports the four-seed mean 0.680. Please clarify which value is used in each comparison.
  2. [§4.1] 'Automatic learning-rate selection' is vague; specify the exact learning rate or the Ultralytics default (e.g., lr0=0.01) to aid reproducibility.
  3. [§5.7] 'All 8,400 detection positions' is unexplained; define how this number derives from the 31 test images and network output.
  4. [Data Availability] No public repository is provided; while the authors state they will provide code on request, a permanent archive (e.g., Zenodo) would strengthen reproducibility.
  5. [Abstract / §7] The sentence 'None of the evaluated configurations exceeded the real-only baseline' could be misread as no individual seed ever exceeded; consider adding 'in mean across seeds' for precision.
  6. [§7] The practical recommendation to collect new object instances rather than transformations is reasonable but contingent on the single fixed split; consider explicitly noting that group-disjoint replication is required before adopting it for other deployments.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: direct empirical comparison with self-contained stats; limitations concern generality, not derivation.

full rationale

The paper is an empirical comparison, not a derivation. The central claim ('No evaluated synthetic/derived configuration surpassed real-only training on the fixed 31-photo test set') is a measured outcome on a fixed split; real-only training is an independent baseline, and each synthetic/derived configuration is a separate training run. The bootstrap percentile intervals are explicitly described as descriptive summaries over four seeds and are not population-level tests, so no fitted parameter is renamed as a prediction. The added-image pool's reuse of ~100 base objects is disclosed and used as an interpretive limitation, not as a result derived from itself. The one contamination event (two hand-composite cutouts from test photos) was detected, discarded, and rerun with training-only cutouts; the paper reports the corrected experiment as its RQ2 answer and retains the discarded result only for traceability, so the final claim does not reduce to leaked test information. The acknowledged limitations (single split, unverified group-disjoint separation, small test set) concern generality and representativeness, not circularity. No load-bearing self-citation or uniqueness theorem is invoked. Therefore no circular step can be exhibited.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

This is an empirical evaluation, not a derivation: the central claim rests on no fitted theory parameters or postulated entities. What it rests on are data assumptions — representativeness of the fixed split, class identity of the unprovenanced synthetic pool, adequacy of the fixed detector platform — and design choices (epoch budgets, thresholds) whose effects the paper mostly discloses. The two hand-picked analysis choices listed do not drive the primary comparison; the RQ1 dose–response and RQ3 transfer conclusions are the most exposed to them.

free parameters (2)
  • Training epoch budget (100 joint / 100+40 sequential) = 100 epochs; 100 synthetic + 40 real
    Chosen by hand in §4.1/§4.3. The unequal budgets between joint and sequential training confound the RQ3 integration-strategy comparison with extra optimization exposure; the paper acknowledges this. It does not drive the primary RQ1/RQ2 claims.
  • Operating-point thresholds (confidence 0.25, NMS IoU 0.7, mAP conf floor 0.001) = 0.25 / 0.7 / 0.001
    Fixed analysis thresholds in §4.1. Thresholded precision/recall and outcome counts depend on them, but the primary mAP comparisons integrate over the full precision–recall curve, so the central claim is insensitive to these choices.
assumptions (4)
  • domain assumption The fixed 31-photo test split is an adequate basis for comparing configurations
    §3.1 splits by random seed 2026 without group-disjoint separation by physical object; §4.2 and §6 state the intervals omit test-sample and split-to-split uncertainty. If the split is unrepresentative, every ranking in the paper shifts.
  • domain assumption The 100 base objects are genuine examples of the four target classes
    §3.2: 'We found no collection record, license, or attribution that establishes who captured these images or how they were created.' Class membership rests on visual inspection, which the authors flag as insufficient for provenance.
  • domain assumption YOLOv8n with COCO pretraining and Ultralytics default training is a controlled platform for comparing training sets
    §4.1: architecture, initialization source, optimizer, and checkpoint rule are held fixed by design; conclusions are bounded to this detector family and are not claimed to transfer to other architectures.
  • standard math Bootstrap percentile intervals over four matched runs are used descriptively
    §4.2 correctly states that four observations cannot support a stable estimate of a training-run distribution and that even a unanimous sign test with n=4 cannot reach p<0.05; the statistics are not over-interpreted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Synthetic and Derived Training Images for Campus Waste Detection: A Multi-Seed Evaluation with YOLOv8n." pith.science (2026). https://pith.science/paper/CWGSWNTC

@misc{pith2026260719535,
  author       = {Pith},
  title        = {Pith review of: Synthetic and Derived Training Images for Campus Waste Detection: A Multi-Seed Evaluation with YOLOv8n},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWGSWNTC}},
  note         = {Machine review of arXiv:2607.19535}
}
read the original abstract

Incorrect disposal can contaminate campus recycling streams, and a bin-mounted camera could provide feedback as an item is discarded. We evaluated whether synthetic and derived images improve a YOLOv8n detector for this view. The real dataset contained 148 campus photographs: 86 for training, 31 for validation, and 31 for testing. Twelve joint-training configurations varied the amount and source of added images. We repeated seven principal settings with four matched seeds and computed bootstrap percentile intervals over those seeds. The real-only model reached a mean mAP@0.5 of 0.691 [0.665, 0.722]. Background replacement reduced the mean to 0.560 [0.499, 0.619], isolated-object images gave 0.680 [0.644, 0.724], and the full augmentation pool gave 0.487 [0.438, 0.537]. We also tested hand-and-forearm composites because every real photo showed a held object. Two cutouts in the initial composite set came from test photographs, so we discarded that experiment, rebuilt the set with training-split cutouts, and reran all four seeds. The corrected paired difference was +0.034 [-0.063, 0.199], which does not support a reliable hand-composite effect. Single-seed transfer experiments produced source-dependent rankings between joint mixing and sequential pretraining. None of the evaluated configurations exceeded the real-only baseline. The reported intervals quantify seed variation; the 31-photo test set remains too small for strong class-specific conclusions.

Figures

Figures reproduced from arXiv: 2607.19535 by the authors.

Figure 1
Figure 1. Study design. We merged and verified three batches of real photos, then trained YOLOv8n [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Dataset composition. (a) The 148 real photos are dominated by plastic; glass is rare, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Examples from the added-image pool. Columns 1–3 show three base objects per class. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Training configurations and image counts. The upper rows vary the amount of added [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Test-set mAP@0.5 as added images are mixed with the 86 real training photographs. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Four-seed results. (a) Mean mAP@0.5 and bootstrap percentile intervals; dots show [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Four-seed mean metrics for background replacement, the corrected training-only hand [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Background-replacement images (top) and their corrected hand-composited counterparts [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Joint mixing and sequential pretrain–fine-tune results for seed 2026. Sequential runs [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Per-class AP@0.5 by image source (four-seed mean [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Model predictions on real test photos (green: ground truth; red: prediction). Row 1: [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Seed-2026 mAP@0.5 for five overlapping condition subsets pooled from validation and [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Four-thread CPU benchmark. The FP32 models have similar size, latency, and mAP. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 4 linked inside Pith

  1. [1]

    Artificial intelligence applications in solid waste management: A systematic research review,

    M. Abdallah, M. Abu Talib, S. Feroz, Q. Nasir, H. Abdalla, and B. Mahfood, “Artificial intelligence applications in solid waste management: A systematic research review,”Waste Management, vol. 109, pp. 231–246, 2020

  2. [2]

    A survey on image data augmentation for deep learning,

    C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmentation for deep learning,” Journal of Big Data, vol. 6, no. 1, Art. no. 60, 2019

  3. [3]

    Cut, paste and learn: Surprisingly easy synthesis for instance detection,

    D. Dwibedi, I. Misra, and M. Hebert, “Cut, paste and learn: Surprisingly easy synthesis for instance detection,” inProc. IEEE Int. Conf. Computer Vision (ICCV), 2017, pp. 1301–1310

  4. [4]

    Synthesizing training data for object detection in indoor scenes,

    G. Georgakis, A. Mousavian, A. C. Berg, and J. Kosecka, “Synthesizing training data for object detection in indoor scenes,” inRobotics: Science and Systems (RSS), 2017

  5. [5]

    Simple copy-paste is a strong data augmentation method for instance segmentation,

    G. Ghiasi, Y. Cui, A. Srinivas, R. Qian, T.-Y. Lin, E. D. Cubuk, Q. V. Le, and B. Zoph, “Simple copy-paste is a strong data augmentation method for instance segmentation,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2021, pp. 2918–2928

  6. [6]

    Playing for data: Ground truth from computer games,

    S. R. Richter, V. Vineet, S. Roth, and V. Koltun, “Playing for data: Ground truth from computer games,” inProc. European Conf. Computer Vision (ECCV), 2016, pp. 102–118

  7. [7]

    The SYNTHIA dataset: A large collection of synthetic images for semantic segmentation of urban scenes,

    G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M. Lopez, “The SYNTHIA dataset: A large collection of synthetic images for semantic segmentation of urban scenes,” inProc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016, pp. 3234–3243

  8. [8]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” inProc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS), 2017, pp. 23–30

Show all 33 references
  1. [9]

    Training deep networks with synthetic data: Bridging the reality gap by domain randomization,

    J. Tremblayet al., “Training deep networks with synthetic data: Bridging the reality gap by domain randomization,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition Workshops (CVPR W), 2018, pp. 969–977

  2. [10]

    Contextual priming for object detection,

    A. Torralba, “Contextual priming for object detection,”International Journal of Computer Vision, vol. 53, no. 2, pp. 169–191, 2003

  3. [11]

    An empirical study of context in object detection,

    S. K. Divvala, D. Hoiem, J. H. Hays, A. A. Efros, and M. Hebert, “An empirical study of context in object detection,” inProc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2009, pp. 1271–1278

  4. [12]

    The elephant in the room,

    A. Rosenfeld, R. Zemel, and J. K. Tsotsos, “The elephant in the room,” arXiv:1808.03305, 2018

  5. [13]

    Modeling visual context is key to augmenting object detection datasets,

    N. Dvornik, J. Mairal, and C. Schmid, “Modeling visual context is key to augmenting object detection datasets,” inProc. European Conf. Computer Vision (ECCV), 2018, pp. 364–380

  6. [14]

    Classification of trash for recyclability status,

    G. Thung and M. Yang, “Classification of trash for recyclability status,” CS229 Course Report, Stanford University, 2016

  7. [15]

    TACO: Trash annotations in context for litter detection,

    P. F. Proen¸ ca and P. Sim˜ oes, “TACO: Trash annotations in context for litter detection,” arXiv:2003.06975, 2020. 19

  8. [16]

    Hierarchical waste detection with weakly supervised segmentation in images from recycling plants,

    D. Yudinet al., “Hierarchical waste detection with weakly supervised segmentation in images from recycling plants,”Engineering Applications of Artificial Intelligence, vol. 128, Art. no. 107542, 2024

  9. [17]

    ZeroWaste dataset: Towards deformable object segmentation in clut- tered scenes,

    D. Bashkirovaet al., “ZeroWaste dataset: Towards deformable object segmentation in clut- tered scenes,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2022, pp. 21147–21157

  10. [18]

    Garbage detection and classifi- cation using a new deep learning-based machine vision system as a tool for sustainable waste recycling,

    S. Jin, Z. Yang, G. Kr´ olczyk, X. Liu, P. Gardoni, and Z. Li, “Garbage detection and classifi- cation using a new deep learning-based machine vision system as a tool for sustainable waste recycling,”Waste Management, vol. 162, pp. 123–130, 2023

  11. [19]

    Faster R-CNN: Towards real-time object detec- tion with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards real-time object detec- tion with region proposal networks,” inAdvances in Neural Information Processing Systems (NeurIPS), 2015, pp. 91–99

  12. [20]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inProc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016, pp. 779–788

  13. [21]

    SSD: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg, “SSD: Single shot multibox detector,” inProc. European Conf. Computer Vision (ECCV), 2016, pp. 21–37

  14. [22]

    MobileNetV2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted residuals and linear bottlenecks,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2018, pp. 4510–4520

  15. [23]

    Ultralytics YOLOv8,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics YOLOv8,” version 8.0, 2023. [Online]. Available:https://github.com/ultralytics/ultralytics

  16. [24]

    Quantization and training of neural networks for efficient integer-arithmetic- only inference,

    B. Jacobet al., “Quantization and training of neural networks for efficient integer-arithmetic- only inference,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2018, pp. 2704–2713

  17. [25]

    Quantizing deep convolutional networks for efficient inference: A whitepa- per,

    R. Krishnamoorthi, “Quantizing deep convolutional networks for efficient inference: A whitepa- per,” arXiv:1806.08342, 2018

  18. [26]

    On pre-trained image features and synthetic images for deep learning,

    S. Hinterstoisser, V. Lepetit, P. Wohlhart, and K. Konolige, “On pre-trained image features and synthetic images for deep learning,” inProc. European Conf. Computer Vision Workshops (ECCVW), 2018, pp. 682–697

  19. [27]

    Approximate statistical tests for comparing supervised classification learning algorithms,

    T. G. Dietterich, “Approximate statistical tests for comparing supervised classification learning algorithms,”Neural Computation, vol. 10, no. 7, pp. 1895–1923, 1998

  20. [28]

    Accounting for variance in machine learning benchmarks,

    X. Bouthillieret al., “Accounting for variance in machine learning benchmarks,” inProceedings of Machine Learning and Systems (MLSys), vol. 3, 2021, pp. 747–769

  21. [29]

    Efron and R

    B. Efron and R. J. Tibshirani,An Introduction to the Bootstrap. New York, NY, USA: Chap- man & Hall, 1993

  22. [30]

    Microsoft COCO: Common objects in context,

    T.-Y. Linet al., “Microsoft COCO: Common objects in context,” inProc. European Conf. Computer Vision (ECCV), 2014, pp. 740–755. 20

  23. [31]

    The PASCAL visual object classes (VOC) challenge,

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman, “The PASCAL visual object classes (VOC) challenge,”International Journal of Computer Vision, vol. 88, no. 2, pp. 303–338, 2010

  24. [32]

    MediaPipe Hands: On-device real-time hand tracking,

    F. Zhang, V. Bazarevsky, A. Vakunov, A. Tkachenka, G. Sung, C.-L. Chang, and M. Grund- mann, “MediaPipe Hands: On-device real-time hand tracking,” arXiv:2006.10214, 2020

  25. [33]

    How transferable are features in deep neural networks?

    J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable are features in deep neural networks?” inAdvances in Neural Information Processing Systems (NeurIPS), 2014, pp. 3320– 3328. 21

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.