Pith. sign in

REVIEW 4 major objections 4 minor 13 references

AdaptoVision: A Multi-Resolution Image Recognition Model for Robust and Scalable Classification

T0 review · 4 major / 4 minor · reviewed 2026-08-16 · 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 6.6-million-parameter CNN, AdaptoVision, can match or beat much larger models on standard benchmarks and reach top reported accuracy on breast-tumor images, all without pretrained weights.

desk verdict The architecture is a plausible mixture of known techniques, but the paper's own tables contradict its SOTA claims and the BreakHis evaluation likely leaks; desk-reject. read the letter →

arxiv 2504.12652 v2 pith:DRMWAKI7 submitted 2025-04-17 cs.CV

Md. Sanaullah Chowdhury Lameya Sabrin This is my paper · ORCID
classification cs.CV
keywords AdaptoVisionefficientCNNimageclassificationdepth-wiseseparableconvolutionenhancedresidualunithierarchicalskipconnectionBreakHisCIFAR
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 claims that a deliberately compact convolutional network, AdaptoVision, can match or exceed much larger models on standard benchmarks (95.3% on CIFAR-10, 85.77% on CIFAR-100) and can reach top reported accuracy on the BreakHis breast-tumor dataset (99.43% magnification-independent), all without pretrained weights. The motivation is practical: if true, high-accuracy image recognition does not require billion-parameter models or transfer learning, which matters for phones, edge devices, and clinical settings with limited compute. The authors attribute the gain to three architectural choices—enhanced residual units, depth-wise separable convolutions in a residual block, and hierarchical skip connections—and support the claim with parameter and FLOP counts far below ResNet-18 and comparable to MobileNetV2.

What carries the argument

The load-bearing components are (1) the Enhanced Residual Unit (ERU), which replaces a shallow residual mapping with a deeper transformation chain $T(x)=W^{(4)}*\sigma(W^{(3)}*B_2(\sigma(W^{(2)}*\sigma(W^{(1)}*x))))$; (2) Block-2, a residual-compatible inception-like unit using $1\times1$ pointwise convolutions around a depth-wise convolution, costing $O(h\cdot w\cdot c\cdot k^2)$; (3) hierarchical skip connections that sum feature maps from prior encoder blocks after projection, keeping gradient flow across depth; and (4) a global average pooling plus reshape aggregation before the final projection. The complexity formula $O_{ERU}=2h w c^2 k^2$, together with hardware-aware constraints, is the quantitative backbone for the efficiency claim.

What would settle it

Re-run the BreakHis evaluation with a patient-level split (all tiles of each patient in one set only) and compare the reported 99.43% magnification-independent accuracy; if accuracy falls markedly—say below 95%—the SOTA claim is inflated by tile leakage.

Watch

Extended reading notes

Core claim

The paper's central claim is that a from-scratch CNN using enhanced residual units, a depthwise-separable 'Block-2', and long-range skip connections can achieve accuracy competitive with or better than much larger pretrained models while using roughly 6.6 million parameters and 4.9 GFLOPs. On BreakHis binary classification the paper reports 99.43% accuracy across magnifications, above every comparison listed; on ISIC 2019 it reports 95.30%; on CIFAR-10/100 it reports 95.30%/85.77%, 'comparable' to strong baselines though below the best listed at 98-99.5% on CIFAR-10. The authors frame the result as evidence that architectural efficiency and information flow, not scale or pretraining, can carry competitive classification.

Load-bearing premise

The BreakHis 99%+ accuracies assume that the 70/30 train/test split keeps tiles originating from the same source image from appearing in both sets; the paper never states or justifies the split level.

Editorial extensions

If this is right

  • If the claims hold, high-accuracy classifiers can be trained from scratch on consumer hardware, removing the reliance on pretrained ImageNet weights.
  • The reported parameter and FLOP reductions imply that similar accuracy can be reached with an order-of-magnitude less compute than ResNet-18, which matters for real-time and embedded deployment.
  • On BreakHis, the architecture would offer a single magnification-independent model that beats specialized medical pipelines without transfer learning.
  • The same design principles—deep residual transforms plus depthwise separable operations—could transfer to other fine-grained medical or benchmark tasks.

Reading between the lines

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

  • Extending the paper's logic, the architecture's efficiency suggests it could serve as a backbone for object detection or segmentation, where parameter counts multiply quickly; the paper does not test this.
  • The strongest test of the BreakHis claim would be re-running the pipeline with a patient-level split; if accuracy stays near 99%, the generalization claim is substantially stronger than the tile-level protocol currently supports.
  • The CIFAR numbers are presented as 'comparable' rather than state-of-the-art; a fair reader inference is that the architecture trades a few accuracy points for large efficiency gains, and the useful comparison is accuracy per FLOP rather than raw accuracy.
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

4 major / 4 minor

Summary. The paper proposes AdaptoVision, a CNN architecture combining enhanced residual units, depth-wise separable convolutions, and hierarchical skip connections. The authors claim state-of-the-art results on BreakHis and ISIC 2019, and accuracy comparable to heavier models on CIFAR-10/100, Flowers-102, and Stanford Cars, while using significantly fewer parameters and FLOPs. The paper reports experiments on these datasets and compares with various baselines.

Significance. If the reported results were valid, a lightweight CNN achieving near-SOTA across multiple benchmarks would be a useful contribution to efficient image recognition. However, the paper's own tables contradict the central claims: the ImageNet result is far below every baseline listed, and the BreakHis SOTA claim is not supported by the numbers in the same table. The reported efficiency gains rest on an incorrect complexity comparison. No code or detailed experimental protocol is provided, further limiting the value of the empirical results.

major comments (4)
  1. [Section 4, Table 7] The text states that the proposed method 'consistently outperforms existing ConvNets' on ImageNet, but Table 7 reports 66.70% top-1 and 86.60% top-5, while every listed baseline achieves at least 76.0% top-1. Moreover, EfficientNet-B0 achieves 77.1% top-1 with 5.3M parameters and 0.3B FLOPs, while AdaptoVision uses 6.7M parameters and 4.9G FLOPs; the proposed model is therefore neither more accurate nor more efficient than this baseline. This contradiction invalidates the paper's efficiency-accuracy trade-off claim.
  2. [Section 3.1.1, Tables 4 and 5] The BreakHis evaluation protocol is not described at the level needed to rule out data leakage. The paper states that images were tiled with a sliding window (step 4 pixels, tile size 126x128) and that the dataset was split 70/30, but it does not state whether the split is at the patient, image, or tile level. Because adjacent tiles overlap heavily, a tile-level split would place near-duplicate crops in both training and test sets and could inflate the reported accuracies to the 98-99% range. Without code or clarification, the SOTA claims in Tables 4 and 5 are not verifiable.
  3. [Section 4, Table 4] The claim that AdaptoVision 'achieved state-of-the-art accuracy in all three magnification images' is contradicted by the same table: Yari et al. report 100% at 40X and 100X, and 99.26% magnification-independent, whereas the proposed method reports 99.37%, 98.67%, 99.33%, 99.17%, and 99.43%. Even accepting the reported numbers, the proposed method is not the top performer in the table, so the SOTA claim is unsupported.
  4. [Section 3.4, Table 2] The complexity comparison is not credible: the table caption says input size 224x224, while the text says 32x32; more importantly, ResNet-18 is listed at 768 GFLOPs, which is roughly 400 times larger than the standard published value (~1.8 GFLOPs at 224x224). With an inflated baseline, the claim that AdaptoVision 'reduces FLOPs by nearly 4x' has no quantitative support. The parameter count comparison (ResNet-18 11.7M vs. proposed 6.6M) is also less favorable than the 'more than 4x' claim suggests.
minor comments (4)
  1. [Section 3.3 and Table 1] The table number 'Table 1' is used both for the dataset summary in Section 3.1 and for the architecture configuration in Section 3.3; renumber to avoid ambiguity.
  2. [References] AlexNet is cited as Hinton et al. [2012], but the canonical citation is Krizhevsky et al. (2012); several references have incomplete bibliographic data (e.g., 'dvvf', 'bhjbjh', 'Unknown Booktitle').
  3. [Figure 2] The text references 'Figure 2: Exploring Accuracy...', but no actual figure appears in the manuscript; check whether the figure is missing or the reference is erroneous.
  4. [Section 4.1, ISIC 2019] The claim of SOTA accuracy on ISIC 2019 is based on Table 6, which lists only one baseline and no statistical significance; the claim should be qualified accordingly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: AdaptoVision's claims are empirical measurements and standard arithmetic, not derivations that reduce to their own inputs.

full rationale

This is an empirical architecture paper. The load-bearing claims are measured accuracies obtained from trained models, not quantities derived from assumed premises that already contain the conclusion. The complexity expressions in Eqs. (8), (15), and (16) are standard FLOP-count formulas applied to the proposed architecture, and Table 2 simply reports the arithmetic result; this is definitional counting, not a fitted parameter renamed as a prediction. No parameter is fitted to a subset of data and then 'predicted' on a closely related quantity. The references are all to external prior work, and no load-bearing self-citation chain or imported uniqueness theorem is used to force the model choice. The possible BreakHis train/test leakage concern raised in the review notes is a data-protocol validity issue, not circularity: even if the 70/30 split is applied at the tile level rather than the patient level, that would inflate empirically measured accuracies, but it would not make an output equivalent to an input by construction. The 'state-of-the-art' language is a comparative claim that may be contestable on the basis of the cited baselines, but disputing a comparison is not exhibiting circularity. Accordingly, no specific circular step can be quoted, and the appropriate score is 0.

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

The paper introduces no new scientific entities, but it rests on several unstated domain assumptions about data splitting, comparability of benchmarks, and evaluation metrics. The architecture hyperparameters are hand-chosen and incompletely specified, so the exact model cannot be reconstructed from the text.

free parameters (6)
  • Initial learning rate = 0.175
    Hand-chosen for SGD; no sensitivity analysis provided (Section 4.1).
  • Learning rate decay factor and period = 0.99 every 12.4 epochs
    Exponential decay schedule stated without justification (Section 4.1).
  • Dropout rate schedule = 0.3 to 0.5 across blocks
    Progressively increased dropout; no rationale or ablation for the schedule (Section 4.1).
  • Width scaling factor alpha and gamma/delta = alpha in {1/2, 1, 2}; gamma and delta unnamed
    Eqs. (7) and (12) define width changes but the actual per-stage values are not listed.
  • Skip connection weights alpha1 and alpha2 = fixed or learnable, not specified
    Eq. (13) uses scalar weights; their values and whether they are learned are not stated.
  • BreakHis tiling parameters = tile size 126x128, step size 4 pixels
    Specific augmentation and tiling choices are given without variation or validation (Section 3.1.1).
assumptions (5)
  • standard math The FLOPs formula O=2*h*w*c^2*k^2 estimates compute cost for a standard convolution.
    Used in Eqs. (8) and (15); this is a standard approximation and is acceptable.
  • domain assumption Accuracy numbers from prior papers are comparable to AdaptoVision even though train/test splits differ (70/30, 80/20, 5-fold, 90/10).
    Tables 4-6 compare against methods with different protocols without correction or discussion.
  • domain assumption Tiles generated from the same BreakHis source image are independent samples.
    Required for the 70/30 split to be valid; no patient-level or image-level split is described in Section 3.1.1.
  • domain assumption ISIC 2019 accuracy computed over an imbalanced set is a meaningful state-of-the-art metric.
    Table 6 reports a single accuracy number (95.30%) without balanced accuracy or per-class metrics, which is a known issue for ISIC 2019.
  • ad hoc to paper The ERU and Block-2 design choices improve accuracy over a vanilla residual network.
    The architecture is justified only by final accuracy; there is no ablation isolating each component (Sections 3.2 and 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of AdaptoVision: A Multi-Resolution Image Recognition Model for Robust and Scalable Classification." pith.science (2026). https://pith.science/paper/DRMWAKI7

@misc{pith2026250412652,
  author       = {Pith},
  title        = {Pith review of: AdaptoVision: A Multi-Resolution Image Recognition Model for Robust and Scalable Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DRMWAKI7}},
  note         = {Machine review of arXiv:2504.12652}
}
read the original abstract

This paper introduces AdaptoVision, a novel convolutional neural network (CNN) architecture designed to efficiently balance computational complexity and classification accuracy. By leveraging enhanced residual units, depth-wise separable convolutions, and hierarchical skip connections, AdaptoVision significantly reduces parameter count and computational requirements while preserving competitive performance across various benchmark and medical image datasets. Extensive experimentation demonstrates that AdaptoVision achieves state-of-the-art on BreakHis dataset and comparable accuracy levels, notably 95.3\% on CIFAR-10 and 85.77\% on CIFAR-100, without relying on any pretrained weights. The model's streamlined architecture and strategic simplifications promote effective feature extraction and robust generalization, making it particularly suitable for deployment in real-time and resource-constrained environments.

Figures

Figures reproduced from arXiv: 2504.12652 by the authors.

Figure 2
Figure 2. Exploring Accuracy Across Image Dimensions using ImageNet Pretrained Weights: Insights into Model [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. BreakHis Binary Classification accuracy graph [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. BreakHis Binary Classification comparison graph [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 3 canonical work pages

  1. [8]

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen

    URLhttps://arxiv.org/abs/1805.09501. Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks,

    arXiv

  2. [10]

    Histopathological breast cancer image classification by deep neural network techniques guided by local clustering.BioMed research international, 2018,

    Abdullah-Al Nahid, Mohamad Ali Mehrabi, and Yinan Kong. Histopathological breast cancer image classification by deep neural network techniques guided by local clustering.BioMed research international, 2018,

    work page 2018

  3. [12]

    Kassem, Khalid M

    Mohamed A. Kassem, Khalid M. Hosny, and Mohamed M. Fouad. Skin lesions classification into eight classes for isic 2019 using deep convolutional neural network and transfer learning.IEEE Access, 8:114822–114832,

    work page 2019

  4. [13]

    doi:10.1109/ACCESS.2020.3003890. 13

    arXiv 2020

  5. [1995]

    Improving neural networks by preventing co-adaptation of feature detectors.arXiv preprint arXiv:1207.0580,

    Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors.arXiv preprint arXiv:1207.0580,

    arXiv

  6. [2010]

    doi:10.1162/NECO_a_00052

    ISSN 0899-7667. doi:10.1162/NECO_a_00052. URLhttps://doi.org/10.1162/NECO_a_00052. Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive cutmix: An enhanced data augmenta- tion approach for deep learning based image classification.arXiv preprint arXiv:2003.13048,

    arXiv

  7. [2012]

    Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556,

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556,

    arXiv

  8. [2016]

    Marc Combalia, Noel C

    doi:10.1109/TBME.2015.2496264. Marc Combalia, Noel C. F. Codella, Veronica Rotemberg, Brian Helba, Veronica Vilaplana, Ofer Reiter, Cristina Carrera, Alicia Barreiro, Allan C. Halpern, Susana Puig, and Josep Malvehy. Bcn20000: Dermoscopic lesions in the wild,

    arXiv 2015

Show all 13 references
  1. [2017]

    Highway networks

    Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway networks. arxiv 2015.arXiv preprint arXiv:1505.00387,

    2015 arXiv

  2. [2018]

    Coca: Contrastive captioners are image-text foundation models.arXiv preprint arXiv:2205.01917,

    Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models.arXiv preprint arXiv:2205.01917,

    arXiv

  3. [2019]

    Mahesh Gour, Sweta Jain, and T Sunil Kumar

    URLhttps://arxiv.org/abs/1801.04381. Mahesh Gour, Sweta Jain, and T Sunil Kumar. Residual learning based cnn for breast cancer histopathological image classification.International Journal of Imaging Systems and Technology, 30:621–635,

    arXiv

  4. [2020]

    doi:10.1109/TCSVT.2019.2935128. Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V . Le. Autoaugment: Learning augmentation policies from data,

    2019

  5. [2022]

    doi:https://doi.org/10.1016/j.iswa.2022.200066

    ISSN 2667-3053. doi:https://doi.org/10.1016/j.iswa.2022.200066. URL https://www.sciencedirect.com/science/article/ pii/S2667305322000072. Md Zahangir Alom, Chris Yakopcic, Mst Shamima Nasrin, Tarek M Taha, and Vijayan K Asari. Breast cancer classification from histopathologica...

    2022

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.