Impact of Tuning Parameters in Deep Convolutional Neural Network Using a Crack Image Dataset

Ho-Jin Choi; Jia Uddin; Mahe Zabin; Md. Monirul Islam

REVIEW 4 major objections 5 minor 21 references

Impact of Tuning Parameters in Deep Convolutional Neural Network Using a Crack Image Dataset

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read For a two-class crack-image dataset, the paper finds that a DCNN with max pooling, the Adam optimizer, and tanh activation reaches 95.42 percent validation accuracy.

desk verdict The paper's headline claim (tanh+Adam+maxpooling is best) is contradicted by its own Table II, and the validation-set selection design makes any ranking unreliable. read the letter →

arxiv 2506.03184 v1 pith:LSD6QEFA submitted 2025-05-30 cs.CV cs.AIcs.LG

Mahe Zabin , Ho-Jin Choi , Md. Monirul Islam , Jia Uddin This is my paper

classification cs.CVcs.AIcs.LG

keywords DeepCNNtuningparameterscrackimageclassificationpoolinglayersoptimizeractivationfunctionsAdamtanh

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

The paper asks how three tuning choices—pooling type, activation function, and optimizer—change the accuracy of a small deep convolutional network on a two-class crack image dataset. It reports that max pooling beats average pooling across almost all combinations, and that the highest training accuracy (1) and validation accuracy (0.9542) come from max pooling, the Adam optimizer, and tanh activation. The practical point is a concrete starting configuration for binary crack detection: if these results hold, practitioners can skip much of the hyperparameter search and begin with max pooling, Adam, and tanh.

What carries the argument

The central object is a shallow DCNN and the tuning grid around it: max pooling selects the maximum value in each $2 \times 2$ window as the way to down-sample, Adam is the optimizer that updates weights using adaptive first and second moments, and tanh is the activation that maps pre-activations to $(-1, 1)$. The paper uses the full cross-product of the three parameter families as the experimental machinery, isolating the impact of each by holding the other two fixed in three staged comparisons.

What would settle it

Show the max-pooling, Adam, and tanh configuration evaluated on a disjoint test set of crack images that was never used to select among configurations; if its accuracy falls materially below 0.9542 or below another configuration, the claimed ranking fails.

Watch

Extended reading notes

Core claim

Using a fixed DCNN with two convolutional layers, two pooling layers, one dropout layer, and a dense layer, the paper sweeps 8 activation functions (relu, tanh, gelu, elu, selu, silu, softmax, softplus), 5 optimizers (SGD, AdaGrad, AdaDelta, RMSProp, Adam), and 2 pooling schemes (max and average). The winning result is claimed for max pooling, Adam, and tanh: perfect training accuracy and 0.9542 validation accuracy on 800 positive and negative crack images. The paper attributes the result to max pooling preserving salient features, Adam combining the strengths of AdaGrad and RMSProp for noisy sparse gradients, and tanh's symmetric range suiting binary classes.

Load-bearing premise

The reported 95.42 percent validation accuracy is treated as an honest measure of how the chosen configuration will generalize, even though the validation split was used to select that configuration and no independent test set was held out.

Editorial extensions

If this is right

For this binary crack-image task and this network size, the sweep singles out max pooling, Adam, and tanh as the highest-accuracy configuration.
Average pooling trails max pooling in nearly every optimizer/activation pair, so the paper recommends max pooling when salient local features matter most.
Softplus results sit near chance level across most combinations, indicating it fails to learn the two-class crack pattern here.
Adam beats SGD, AdaGrad, AdaDelta, and RMSProp under max pooling and tanh, which the paper attributes to adaptive moment estimation and robustness to noisy gradients.

Reading between the lines

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

Because the validation split was used to pick the winning configuration, the reported 0.9542 is likely optimistic relative to truly unseen data; an independent test set would quantify the gap.
With 800 images and single-run accuracies, differences between the top few configurations—such as relu's 0.9667 validation accuracy versus tanh's 0.9542—may be within run-to-run noise, so rankings should be checked with repeated runs or cross-validation.
The recommended combination is specific to this dataset and architecture; deeper networks or multi-class damage classification would require their own tuning sweeps before the same choice is assumed.

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's headline result is contradicted by its own Table II, and the experimental design doesn't support any ranking anyway.

What's actually here: it's a straightforward grid search over 8 activations, 5 optimizers, and 2 pooling types on a public 800-image crack dataset with a small two-conv-layer CNN. That is a legitimate thing to document, and the tables are readable. The writing is clear and the references are mostly appropriate. If the numbers were internally consistent, this might be a minor empirical data point for that dataset and architecture.

Where it falls apart: the abstract and conclusion claim tanh+Adam+maxpooling gives the best validation accuracy, 0.9542. Table II shows relu+Adam and gelu+Adam both hit 0.9667 with max pooling. The paper even says in Section 3 that the highest validation accuracy is for relu. So the main claim is false on the paper's own evidence. That is not a subtle issue; it is the central empirical assertion.

The deeper problem is that the authors pick the best configuration by looking at validation accuracy, then report that same validation accuracy as the result. There is no held-out test set, no error bars, and no mention of hyperparameters like learning rate, batch size, or epoch count. That makes any ranking selection-biased. Even if the internal contradiction were fixed, the headline number would not be an unbiased estimate of generalization.

Minor issues: the claim that 'max-pooling works better in all cases' is not directly shown in a table, though Table II does support it for most rows. The paper has a few typos and sloppy phrases, but nothing that matters compared to the above.

Who this is for: no one, really. A reader looking for guidance on activation/optimizer choice for crack detection won't get a reliable answer because the paper's own numbers don't support the answer it gives. This is a desk reject in my view. It would need a complete re-run with a proper train/validation/test split and corrected reporting before it could be taken seriously.

Referee Report

4 major / 5 minor

Summary. The paper reports an empirical study of tuning parameters for a small deep convolutional neural network (2 convolutional layers, 2 pooling layers, 1 dropout, 1 dense layer) applied to a two-class crack image dataset of 800 images. It compares 8 activation functions, 5 optimizers, and 2 pooling types, reporting training and validation accuracies for each combination in Table II. The abstract and conclusion claim that the best configuration is max pooling with the Adam optimizer and tanh activation, with a validation accuracy of 0.9542. The manuscript also includes ablations in Tables III–V to motivate this choice.

Significance. Hyperparameter selection is a practically relevant topic, and the paper's systematic sweep over 8 activations, 5 optimizers, and 2 pooling types is a reasonable experimental design. The full accuracy tables are transparent and could serve as a useful comparison if the results were internally consistent and methodologically sound. However, the central claim is directly contradicted by the paper's own data, and the evaluation protocol lacks an independent test set, so the reported numbers cannot support the stated conclusion. The paper also omits essential training details, preventing reproducibility and meaningful comparison with other studies.

major comments (4)

[Abstract and Section 4 (Conclusion) versus Table II] The central claim that max pooling, Adam, and tanh yield the highest validation accuracy (0.9542) is contradicted by Table II: for max pooling, ReLU+Adam and GELU+Adam both achieve validation accuracy 0.9667, strictly higher than 0.9542. The conclusion's statement that 'highest ... validation accuracy 0.9542' was observed for tanh is therefore false under the paper's own reported numbers.
[Section 3, Table II] The claims that 'the max-pooling works better in all cases' and that 'Adam optimizer demonstrated the highest accuracy both in training and validation' are contradicted by the table. For example, softplus+adagrad has max-pooling validation accuracy 0.4958 but average-pooling validation accuracy 0.8625, so average pooling is better in that case. For max pooling, ReLU+rmsprop training accuracy is 0.9946, exceeding ReLU+adam's 0.9181. These overgeneralizations form the basis for selecting max pooling and Adam in subsequent steps, so the selection rationale is invalid.
[Sections 3.2, 3.3, and Tables IV, V] The configuration is selected by inspecting validation accuracies in Tables II, IV, and V, and then the same validation accuracies are presented as the final results. No independent test set is held out. The reported 0.9542 for the selected model is therefore an optimistic, selection-biased estimate rather than an unbiased evaluation of generalization. The paper should re-evaluate the chosen configuration on a separate test set or at a minimum use a nested or cross-validated procedure.
[Section 2 (Methodology)] The manuscript does not report the learning rate, number of epochs, batch size, dropout rate, number of convolutional filters, kernel sizes, or the train/validation split ratio. Without these details, the experiments cannot be reproduced, and the observed differences between configurations cannot be attributed to the manipulated variables alone. This is a load-bearing omission for a study whose entire aim is to compare tuning parameters.

minor comments (5)

[Section 3, Table II] The row 'Relu adam' in Table II mixes case and abbrevations inconsistently ('Relu' vs 'relu' elsewhere); please use consistent capitalization throughout.
[Section 3.1, Table III] Table III only compares tanh+Adam across the two pooling types, which is a single slice of Table II. The narrative that 'max-pooling demonstrates the highest score when the activation function and optimizer are the same' is not supported by a comprehensive comparison; Table II already contains all pairwise comparisons and should be cited instead.
[Section 3, text after Table II] The explanation of tanh versus ReLU is informal and contains technical imprecision (e.g., the claim that tanh leads to a 'fading gradient problem' for inputs outside [-2, 2] is a statement about saturated regions, not a well-defined property). Consider clarifying the gradient behavior and avoiding the term 'dead neuron' unless referring specifically to ReLU.
[References] Reference [5] is described as 'adam activation and softmax optimizers'; this appears to be a misuse of terminology (Adam is an optimizer, not an activation, and softmax is usually an output activation, not an optimizer). Please correct the reference summary.
[Throughout] Several typographical errors and inconsistent capitalizations appear (e.g., 'demonestrated', 'tems', 'pooing', 'Deep CNN' vs 'DCNN'); a careful proofreading pass is needed.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed best configuration is selected by the validation accuracy in Table V and then the same validation accuracy is reported as the concluding result; no independent test set exists, and the 'highest' claim contradicts Table II.

fitted input called prediction [Section 3.3 (Table V) and Section 4 (Conclusion)]
"In depicted in Table V, the adam optimizer demonestrated the highest classification accuracy for the tanh activation function and max-pooling. ... we have observed the highest training accuracy 1 and validation accuracy 0.9542 for the max pooling, adam optimizer, and tanh function for the crack image dataset having 2 classes."

The configuration (max pooling, Adam, tanh) is chosen because Table V reports it as having the highest validation accuracy, and the conclusion then reports that same Table V validation accuracy (0.9542) as the paper's result. The selection criterion and the reported outcome are the same quantity: no held-out test set independently evaluates the chosen configuration. Thus the 'best configuration' claim is forced by construction from the validation table rather than confirmed by any separate prediction. This is compounded by Table II, where relu+Adam and gelu+Adam under max pooling both reach validation accuracy 0.9667, above the claimed 0.9542, so the selected row is not even the maximum in the paper's own data.

full rationale

This paper is an empirical ablation study rather than a derivation, so most circularity patterns (self-citation chains, imported uniqueness theorems, ansatz-by-citation, renaming known results) are not present. The one load-bearing circularity is the selection-evaluation identity: Adam/tanh/max pooling is declared best because it has the highest validation accuracy in Table V, and the conclusion cites the same 0.9542 validation accuracy as the observed optimum. Since the validation set is used both to choose the configuration and to quantify its performance, the reported 'highest' accuracy is the selection criterion itself, not an unbiased or independent estimate. The paper also contains an internal inconsistency: Table II lists relu+Adam and gelu+Adam at 0.9667 validation accuracy under max pooling, and Section 3 states the highest validation accuracy is exhibited for relu, while the conclusion assigns the highest validation accuracy to tanh. That inconsistency is a correctness defect rather than a circularity, but it further undermines the central claim. Self-citations, such as reference [15] for the accuracy formula, are standard and not load-bearing. Overall, the per-configuration accuracy measurements retain independent empirical content, but the headline recommendation reduces by construction to the validation-selection step, giving a partial circularity score of 6.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper rests on an unstated and partially violated validation protocol, unspecified training hyperparameters, and single-run accuracy measurements. The central claim is therefore supported only by numbers that are themselves dependent on the missing choices.

free parameters (6)

learning rate = not reported
The optimizer learning rate is required for training but is never stated; all results depend on it.
number of epochs = not reported
Training length determines convergence and accuracy values; omitted.
batch size = not reported
Batch size affects gradient noise and optimizer behavior; omitted.
dropout rate = not reported
The architecture includes one dropout layer, but the dropout probability is not given.
convolutional filter counts and kernel sizes = not reported
Layer capacity is unspecified; only '2 convolutional layers' are mentioned.
train/validation split ratio = not reported
800 images are used, but the number of training vs validation samples is not stated.

assumptions (4)

domain assumption The labels in the crack image dataset [12] are correct and representative.
The paper does not audit the dataset; all accuracy measurements inherit this assumption.
domain assumption The validation set used for model selection was not also used to report final performance.
Section 3.3 selects Adam because it 'demonstrated the highest classification accuracy' on validation and then reports that same validation accuracy as the result, violating this assumption.
domain assumption The chosen small architecture (2 conv, 2 pool, dropout, dense) is an appropriate model class for this task.
No justification or comparison to other architectures is given.
domain assumption Accuracy from a single training run is a stable measure of performance.
No repeats or variance estimates are reported; this is an unstated background assumption for all comparisons.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Impact of Tuning Parameters in Deep Convolutional Neural Network Using a Crack Image Dataset." pith.science (2026). https://pith.science/paper/LSD6QEFA

@misc{pith2026250603184,
  author       = {Pith},
  title        = {Pith review of: Impact of Tuning Parameters in Deep Convolutional Neural Network Using a Crack Image Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSD6QEFA}},
  note         = {Machine review of arXiv:2506.03184}
}

read the original abstract

The performance of a classifier depends on the tuning of its parame ters. In this paper, we have experimented the impact of various tuning parameters on the performance of a deep convolutional neural network (DCNN). In the ex perimental evaluation, we have considered a DCNN classifier that consists of 2 convolutional layers (CL), 2 pooling layers (PL), 1 dropout, and a dense layer. To observe the impact of pooling, activation function, and optimizer tuning pa rameters, we utilized a crack image dataset having two classes: negative and pos itive. The experimental results demonstrate that with the maxpooling, the DCNN demonstrates its better performance for adam optimizer and tanh activation func tion.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 19 canonical work pages

[1]

SAPIENTIAE, INFORMATICA, 10 (1), 26–42, pp

Mureşan, H., Oltean, M.: Fruit recognition from images using deep learning ACTA UNIV. SAPIENTIAE, INFORMATICA, 10 (1), 26–42, pp. 1-22 (2018)

work page 2018
[2]

International conference on machine learning

Hayou, S., Arnaud, D., Judith, R.: On the impact of the activation function on deep neural networks training. International conference on machine learning. PMLR (2019)

work page 2019
[3]

On the Selection of Initialization and Activation Function for Deep Neural Networks

Hayou, S., Arnaud, D., Judith, R.: On the selection of initialization and activation fun ction for deep neural networks. arXiv preprint arXiv:1805.08266 (2018)

work page Pith review arXiv 2018
[4]

International Journal of Pattern Recognition and Artificial Intelligence 34.02 pp

Postalcıoğlu, S.: Performance analysis of different optimizers for deep learning-based image recognition. International Journal of Pattern Recognition and Artificial Intelligence 34.02 pp. 2051003 (2020)

work page 2020
[5]

International Conference on Intelligent Computing and Control Systems (ICCS), pp

Mehta, S., Paunwala, C., Vaidya, B.: CNN based Traffic Sign Classification using Adam Optimizer. International Conference on Intelligent Computing and Control Systems (ICCS), pp. 1293-1298, doi: 10.1109/ICCS45141.2019.9065537 (2019)

work page arXiv 2019
[6]

Proceedings of the 11th International Conference on Distributed Smart Cameras-ICDSC 2017 (2017)

Abdelouahab, K., Maxime, P., François, B.: PhD Forum: Why TanH can be a Hardware Friendly Activation Function for CNNs. Proceedings of the 11th International Conference on Distributed Smart Cameras-ICDSC 2017 (2017)

work page 2017
[7]

Postharvest Biology and Technology 166, pp

Momeny, M., et al.: Accurate classification of cherry fruit using deep CNN based on hybrid pooling approach. Postharvest Biology and Technology 166, pp. 111204 (2020)

work page 2020
[8]

Mathematical Problems in Engineering 2018 (2018)

Li, J ., et al.: Deep convolutional neural network based ECG classification system using information fusion a nd one -hot encoding techniques. Mathematical Problems in Engineering 2018 (2018)

work page 2018

Show all 21 references

[9]

et al.: A novel softplus linear unit for deep convolutional neural networks

Zhao, H., Liu, F., Li, L. et al.: A novel softplus linear unit for deep convolutional neural networks. Appl Intell 48, 1707–1720. https://doi.org/10.1007/s10489-017-1028-7 (2018)

2018 doi
[10]

IEEE Access, vol

Sun, Y., et al.: Convolutional Neural Network Based Models for Impr oving Super - Resolution Imaging. IEEE Access, vol. 7, pp. 43042 -43051, doi: 10.1109/ACCESS.2019.2908501 (2019)

2019
[11]

Devaraj, A.F.S., et al.: An Efficient Framework for Secure Image Archival and Retrieval System Using Multipl e Secret Share Creation Scheme, IEEE Access, vol. 8, pp. 144310 - 144320, doi: 10.1109/ACCESS.2020.3014346 (2020)

2020
[12]

Ozgenel, Ç .F., Gonenç Sorguç, A.: Performance Comparison of Pretrained Convolutional Neural Networks on Crack Detection in Buildings, ISARC Berlin (2018)

2018
[13]

T., Hasan, M

Nakib, M., Khan, R. T., Hasan, M. S., Uddin, J.: Crime Scene Prediction by Detecting Threatening Objects Using Convolutional Neural Network. International Conference on Computer, Communication, Chemical, Material and Electronic Engineering (IC4ME2), , pp. 1-4, doi: 10.1109/IC4...

2018
[14]

International Conference on Engineering and Technology (ICE T), pp

Albawi, S., Mohammed, T.A., Al-Zawi, S.: Understanding of a convolutional neural network. International Conference on Engineering and Technology (ICE T), pp. 1 -6, doi: 10.1109/ICEngTechnol.2017.8308186 (2017)

2017
[15]

IAES International Journal of Artificial Intelligence (IJ-AI), vol

Islam, M.M., Kashem, M.A, Uddin, J.: Fish survival prediction in an aquatic environment using random forest model. IAES International Journal of Artificial Intelligence (IJ-AI), vol. 10, no. 3, pp. 614-622, http://doi.org/10.11591/ijai.v10.i3.pp614-622 (2021)

2021 doi
[16]

2019 International Conference on Document Analysis and Recognition (ICDAR)

Christlein, V., et al.: Deep generalized max poo ling. 2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, (2019)

2019
[17]

International Journal of I maging Systems and Technology (2021)

Kakarla, J., et al.: Three‐class classification of brain magnetic resonance images using average‐pooling convolutional neural network. International Journal of I maging Systems and Technology (2021)

2021
[18]

Journal of machine learning research, vol

Duchi, J., Hazan , E., Singer , Y.: Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, vol. 12(Jul), pp. 2121-59 (2011)

2011
[19]

: Adadelta: an adaptive learning rate method

Zeiler M.D. : Adadelta: an adaptive learning rate method . arXiv preprint, vol. 12125701 http://arxiv.org/abs/1212.5701 (2012)

2012 arXiv
[20]

arXiv preprint , vol

Kingma, D.P., Ba , J.: Adam: A method for stochastic optimization. arXiv preprint , vol. 14126980 (2014)

2014
[21]

International Journal of Internet, Broadcasting and Communication, vol

Eom, S.H.: Developing Sentimental Analysis System Based on Various Optimizer. International Journal of Internet, Broadcasting and Communication, vol. 13, no. 1, pp. 100- 106 (2021) View publication stats

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

SAPIENTIAE, INFORMATICA, 10 (1), 26–42, pp

Mureşan, H., Oltean, M.: Fruit recognition from images using deep learning ACTA UNIV. SAPIENTIAE, INFORMATICA, 10 (1), 26–42, pp. 1-22 (2018)

work page 2018

[2] [2]

International conference on machine learning

Hayou, S., Arnaud, D., Judith, R.: On the impact of the activation function on deep neural networks training. International conference on machine learning. PMLR (2019)

work page 2019

[3] [3]

On the Selection of Initialization and Activation Function for Deep Neural Networks

Hayou, S., Arnaud, D., Judith, R.: On the selection of initialization and activation fun ction for deep neural networks. arXiv preprint arXiv:1805.08266 (2018)

work page Pith review arXiv 2018

[4] [4]

International Journal of Pattern Recognition and Artificial Intelligence 34.02 pp

Postalcıoğlu, S.: Performance analysis of different optimizers for deep learning-based image recognition. International Journal of Pattern Recognition and Artificial Intelligence 34.02 pp. 2051003 (2020)

work page 2020

[5] [5]

International Conference on Intelligent Computing and Control Systems (ICCS), pp

Mehta, S., Paunwala, C., Vaidya, B.: CNN based Traffic Sign Classification using Adam Optimizer. International Conference on Intelligent Computing and Control Systems (ICCS), pp. 1293-1298, doi: 10.1109/ICCS45141.2019.9065537 (2019)

work page arXiv 2019

[6] [6]

Proceedings of the 11th International Conference on Distributed Smart Cameras-ICDSC 2017 (2017)

Abdelouahab, K., Maxime, P., François, B.: PhD Forum: Why TanH can be a Hardware Friendly Activation Function for CNNs. Proceedings of the 11th International Conference on Distributed Smart Cameras-ICDSC 2017 (2017)

work page 2017

[7] [7]

Postharvest Biology and Technology 166, pp

Momeny, M., et al.: Accurate classification of cherry fruit using deep CNN based on hybrid pooling approach. Postharvest Biology and Technology 166, pp. 111204 (2020)

work page 2020

[8] [8]

Mathematical Problems in Engineering 2018 (2018)

Li, J ., et al.: Deep convolutional neural network based ECG classification system using information fusion a nd one -hot encoding techniques. Mathematical Problems in Engineering 2018 (2018)

work page 2018