Pith. sign in

REVIEW 4 major objections 8 minor 19 references

Planarian Neural Networks: Evolutionary Patterns from Basic Bilateria Shaping Modern Artificial Neural Network Architectures

T0 review · 4 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Periodically swapping early-layer weights between two parallel networks lowers error below an ordinary ensemble on CIFAR-10 and CIFAR-100.

desk verdict A clearly described ensemble variant with a small, possibly real gain, but the headline claim leans on a post-selection p-value and no code. read the letter →

arxiv 2501.04700 v1 pith:7KEFIFGE submitted 2025-01-08 cs.NE cs.AIcs.CVcs.LG

Ziyuan Huang , Mark Newman , Maria Vaida , Srikar Bellur , Roozbeh Sadeghian , Andrew Siu , Hui Wang , Kevin Huggins This is my paper · ORCID
classification cs.NEcs.AIcs.CVcs.LG
keywords planarianneuralnetworkscross-networkcommunicationStemBlockweightexchangepatiencegateensemblelearningwideresidualCIFARimageclassificationbiologicallyinspiredcomputing
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 coupling two neural networks the way a planarian couples its brain with its two nerve cords improves image classification accuracy. The two networks train in parallel, and every few epochs the weights of their early StemBlocks are exchanged through an artificial brain, with the exchange frequency set by a patience gate. The coupled pair, called a PNN, is compared with an ordinary soft-voting ensemble of the same two networks, and the paper reports lower average test error on both datasets: 5.81 versus 5.98 percent on CIFAR-10, and 20.67 versus 20.79 percent on CIFAR-100. The CIFAR-10 difference is reported as statistically significant; the CIFAR-100 difference is not. The aim is to show that cross-network communication during training, not extra neurons or a stronger final combination, is what drives the gain.

What carries the argument

The mechanism is a dual-network training loop with a StemBlock portal and a patience gate. A StemBlock is the shallow initial convolutional block that the PNN uses as the weight-exchange site; each global epoch, the two nerve-cord networks train for their own subepochs, and when the patience gate fires, the StemBlock weights of both networks are copied into an artificial brain and interchanged. The patience gate fires only after validation accuracy fails to improve for a set number of consecutive epochs, so swapping is adaptive rather than fixed-frequency. At test time the two networks' prediction probabilities are merged by soft voting. The comparison to a static ensemble isolates the swap: identical network members, identical total neuron count, with only the gated StemBlock exchange differing.

What would settle it

Retrain the paired networks with the patience gate fixed before any test-set evaluation, or chosen on a separate validation split; if PNN15 or PNN20 no longer beat their ensembles, the reported advantage was an artifact of test-set model selection. Alternatively, train the same two networks with identical schedules but no weight swapping and soft-vote them; if the no-swap control matches the PNN's test error, the exchange contributes nothing.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that periodic StemBlock weight exchange between two parallel networks improves test accuracy over both the individual networks and a conventional ensemble. PNN15, the best PNN on CIFAR-10, reaches 5.81 percent mean test error against 5.98 percent for the soft-voting ensemble, and the difference has a Mann-Whitney U p-value of 0.04. On CIFAR-100, PNN20 reaches 20.67 percent against 20.79 percent for the ensemble, a smaller difference with p = 0.64. Because the PNN and the ensemble contain the same total number of neurons, the paper attributes the improvement to the cross-communication mechanism rather than to capacity. The paper also reports that individual networks embedded in a PNN match or slightly beat their baselines, with the best single member improving from 7.44 to 7.27 percent mean error on CIFAR-10.

Load-bearing premise

The central claim stands on the premise that the small error advantage of PNN15 on CIFAR-10 and PNN20 on CIFAR-100 comes from the periodic StemBlock weight exchange itself, and not from the fact that the patience gate values were selected using the same test data that produced the highlighted error rates.

Editorial extensions

If this is right

  • A paired ResNet and WideResNet trained with gated StemBlock exchange can beat a soft-voting ensemble of the same pair without adding parameters.
  • Increasing the patience gate value from 5 to 15 on CIFAR-10 and from 10 to 20 on CIFAR-100 lowered the average test error in the tested range, pointing to less frequent exchange as beneficial.
  • The best coupled PNN member on CIFAR-10, PNN15's ResNet20, improves mean test error from 7.44 to 7.27 percent over its baseline, so the mechanism helps not just the combined system but the individual learners.
  • On CIFAR-100 the dual-network gain is directionally consistent with CIFAR-10 but not statistically significant with seven runs, so a larger seed count is the natural next check.

Reading between the lines

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

  • A natural testable extension is to swap weights at different depths or to replace discrete swaps by an exponential moving average; if the gain depends specifically on early-layer synchronization, depth and averaging schedule should matter.
  • The monotone trend with patience values suggests mapping the full patience curve beyond 15 on CIFAR-10 and 20 on CIFAR-100; an optimum would show where the reported trend turns around.
  • If the benefit is real, a no-swap control with identical training schedules but ordinary soft voting should reproduce only the ensemble baseline; if that control equals the PNN's accuracy, the exchange itself is not the active ingredient.
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 / 8 minor

Summary. The paper proposes Planarian Neural Networks (PNNs), a dual-network architecture inspired by the planarian nervous system, in which two ResNet variants (a deep network and a wide network) are trained in parallel and periodically exchange StemBlock weights through an artificial brain, with the swap frequency controlled by a patience gate. The authors evaluate PNNs against single-network baselines and a soft-voting ensemble baseline on CIFAR-10 (five seeds) and CIFAR-100 (seven seeds), reporting mean error rates and nonparametric significance tests. They claim that cross-network communication improves test accuracy, with PNN15 on CIFAR-10 and PNN20 on CIFAR-100 outperforming all other models, and they attribute the improvement to the weight-exchange mechanism rather than to increased neuron count.

Significance. If the central claim were established, the paper would offer a modest but interesting biologically motivated training mechanism: periodically swapping early-layer weights between two networks during training, gated by validation-stagnation, could yield a small accuracy gain over a matched ensemble without increasing parameter count. The manuscript has genuine strengths: it provides per-seed result tables for all models, uses nonparametric tests appropriate for small samples, matches the parameter counts between ensemble and PNN, and reports both significant and non-significant results rather than only favorable ones. However, the headline claim is not currently supported because the main favorable result is a single borderline p-value obtained after selecting the best patience value on the test set, and the CIFAR-100 result is not statistically significant. The paper also lacks an ablation that isolates the weight-swap mechanism from the altered training schedule (subepochs and patience gating).

major comments (4)
  1. [IV.A and V.A (CIFAR-10 dual-network U test)] The U test comparing PNN15 with Ensemble (p = 0.04) is not a valid test of a pre-specified hypothesis. Section IV.A states that 'after the PNN models were fine-tuned, we presented the PNN5, PNN10, and PNN15 models,' and the same test set is then used both to select the best patience value (PNN15) and to compute the highlighted error rate. Because three patience values were compared on the same test data, the reported p-value is a post-selection artifact unless the authors demonstrate that the patience value was chosen on the validation split before any test evaluation; the manuscript provides no such protocol or code. The paper should report a held-out evaluation of a pre-selected architecture, or apply a multiplicity correction, or present the comparison of all three patience values as exploratory.
  2. [V.B (CIFAR-100 dual-network tests)] The conclusion in Section VI that PNN20 'outperformed the other models' on CIFAR-100 is not supported by the statistical tests reported in Section V.B: the H test across the four dual networks gives p = 0.54 and the U test for PNN20 versus Ensemble gives p = 0.64, both far above alpha = 0.05. The authors themselves note that the improvements were not statistically significant, yet the Conclusions section still asserts outperformance. This contradiction between the stated results and the concluding claim must be resolved, for example by reframing the CIFAR-100 result as a non-significant trend or by providing additional evidence.
  3. [III.A and IV (ablation isolating the swap mechanism)] The paper's core attribution claim, stated in Section III, is that 'the improvements in test accuracy were induced by the inclusion of cross-network communication rather than an increase in the number of neurons.' The comparison between PNN and Ensemble holds neuron count constant, but the two procedures differ in multiple ways: the PNN trains each network in alternating subepochs, uses a patience gate that changes the effective learning-rate schedule, and swaps StemBlock weights. Without an ablation that trains two networks in the same subepoch schedule with the same patience gating but without the weight swap (or, conversely, with weight swap but no soft voting), it is impossible to attribute the observed difference specifically to cross-network communication. Such an ablation is load-bearing for the paper's central claim and should be added.
  4. [IV (fine-tuning and test-set reuse)] The manuscript does not specify a validation-based model-selection procedure. Section II.A mentions that 10% of the training images are used as validation, but Section IV.A's phrase 'after the PNN models were fine-tuned' suggests that the patience values (5, 10, 15 on CIFAR-10; 10, 15, 20 on CIFAR-100) were tuned using the test error. If the patience value was chosen on validation, the paper must state this explicitly, describe the selection rule, and account for the selection in the statistical analysis. If it was chosen on test error, then the reported significance levels overstate the evidence and a held-out evaluation is required.
minor comments (8)
  1. [Index Terms] The index term 'plana1rian neural networks' contains a typo; it should read 'planarian neural networks.'
  2. [IV.B (CIFAR-100 model name)] In Section IV.B, the text refers to 'PNN (ResNet164 + WideResNet108)' and later 'WideResNet108,' but the models described and listed in Tables V and VI are WideResNet110; the naming is inconsistent and should be corrected.
  3. [References] Reference [19] is cited as 'University of Utah, 2023, Mann-Whitney U test' but the URL points to Wikipedia; the reference should be corrected to the actual source.
  4. [Section II.B] The phrase 'CIFAR0-10' in Section II.B appears to be a typo for 'CIFAR-10.'
  5. [Section III] The citation 'Eguyen et al.' in Section III appears to be a misspelling of 'Nguyen et al.' (reference [11]); please correct the in-text name.
  6. [Section II.B and IV.A] The term 'patience gate' is sometimes written as 'patient gate' (e.g., Section IV.B: 'fine-tuned using patient gate values'); unify the terminology to 'patience gate' throughout.
  7. [Algorithm 2 (Table II)] Algorithm 2 resets patience_level to 0 after a swap in the pseudocode, but the surrounding text states that 'each weight swap reset the patience level to 1.' This discrepancy between the pseudocode and the prose should be resolved.
  8. [General] The paper does not provide a statement on code availability or reproducibility details beyond the seed-selection description; sharing the training and evaluation code, including the exact weight-swap and validation logic, would strengthen the reproducibility of the results.

Circularity Check

1 steps flagged · score 6.0 of 10

The 'best' PNN is selected from the same test-error tables that are then used to claim statistical superiority, so the headline outperformance claim is partly a post-selection artifact.

  1. fitted input called prediction [Section IV.A (Dual Network Results, Tables IV and VI) and Section V.A (Dual Network Architecture); echoed in Section VI Conclusions]
    "After the PNN models were fine-tuned, we presented the PNN5, PNN10, and PNN15 models ... with fine-tuned patience gate values of 5, 10, and 15, respectively. The five-run average errors ... were 5.98, 6.01, 5.94, and 5.81, respectively. ... Among the four dual networks listed in TABLE IV, PNN15 ... exhibited the best five-run average error of 5.81 compared with the baseline model's 5.98. ... Next, we evaluated the significance of the differences between the best fine-tuned model, PNN15, and the baseline Ensemble model on CIFAR-10 using the U test."

    The patience gate value is a hyperparameter described as 'fine-tuned,' but the paper never states that the choice among 5, 10, and 15 was made on validation data before any test evaluation. The test-error table is used both to identify PNN15 as the best model and then to run the U test against the Ensemble baseline. The p=0.04 therefore tests a hypothesis that was generated from the very data used to compute it, after best-of-three selection with no multiplicity correction. The claimed 'prediction' that cross-communication-based PNNs outperform others is, in this part, a description of the selected maximum rather than an independent, out-of-sample confirmation.

full rationale

This paper contains no mathematical derivation chain, so there is no equation-level circularity of the self-definitional type. The architecture comparison itself is controlled: the PNN and Ensemble have the same aggregate neuron count (530,932 on CIFAR-10; 3,364,712 on CIFAR-100), so the claim that cross-network communication, rather than extra neurons, drives the difference is a legitimate experimental contrast. However, the central statistical claim of superiority is partially circular. The patience gate values are 'fine-tuned' and the best PNN (PNN15 on CIFAR-10, PNN20 on CIFAR-100) is identified from the same test-error tables later used for the significance tests. No validation-only selection protocol or multiplicity adjustment is reported. On CIFAR-10 the only positive result is a single p=0.04 U test performed after this post-selection; on CIFAR-100 the corresponding U test is not significant (p=0.64), and the H test across all four dual networks is also not significant (p=0.54). Thus the headline that the PNN models 'outperformed the other models' reduces in part to choosing the best of several patience values on the test set, not to a pre-specified prediction from the biological analogy. This is a partial circularity, not a full one, because the raw mean differences and the controlled neuron counts remain independent empirical content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The central empirical claim rests mainly on the training setup and evaluation choices: the patience gate is a tuned hyperparameter, weight exchange is assumed to be the active ingredient, and the statistical tests assume independence across runs. There is no mathematical derivation in the paper.

free parameters (3)
  • patience gate max_patience (n) = 15 for CIFAR-10 dual; 20 for CIFAR-100 dual; 5 and 10 also tested
    Controls frequency of StemBlock weight swaps; values were selected by comparing test errors across PNN5/10/15 and PNN10/15/20, so the highlighted result is partially fit to the test set.
  • subepoch count per nerve cord = not stated
    Algorithm 1 loops over sub_epochs_1 and sub_epochs_2, but the paper never gives these values, so the effective training budget per nerve cord is an unspecified hyperparameter.
  • StemBlock weight_swap_function = not specified
    Algorithm 1 calls a weight_swap_function without defining whether weights are averaged, copied, or mixed; this is a design choice that could materially affect results.
assumptions (4)
  • domain assumption StemBlock weight exchange is a faithful and sufficient mechanism for cross-network communication.
    Section III.A chooses StemBlocks as the only weight-exchange portal; the paper never tests whether this specific exchange is what produces the accuracy difference.
  • ad hoc to paper Validation accuracy stagnation is a valid signal for scheduling weight swaps.
    The patience gate in Algorithm 2 is an arbitrary heuristic; the paper provides no evidence that this schedule is better than fixed-interval or random swapping.
  • standard math The independence assumption of the nonparametric tests holds across runs.
    Section V states each model generated its results independently, but all runs share the same codebase, data pipeline, and hyperparameters, so independence is approximate.
  • domain assumption Training budgets and data augmentation are matched across baseline and PNN models.
    The paper asserts identical hyperparameters, but subepoch counts for each nerve cord are not reported in Algorithm 1, so the effective training budget per network is not verified.
invented entities (2)
  • Artificial brain (AANS) and artificial nerve cords (ABNS)
    purpose: Conceptual labels for the central weight-exchange module and the parallel network trunks
    These are descriptive metaphors for existing PyTorch modules; they generate no falsifiable biological prediction outside the architecture itself.
  • Patience gate
    purpose: A counter that triggers weight swaps after validation accuracy stalls
    This is a hyperparameterized heuristic, not an entity with independent evidence; its values are tuned on the benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Planarian Neural Networks: Evolutionary Patterns from Basic Bilateria Shaping Modern Artificial Neural Network Architectures." pith.science (2026). https://pith.science/paper/7KEFIFGE

@misc{pith2026250104700,
  author       = {Pith},
  title        = {Pith review of: Planarian Neural Networks: Evolutionary Patterns from Basic Bilateria Shaping Modern Artificial Neural Network Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7KEFIFGE}},
  note         = {Machine review of arXiv:2501.04700}
}
read the original abstract

This study examined the viability of enhancing the prediction accuracy of artificial neural networks (ANNs) in image classification tasks by developing ANNs with evolution patterns similar to those of biological neural networks. ResNet is a widely used family of neural networks with both deep and wide variants; therefore, it was selected as the base model for our investigation. The aim of this study is to improve the image classification performance of ANNs via a novel approach inspired by the biological nervous system architecture of planarians, which comprises a brain and two nerve cords. We believe that the unique neural architecture of planarians offers valuable insights into the performance enhancement of ANNs. The proposed planarian neural architecture-based neural network was evaluated on the CIFAR-10 and CIFAR-100 datasets. Our results indicate that the proposed method exhibits higher prediction accuracy than the baseline neural network models in image classification tasks. These findings demonstrate the significant potential of biologically inspired neural network architectures in improving the performance of ANNs in a wide range of applications.

Figures

Figures reproduced from arXiv: 2501.04700 by the authors.

Figure 2
Figure 2. StemBlock-based weight exchange in PNN. In the proposed PNN, the two nerve cords were trained over subepochs, and their learned knowledge was balanced over global epochs in the artificial brain (as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Ensemble (ResNet + WideResNet). PNN (ResNet20 + WideResNet14): PNN (ResNet20 + WideResNet14) was implemented via ensemble learning ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Error rates of the models on CIFAR-10. Single Network Architecture: The single-network PNN models exhibited certain advantages over the baseline models ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Error rates of single-network models on CIFAR-10. The H-test hypotheses on single network models during experiments on CIFAR-10 are listed below: 𝐻0 : 𝑥̃1 = 𝑥̃2 = 𝑥̃3 = 𝑥̃4 = 𝑥̃5 = 𝑥̃6 = 𝑥̃7 = 𝑥̃8 𝐻1 : 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 𝑚𝑒𝑑𝑖𝑎𝑛 𝑣𝑎𝑙𝑢𝑒 𝑖𝑠 𝑑𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑡ℎ𝑜𝑠𝑒 𝑜𝑓 𝑜𝑡ℎ…
Figure 6
Figure 6. Figure 6: Error Rates of Dual Network Models on CIFAR-10. The H test hypotheses for dual network models on CIFAR￾10 are listed below: 𝐻0 : 𝑥̃1 = 𝑥̃2 = 𝑥̃3 = 𝑥̃4 𝐻1 : 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 𝑚𝑒𝑑𝑖𝑎𝑛 𝑖𝑠 𝑑𝑖𝑓𝑓𝑒𝑟𝑛𝑡 𝑓𝑟𝑜𝑚 𝑡ℎ𝑜𝑠𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑜𝑡ℎ𝑒𝑟 𝑚𝑜𝑑𝑒𝑙𝑠 Based on (1) and using SciPy’s statisti…
Figure 8
Figure 8. Figure 8: Error rates of single-network models on CIFAR-100. In the H test on single network models on CIFAR-100, we assumed the following: 𝐻0 : 𝑥̃1 = 𝑥̃2 = 𝑥̃3 = 𝑥̃4 = 𝑥̃5 = 𝑥̃6 = 𝑥̃7 = 𝑥̃8 𝐻1 : 𝑎𝑡 𝑙𝑒𝑎𝑠𝑡 𝑜𝑛𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 𝑚𝑒𝑑𝑖𝑎𝑛 𝑖𝑠 𝑑𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑡ℎ𝑜𝑠𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑜𝑡ℎ𝑒𝑟 𝑚𝑜𝑑𝑒𝑙𝑠 The null…
Figure 9
Figure 9. Figure 9: Error Rates of Dual Network Experiments using CIFAR￾100. An H test was performed to gauge the statistical significance of the differences between the Ensemble (ResNet164+WideResNet110), PNN10 (ResNet164+WideResNet110), PNN15 (ResNet164+WideResNet110), and PNN20 (ResNet…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 11 canonical work pages

  1. [1]

    González and C

    G. González and C. L. Evans, “Biomedical Image processing with Containers and Deep Learning: An Automated Analysis Pipeline: Data architecture, artificial intelligence, automated processing, containerization, and clusters orchestration ease the transit ion from data acquisition to insights in medium -to-large datasets,” BioEssays, vol. 41, no. 6, p. e1900...

    work page doi:10.1002/bies.201900004 2019

  2. [2]

    Land cover changes analysis based on deep machine learning technique,

    N. N. Kussul et al., “Land cover changes analysis based on deep machine learning technique,” J. Autom. Inf. Sci., vol. 48, no. 5, pp. 42–54, 2016, doi: 10.1615/JAutomatInfScien.v48.i5.40

    work page doi:10.1615/jautomatinfscien.v48.i5.40 2016

  3. [3]

    Railway track fastener defect detection based on image processing and deep learning techniques: A comparative study,

    X. Wei et al. , “Railway track fastener defect detection based on image processing and deep learning techniques: A comparative study,” Eng. Appl. Artif. Intell. , vol. 80, pp. 66 –81, 2019, doi: 10.1016/j.engappai.2019.01.008

    work page doi:10.1016/j.engappai.2019.01.008 2019

  4. [4]

    Deep residual learning for image recognition

    K. He et al., “Deep residual learning for image recognition.” IEEE Conf. on Comput. Vis. and Pattern Recognit., (CVPR), Las Vegas, NV, USA, 2016, 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

  5. [5]

    ‘Identity mappings in deep residual networks,’ Vis. ECCV,

    K. He et al., “‘Identity mappings in deep residual networks,’ Vis. ECCV,” in Proc. Part IV: 14th Eur. Conf. on comput. vis., Amsterdam, The Netherlands, October 11–14, 2016, pp. 630–645

    work page 2016

  6. [6]

    ImageNet classification with deep convolutional neural networks,

    A. Krizhevsky et al. , “ImageNet classification with deep convolutional neural networks,” Commun. ACM, vol. 60, no. 6, pp. 84 –90, 2017, doi: 10.1145/3065386

    doi:10.1145/3065386 2017

  7. [7]

    Deeply-supervised nets,

    C.-Y. Lee et al., “Deeply-supervised nets,” Artif. Intell. Stat, pp. 562–570, 2015

    work page 2015

  8. [8]

    Krizhevsky, Learning Multiple Layers of Features from Tiny Images , 2009, pp

    A. Krizhevsky, Learning Multiple Layers of Features from Tiny Images , 2009, pp. 32–33

    work page 2009

Show all 19 references
  1. [9]

    From nerve net to nerve ring, nerve cord and brain — Evolution of the nervous system,

    D. Arendt et al., “From nerve net to nerve ring, nerve cord and brain — Evolution of the nervous system,” Nat. Rev. Neurosci., vol. 17, no. 1, pp. 61–72, 2016, doi: 10.1038/nrn.2015.15

    2016 doi

  2. [10]

    Pros and cons of different sampling techniques,

    G. Sharma, “Pros and cons of different sampling techniques,” Int. J. Appl. Res., vol. 3, no. 7, pp. 749–752, 2017

    2017

  3. [11]

    Do wide and deep networks learn the same things? Uncovering how neural network representations vary with width and depth,

    T. Nguyen et al., 2020, “Do wide and deep networks learn the same things? Uncovering how neural network representations vary with width and depth,” ArXiv Preprint ArXiv:2010.15327

    2020 arXiv

  4. [12]

    Deep learning scaling is predictable, empirically,

    J. Hestness et al., 2017, “Deep learning scaling is predictable, empirically,” ArXiv Preprint ArXiv:1712.00409

    2017 arXiv

  5. [13]

    Scaling laws for neural language models,

    J. Kaplan et al., 2020, “Scaling laws for neural language models,” ArXiv Preprint ArXiv:2001.08361

    2020 arXiv

  6. [14]

    In search of the real inductive bias: On the role of implicit regularization in deep learning,

    B. Neyshabur et al., 2014, “In search of the real inductive bias: On the role of implicit regularization in deep learning,” ArXiv Preprint ArXiv:1412.6614

    2014 arXiv

  7. [15]

    Towards understanding the role of over - parametrization in generalization of neural networks,

    B. Neyshabur et al. , 2018, “Towards understanding the role of over - parametrization in generalization of neural networks,” ArXiv Preprint ArXiv:1805.12076

    2018 arXiv

  8. [16]

    Pelee: A real -time object detection system on mobile devices,

    R. J. Wang et al., “Pelee: A real -time object detection system on mobile devices,” Adv. Neural Inf. Process. Syst., vol. 31, 2018

    2018

  9. [17]

    Wide residual networks,

    S. Zagoruyko and N. Komodakis, “Wide residual networks,” in Proceedings of the British Machine Vision Conference 2016 . British Machine Vision Association, 2016, pp. 87.1–87.12, doi: 10.5244/C.30.87

    2016 doi

  10. [18]

    Getting started with the Kruskal-Wallis test

    S. Lomuscio, 2021, “Getting started with the Kruskal-Wallis test.” Available at: https://data.library.virginia.edu/getting -started-with-the- kruskal-wallis-test/

    2021

  11. [19]

    Mann -Whitney U test

    University of Utah, 2023, “Mann -Whitney U test.” Available at: https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.