Pith. sign in

REVIEW 4 major objections 6 minor 27 references

Exploring Active Learning for Semiconductor Defect Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a combination of contrastive pretraining on unlabelled X-ray scans and a rareness-aware acquisition function lets active learning reach 98% of fully supervised defect-segmentation performance using only about 4.9%…

desk verdict Solid applied AL study; the contrastive pretraining result is real, but the rareness-aware acquisition term is not yet supported by the evidence. read the letter →

arxiv 2507.17359 v1 pith:JCAEP5XQ submitted 2025-07-23 cs.CV

Lile Cai , Ramanpreet Singh Pahwa , Xun Xu , Jie Wang , Richard Chang , Lining Zhang , Chuan-Sheng Foo This is my paper · ORCID
classification cs.CV
keywords activelearningsemanticsegmentationsemiconductordefectdetectionX-raymicroscopycontrastivepretrainingclassimbalancerareness-awareacquisitionhighbandwidthmemory
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 active learning can cut the annotation cost of training deep segmentation models for semiconductor inspection. It identifies two obstacles specific to X-ray microscopy scans: a large domain shift from natural-image pretraining, and severe class imbalance in which the defect class (void) occupies only a tiny area. To counter both, the authors pretrain a U-Net with SimCLR-style contrastive learning on the unlabelled scans and re-use those weights to initialize the model each active-learning cycle, and they add a rareness term to the acquisition function that favors images whose pseudo-labels contain rare classes. On a dataset of X-ray scans of high bandwidth memory structures, the method reaches 78.18% mIoU with 200 annotated images, about 4.9% of the training set, which is 98% of the fully supervised model's score. The paper's claim is that this combination is the current state of the art for this data.

What carries the argument

The load-bearing machinery is the rareness-aware acquisition function $s(I)=r(I)+u(I)+d(I,L)$, where $r(I)$ aggregates pixel-wise rareness scores $r(x)=e^{-p(\hat{y}(x))}$ computed from pseudo-labels of the current model, $u(I)$ is predictive entropy, and $d(I,L)$ is the distance to already selected images, with max pooling used for aggregation. The other half is contrastive pretraining: the segmentation model's decoder output is globally pooled and passed through a two-layer projection head, and an InfoNCE loss is minimized on the unlabelled training set; the resulting encoder-decoder weights initialize every active-learning cycle.

What would settle it

Take the proposed pipeline but construct the first annotated batch so that it contains no void pixels. If, over several active-learning cycles, the rareness term does not pull void-containing images into the labelled set and void mIoU stays near zero, the central claim that rareness-aware selection drives rare-class performance would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that both components---contrastive pretraining on the target unlabelled data and rareness-aware selection---are needed to make active learning work under large domain shift and class imbalance. With SimCLR weights instead of ImageNet weights, early-cycle segmentation improves substantially; with the rareness term added to entropy and feature-distance scores, mIoU improves by about 1.06 percentage points over entropy alone at a budget of 200 images. The authors attribute part of the gain to max aggregation of pixel rareness scores, which lets a single rare void pixel drive image selection. At the final tested budget the full pipeline reaches 78.18% mIoU, about 98% of the fully supervised performance, while using only about 4.9% of the training labels.

Load-bearing premise

The rareness score is built from pseudo-labels produced by the current model, so if that model never predicts the rare void class, the rareness term cannot tell which images actually contain voids and selection falls back to uncertainty and diversity.

Editorial extensions

If this is right

  • With the proposed pipeline, annotating roughly 4.9% of the training images (200 of 4,086) yields 78.18% mIoU, about 98% of the fully supervised upper bound.
  • Contrastive pretraining on unlabelled domain data matters most when the labelling budget is small, exactly the regime active learning targets.
  • Adding rareness to uncertainty-plus-diversity selection improves rare-class segmentation on imbalanced semiconductor data without sacrificing overall mIoU.
  • The method outperforms Random, Entropy, CoreSet, CoreGCN, VAAL, and BADGE at all tested labelling budgets on this dataset.

Reading between the lines

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

  • A stress test outside the paper's experiments would start a cycle from a first batch containing no void images; because pseudo-labels would assign zero probability to the void class, the rareness term could not rank void-containing images, so any remaining gain would come from uncertainty and diversity alone.
  • The mechanism should transfer to other tiny-defect inspection domains, such as medical or materials imaging, but only if the segmentation model's pseudo-labels are confident enough that the estimated class frequencies $p(c)$ are informative.
  • The contrastive pretraining step costs about 3.5 hours on one V100 GPU; a deployment decision would weigh that fixed cost against the annotation savings, a trade-off the paper does not quantify.
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

4 major / 6 minor

Summary. The manuscript addresses active learning for semantic segmentation of semiconductor XRM images, motivated by annotation cost and by two challenges: domain shift from natural images and severe class imbalance. The authors propose (i) SimCLR-style contrastive pretraining on unlabeled scans to initialize a U-Net at each active-learning cycle, and (ii) a rareness-aware acquisition score that combines a pseudo-label-based pixel-rareness term with entropy uncertainty and feature-space diversity. Experiments on a 4,086-image training set split at scan level compare the method with Random, Entropy, CoreSet, CoreGCN, VAAL, and BADGE; with 200 labeled images (about 4.9% of training data) the method reaches 78.18% mIoU, about 98% of the fully supervised model. The paper reports that contrastive pretraining markedly outperforms ImageNet initialization, and that the rareness term gives a small additional gain over entropy plus diversity.

Significance. If fully supported, the paper would be practically valuable: it applies active learning to a real industrial inspection task, evaluates on a scan-level split that avoids leakage, uses five-run means, and compares against several standard AL baselines. The contrastive-pretraining result is particularly encouraging, with large and likely robust gains in the low-budget regime. However, the novelty attributable to the rareness-aware term is not yet established because the reported gains are small relative to run-to-run variability and because the max-aggregated rareness score degenerates under the rare-void geometry. The paper therefore needs additional analysis before the state-of-the-art claim can be accepted.

major comments (4)
  1. [Sec. 3.2, Eq. (3)] The max aggregation makes the rareness term a near-binary presence indicator rather than a graded rarity signal. With p(void) close to zero for the rare void class, any image in which the current model predicts at least one void pixel receives r(I) approximately exp(-0) = 1, whereas the same term for a void-free image is bounded by exp(-p(common)), typically well below 1. Consequently, r(I) cannot order images by void count or area, and selection among void-containing images is driven only by u(I) + d(I,L). This means the mechanism claimed in the abstract and Section 3.2 is not what Eq. (4) actually implements. Please report the fraction of unlabeled images with at least one predicted void pixel per cycle, compare the acquisition function against a variant with a binary void-presence indicator, and consider a sum or class-balanced aggregation if a graded rareness signal is intended.
  2. [Table 1a and text after Sec. 4.2] The ablation table does not unambiguously support the sentence "rareness term improves mIoU by 1.06% over Entropy, and 0.19% over Entropy+Feature." The row "✓ ✓ 78.00" is not labeled, and the reported differences are smaller than the standard deviations (e.g., 78.19 ± 0.40 versus 77.92 ± 0.79). Without paired significance tests (for example, paired permutation or bootstrap across the five runs, or additional runs), the incremental contribution of the rareness term over Entropy+Feature is within noise and cannot be claimed. Please label each row explicitly and provide significance tests for all budgets displayed in Fig. 3b.
  3. [Fig. 3b and Sec. 4.2] The paper does not report numeric mIoU means and standard deviations for the baseline methods at each budget, nor significance tests for the pairwise comparisons. Since some differences at budget=200 are likely within one standard deviation, the statement that the method "consistently outperforms" other methods is not quantitatively supported. Please include a table with all methods and budgets, report means and standard deviations, and perform paired tests between the proposed method and each baseline.
  4. [Sec. 3.2, pseudo-label reliability] The rareness score depends on pseudo-labels from the model trained in the previous cycle. In early cycles with very few labels, the model may never predict the rare void class, so p(void)=0 and the rareness term cannot guide selection toward true void-containing images. The paper does not analyze this failure mode or ablate the effect of cycle number. Please report pseudo-label void recall at each cycle and, if possible, the performance of an oracle variant that uses ground-truth void frequency.
minor comments (6)
  1. [Sec. 3.1] There is a typo: "constrastive loss" should be "contrastive loss."
  2. [Table 1a] The row labels are ambiguous; please use explicit labels such as "Entropy", "Entropy+Feature", "Entropy+Rareness", and "Entropy+Feature+Rareness" in the table or caption.
  3. [Sec. 3.2, Eq. (2)] Please clarify whether p(c) is computed over the unlabeled pool, the labeled set, or all pixels in the training split; the text says "training set" but pseudo-labels are used for querying unlabeled images.
  4. [Sec. 4.1] The description "other augmentation techniques and hyper-parameters are kept the same as used in [21]" is vague; please list the augmentation operations and hyper-parameter values explicitly.
  5. [Conclusion] There is a typo: "state-of-art-performance" should be "state-of-the-art performance."
  6. [Fig. 3] The error bars are difficult to read at the printed figure size; consider enlarging them or providing a numeric table of all plotted values.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: empirical active learning evaluation with external benchmarks; self-citation is motivational, not load-bearing.

full rationale

The paper makes no first-principles derivation claim; all reported numbers are empirical evaluations on a held-out test split, with training/testing separated at the 3D scan level to avoid leakage. Contrastive pretraining (Eq. 1) uses only unlabeled data, and the rareness-aware acquisition function (Eqs. 2-4) is a heuristic scoring rule whose inputs are pseudo-label frequencies from the current model. This pseudo-label feedback is self-referential in the sense of being model-dependent, but it is not circular: the method is evaluated against held-out manual annotations and compared to external baselines (Random, Entropy, CoreSet, CoreGCN, VAAL, BADGE) under identical pretrained weights and initial batch. The only self-citation is [17], used to motivate the design choice that labeling rare-class samples improves learning on imbalanced data; it is a motivational citation with independent peer-reviewed support, not used to derive any equation or to validate a predicted output. The potential degenerate behavior of the max-aggregated rareness score (reduction to a near-binary void-presence indicator) is a robustness/effectiveness concern, not circularity, because no claimed result is assumed by construction. Hence no significant circularity is found.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims are empirical and do not require new physical entities. The main burden falls on hyperparameter choices and the pseudo-label reliability assumption.

free parameters (6)
  • InfoNCE temperature tau = default from [21] (value not reported)
    Temperature in Eq. 1 controls the sharpness of the contrastive loss; adopted from the self-damaging contrastive learning library without tuning on the semiconductor data.
  • Contrastive pretraining schedule = 2000 epochs, batch size 256, lr 0.5 with cosine
    Taken from [21]; no sensitivity analysis is reported for these settings on the target dataset.
  • Rareness function r(x) = exp(-p(c)) = exponential decay
    Chosen in Eq. 2 without derivation or comparison to other monotonic functions such as 1-p(c) or -log p(c); the functional form affects which images are selected.
  • Acquisition term weights = 1, 1, 1 for r, u, d
    Eq. 4 sums rareness, uncertainty, and diversity with equal weights; no ablation or tuning of these weights is reported.
  • Aggregation function faggr = max (vs mean)
    Table 1b shows max outperforms mean; if this choice was made using test set results, it constitutes test-set model selection.
  • Batch budget K per AL cycle = 50 images
    The annotation budget increments by 50 (Fig. 3); the reported 98% figure depends on this arbitrary budget size.
assumptions (5)
  • standard math InfoNCE contrastive loss (Eq. 1) is a valid objective for learning useful representations from unlabeled images.
    The method builds entirely on this loss; it is a well-established result from prior work [15].
  • domain assumption The unlabeled XRM training data used for contrastive pretraining is drawn from the same distribution as the test data.
    The train/test split is at the 3D scan level (Section 4.1), but the similarity of the unlabeled pool to the test distribution is not measured.
  • domain assumption Pseudo-labels from the model trained in the previous AL cycle are reliable enough to estimate class frequencies in Eq. (2).
    If the model never predicts the rare void class, p(void)=0 and the rareness term cannot identify void-containing images. The paper does not validate this assumption.
  • domain assumption Labeling more samples from rare classes improves model performance on imbalanced segmentation.
    The motivation for the rareness term relies on the authors' prior work [17]; the paper does not re-test this on the semiconductor data.
  • domain assumption U-Net with a ResNet-18 backbone and weighted cross-entropy loss is an appropriate segmentation model for XRM scans.
    No comparison to other architectures is made; the choice is standard but unvalidated for this domain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Active Learning for Semiconductor Defect Segmentation." pith.science (2026). https://pith.science/paper/JCAEP5XQ

@misc{pith2026250717359,
  author       = {Pith},
  title        = {Pith review of: Exploring Active Learning for Semiconductor Defect Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCAEP5XQ}},
  note         = {Machine review of arXiv:2507.17359}
}
read the original abstract

The development of X-Ray microscopy (XRM) technology has enabled non-destructive inspection of semiconductor structures for defect identification. Deep learning is widely used as the state-of-the-art approach to perform visual analysis tasks. However, deep learning based models require large amount of annotated data to train. This can be time-consuming and expensive to obtain especially for dense prediction tasks like semantic segmentation. In this work, we explore active learning (AL) as a potential solution to alleviate the annotation burden. We identify two unique challenges when applying AL on semiconductor XRM scans: large domain shift and severe class-imbalance. To address these challenges, we propose to perform contrastive pretraining on the unlabelled data to obtain the initialization weights for each AL cycle, and a rareness-aware acquisition function that favors the selection of samples containing rare classes. We evaluate our method on a semiconductor dataset that is compiled from XRM scans of high bandwidth memory structures composed of logic and memory dies, and demonstrate that our method achieves state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2507.17359 by the authors.

Figure 4
Figure 4. Our method (Rareness-Aware) is able to segment the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages

  1. [1]

    Exploring Active Learning for Semiconductor Defect Segmentation

    INTRODUCTION The development of X-Ray microscopy (XRM) technology has enabled non-destructive techniques (NDT) applications in inspection of semiconductor structures. Facilitated by machine learning and sophisticated image processing, it is now possible to automatically identify important structures in semiconductor XRM scans. A use case is illustrated in...

    work page Pith review arXiv 2025

  2. [2]

    Uncertainty- based methods select samples that the current model is most uncertain about to label

    RELA TED WORK Active Learning Based on the criterion used to query sam- ples, AL methods can be broadly categorized into uncertainty- based, diversity-based and hybrid methods. Uncertainty- based methods select samples that the current model is most uncertain about to label. Ensemble-based method [4] has shown to provide more calibrated uncertainty estima...

    work page

  3. [3]

    rareness

    METHOD In this section, we first introduce our method to perform con- trastive pretraining with segmentation models, followed by description on how we perform rareness-aware sampling to select samples for annotation. 3.1. Contrastive Pretraining for Segmentation Models The loss function in contrastive learning measures the simi- larities of sample pairs i...

    work page

  4. [4]

    Experimental Setup Datasets Our dataset is compiled from 3D XRM scans of high bandwidth memory structures composed of logic and memory dies

    EXPERIMENTS 4.1. Experimental Setup Datasets Our dataset is compiled from 3D XRM scans of high bandwidth memory structures composed of logic and memory dies. The logic die consists of three classes, namely, copper pillar, solder and void; the memory die contains one addi- tional class named copper pad. The dataset contains 25 3D scans for logic die, and 5...

    work page

  5. [5]

    CONCLUSIONS In this work, we explored active learning for semiconductor defect segmentation. We proposed using contrastive pretrain- ing for initializing the segmentation model, and proposed a rareness-aware acquisition function to prioritize samples con- taining minority classes for labelling. When benchmarked on a semiconductor dataset composed of XRM s...

    work page

  6. [6]

    Active learning for convolutional neural networks: A core-set approach,

    Ozan Sener and Silvio Savarese, “Active learning for convolutional neural networks: A core-set approach,” arXiv preprint arXiv:1708.00489, 2017

    arXiv 2017

  7. [7]

    Deep Learning Analysis of 3D X-ray Images for Automated Object Detection and Attribute Measure- ment of Buried Package Features,

    Ramanpreet Singh Pahwa, Ma Tin Lay Nwe, Richard Chang, Wang Jie, Oo Zaw Min, David Ho Soon Wee, Ren Qin, Vempati Srinivasa Rao, Yanjing Yang, Jens Timo Neumann, Ramani Pichumani, and Tom Gre- gorich, “Deep Learning Analysis of 3D X-ray Images for Automated Object Detection and Attribute Measure- ment of Buried Package Features,” in 22nd Electronics Packag...

    work page 2020

  8. [8]

    Au- tomated attribute measurements of buried package fea- tures in 3D X-ray images using deep learning,

    Ramanpreet Singh Pahwa, Ma Tin Lay Nwe, Richard Chang, Oo Zaw Min, Wang Jie, Saisubramaniam Gopalakrishnan, David Ho Soon Wee, Ren Qin, Vem- pati Srinivasa Rao, Haiwen Dai, Jens Timo Neu- mann, Ramani Pichumani, and Tom Gregorich, “Au- tomated attribute measurements of buried package fea- tures in 3D X-ray images using deep learning,” in 71st Electronic C...

    work page 2021

Show all 27 references
  1. [9]

    Automated Detection and Segmentation of HBMs in 3D X-ray Images using Semi-Supervised Deep Learning,

    Ramanpreet Singh Pahwa, Richard Chang, Wang Jie, Xu Xun, Oo Zaw Min, Foo Chuan Sheng, Chong Ser Choong, and Vempati Srinivasa Rao, “Automated Detection and Segmentation of HBMs in 3D X-ray Images using Semi-Supervised Deep Learning,” in 72nd Electronic Components and Technolog...

    2022

  2. [10]

    The power of ensembles for active learning in image classification,

    William H Beluch, Tim Genewein, Andreas N¨urnberger, and Jan M K ¨ohler, “The power of ensembles for active learning in image classification,” in CVPR, 2018, pp. 9368–9377

    2018

  3. [11]

    Learning loss for active learning,

    Donggeun Yoo and In So Kweon, “Learning loss for active learning,” in CVPR, 2019, pp. 93–102

    2019

  4. [12]

    Exploring simple siamese representation learning,

    Xinlei Chen and Kaiming He, “Exploring simple siamese representation learning,” in CVPR, 2021, pp. 15750–15758

    2021

  5. [13]

    Sequential graph convolutional network for ac- tive learning,

    Razvan Caramalau, Binod Bhattarai, and Tae-Kyun Kim, “Sequential graph convolutional network for ac- tive learning,” in CVPR, 2021, pp. 9583–9592

    2021

  6. [14]

    Variational adversarial active learning,

    Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell, “Variational adversarial active learning,” inICCV, 2019, pp. 5972–5981

    2019

  7. [15]

    Deep batch active learning by diverse, uncertain gradient lower bounds,

    Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal, “Deep batch active learning by diverse, uncertain gradient lower bounds,” arXiv preprint arXiv:1906.03671, 2019

    1906 arXiv

  8. [16]

    A simple framework for contrastive learning of visual representations,

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton, “A simple framework for contrastive learning of visual representations,” inInternational con- ference on machine learning . PMLR, 2020, pp. 1597– 1607

    2020

  9. [17]

    Momentum contrast for unsupervised visual representation learning,

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick, “Momentum contrast for unsupervised visual representation learning,” in CVPR, 2020, pp. 9729–9738

    2020

  10. [18]

    U-net: Convolutional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Med- ical image computing and computer-assisted interven- tion. Springer, 2015, pp. 234–241

    2015

  11. [19]

    During each AL cycle, the model is trained with RMSprop optimizer with weighted cross entropy loss

    backbone. During each AL cycle, the model is trained with RMSprop optimizer with weighted cross entropy loss. The weight for each class is set inversely to the class fre- quency in current labelled data. Hyper-parameters are set as follows: number of epochs = 50, learning rate...

    2000

  12. [20]

    Rethinking deep active learning: Using unlabeled data at model training,

    Oriane Sim ´eoni, Mateusz Budnik, Yannis Avrithis, and Guillaume Gravier, “Rethinking deep active learning: Using unlabeled data at model training,” inICPR. IEEE, 2021, pp. 1220–1227

    2021

  13. [21]

    Deep active learning for biased datasets via fisher kernel self-supervision,

    Denis Gudovskiy, Alec Hodgkinson, Takuya Yam- aguchi, and Sotaro Tsukizawa, “Deep active learning for biased datasets via fisher kernel self-supervision,” in CVPR, 2020, pp. 9041–9049

    2020

  14. [22]

    Representation learning with contrastive predictive coding,

    Aaron Van den Oord, Yazhe Li, and Oriol Vinyals, “Representation learning with contrastive predictive coding,” arXiv e-prints, pp. arXiv–1807, 2018

    2018

  15. [23]

    Improved baselines with momentum contrastive learning,

    Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He, “Improved baselines with momentum contrastive learning,” arXiv preprint arXiv:2003.04297, 2020

    2003 arXiv

  16. [24]

    Revisiting superpixels for active learning in semantic segmentation with realistic annotation costs,

    Lile Cai, Xun Xu, Jun Hao Liew, and Chuan Sheng Foo, “Revisiting superpixels for active learning in semantic segmentation with realistic annotation costs,” in CVPR, 2021, pp. 10988–10997

    2021

  17. [25]

    Deep residual learning for image recognition,

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

    2016

  18. [26]

    Segmentation models py- torch,

    Pavel Yakubovskiy, “Segmentation models py- torch,” https://github.com/qubvel/ segmentation$\_$models.pytorch, 2020

    2020

  19. [27]

    Self-damaging contrastive learn- ing,

    Ziyu Jiang, Tianlong Chen, Bobak Mortazavi, and Zhangyang Wang, “Self-damaging contrastive learn- ing,” in International Conference on Machine Learning, 2021

    2021

Pith tools

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