Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Weight magnitude alone does not determine neuron importance in image classifiers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:09 UTC pith:U45WXJQB

load-bearing objection The experiments operate on scalar weights, not CNN neurons, so the headline claim about 'high-weight neurons' doesn't follow; and the '45–80% accuracy degradation' numbers are a normalized effect index, not measured accuracy drops. the 3 major comments →

arxiv 2607.25529 v1 pith:U45WXJQB submitted 2026-07-28 cs.AI

Are the High-weight Neurons the Important Ones in Image Classification Neural Networks?

Qitao Chen , Dongfu Yin , F. Richard Yu This is my paper
classification cs.AI
keywords neuron importanceweight pruningweight-importance relationshipablation studyperturbation analysisimage classificationCIFAR-10Mini-ImageNet
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tests the common assumption that neurons with the largest absolute weights are the most important units in an image classification network. Using three operational probes—overlap between high-weight and accuracy-critical neurons, perturbation sensitivity, and ablation followed by retraining—on CIFAR-10 and Mini-ImageNet, it finds that top-10% weight neurons overlap with truly impactful neurons only about a quarter of the time. Some top-decile weights are nearly inert, while some low-weight neurons degrade accuracy as much as mid-range ones when disturbed. The authors conclude that the weight–importance relationship is nonlinear and that applications such as pruning, encryption, and backdoor defense should not treat high magnitude as synonymous with high importance.

Core claim

On the paper's own terms, 'neuron' means a single scalar weight: after training, the model's weights are flattened, absolute-valued, sorted descending, and cut into 10% intervals. Three probes are then run. Per-weight ablation finds which weights, individually removed, cost the most test accuracy; overlap between that set and the largest-weight set peaks around 25% in the top decile and is worse elsewhere. Perturbation tests (adding/subtracting a constant or multiplying/dividing by a factor) show the top 10% interval is far more fragile than random selection—45–80% accuracy loss versus 3–7%—but roughly a third of top-decile weights barely matter, and some low-weight intervals lose 10–17% und

What carries the argument

Central machinery: the sorted absolute-weight interval decomposition. Every scalar weight in the trained network is flattened into a vector, absolute-valued, sorted descending, and split into non-overlapping intervals (deciles, with the top decile subdivided into finer bands). Three operations are then applied interval-by-interval: single-weight ablation for overlap, additive/multiplicative perturbation for robustness, and masked zeroing with retraining for recoverability. Together these translate 'importance' from a magnitude claim into a behavioural claim—what actually happens to accuracy when the unit is removed, disturbed, or blocked from learning.

Load-bearing premise

The load-bearing premise is that a single scalar weight in a flattened, absolute-sorted vector counts as a 'neuron'; if the functional unit is a filter or channel made of many weights, the measured nonlinearity applies to scalar weights, not to neurons as usually understood.

What would settle it

Take a fixed ResNet-18 trained on CIFAR-10 and compute the overlap between the 10% largest absolute scalar weights and the 10% of weights whose individual zeroing costs the most test accuracy; the paper's claim predicts this overlap stays near 25% and not, say, above 60%. Independently, if one groups weights by convolutional filter and ranks filters by summed absolute magnitude, high-magnitude filters may overlap with accuracy-critical filters much more strongly—which would show the 'nonlinearity' is an artifact of splitting weights inside a filter rather than a property of neurons.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A pruning rule that removes the smallest absolute weights may discard functionally significant units, because some low-weight intervals affect accuracy as much as mid-range high-weight ones.
  • High weight alone does not justify cryptographic or defensive priority: a substantial subset of the top decile shows minimal accuracy impact under perturbation, so protecting all high-weight scalars would waste effort.
  • The top decile is not a single class: removing all of it prevents recovery after retraining, but finer subdivision shows many subintervals are recoverable, so irreplaceability is concentrated in a small extreme subset of weights.
  • Applications that rely on weight-importance equivalence—pruning, backdoor defense, interpretability—should incorporate behavioural measures rather than magnitude alone.
  • The same qualitative pattern appeared in a financial OHLC image generalization experiment, so the phenomenon is not unique to natural-image benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: because each 'neuron' is a scalar weight rather than a convolutional filter, the result is best read as a statement about weight-level sensitivity. Repeating the three probes at filter or channel granularity would show how much of the nonlinearity survives when the unit is a real neuron.
  • Beyond the paper: the perturbation pattern suggests a testable 'criticality spectrum'—extreme high weights are individually fragile, mid-range weights are redundant, and very low weights matter only in groups. A direct next experiment is to ablate random groups of low-weight neurons of growing size and check whether their collective contribution is linear or sharply thresholded.
  • Beyond the paper: the results imply a cheap behavioural importance score—marginal accuracy loss from one-shot ablation or small perturbation—that could replace magnitude ranking in pruning and be benchmarked against magnitude pruning at matched sparsity.
  • Beyond the paper: the low-weight 'ecosystem' idea implies interpretability and mechanistic analyses should not focus only on high-activation or high-magnitude units; distributed low-weight codes may matter for generalization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper asks whether high-magnitude weights identify functionally important neurons in image-classification CNNs. The method flattens all trainable weight matrices, sorts scalar weights by absolute value, partitions them into 10% intervals, and conducts three analyses on CIFAR-10 and Mini-ImageNet with LeNet-5, ResNet-18, and ResNet-50: (i) overlap between the top-n weight set and the top-n accuracy-impact set obtained by one-at-a-time ablation; (ii) perturbation of weights in each interval via additive/proportional operations; and (iii) ablation of top-n weights with gradient-masked retraining. The authors report that top-10% weight intervals overlap with accuracy-impacting sets by at most about 25%, that high-weight intervals dominate their normalized perturbation-effect measure while some low-weight intervals show non-negligible effects, and that top-10% ablation cannot be recovered by retraining whereas top-0.1% ablation can. They conclude that weight magnitude alone does not determine neuron importance and that low-weight neurons contribute significantly.

Significance. If the central claim were established at the level of neurons, the paper would be a useful challenge to magnitude-based pruning and would inform interpretability and robustness work. The paper has concrete strengths: a clear experimental protocol, multiple architectures and datasets, and a clean ablation-retraining design with gradient masking. Read as a study of individual scalar weights, the results are a plausible caution that fine-grained magnitude ranking does not align with causal impact. However, the paper's central claim is framed at the neuron level, while the experiments never manipulate a neuron (filter/channel); they manipulate single connection weights. The quantitative claims in the abstract and Section 4.3 also confuse a normalized effect share with an absolute accuracy drop. The first issue is load-bearing and cannot be repaired by editing; it requires rerunning the experiments at unit/filter granularity. The paper's contribution in its current form is therefore substantially narrower than claimed.

major comments (3)
  1. [§3.1–§3.2, Fig. 1] Throughout the method, 'neuron' is operationalized as a scalar entry in a flattened weight tensor. In the CNNs used here (LeNet-5, ResNet-18/50), the elementary functional unit is the filter/channel, not an individual connection weight. Setting one element of a 3×3 convolutional kernel to zero does not ablate a neuron; it partially corrupts one connection, and the effect is diluted across spatial positions and filters. Consequently, the overlap sets S_w and S_a in Eq. (1), the perturbation intervals in Tables 1–2, and the ablation-retraining in Algorithm 1 all concern individual scalar weights, not neurons. A filter with a large norm can contribute many top-10% scalar weights, so the top-10% scalar-weight set is not a top-10% neuron set. Section 4.5 explicitly defers module/conv-layer analysis to future work, confirming that the present experiments do not test the paper's stated unit of
  2. [§3.3, Eqs. (5)–(6), Tables 1–2] Eq. (5) defines E(p_i) = max(p)+ε−p_i and Eq. (6) normalizes by the sum of E over intervals. The numbers in Tables 1–2 therefore are shares of the total perturbation effect; in each row the interval entries plus the Random column sum to approximately 100%. They are not measured accuracy degradations. Yet the abstract states '45–80% accuracy degradation' and §4.3 states the 'Divide' operation 'caused 79.33% accuracy degradation' in ResNet-18 on CIFAR-10. Those sentences confuse a normalized index with an absolute drop. The qualitative finding that the top interval has a disproportionate share may survive renormalization, but the quantitative claims as written are unsupported. Also, the Mini-ImageNet 'minus' row in Table 1 lists 21.37 for three adjacent intervals, which appears to be a data-entry error rather than a real effect.
  3. [§4.2, Eq. (1)] The overlap analysis is the first pillar of the paper's conclusion, but Eq. (1) is not connected to the reported numbers in a reproducible way. The text describes the red line in Fig. 3 as 'the proportion of neurons that are both high-weight and important within the same interval,' while Eq. (1) defines a ratio in terms of differences of cumulative intersections divided by interval width. The symbols S_{a,m}, S_{w,m}, and N_m are not defined precisely enough to determine whether the denominator is interval size, the number of ablated neurons, or the size of the TOP-n sets. The reader cannot reconstruct the overlap percentages from the method as written.
minor comments (5)
  1. [§4.2–§4.4, figure numbering] The in-text figure references are inconsistent with the captions. §4.2 refers to Figure 2A/2C for overlap results, but Figure 2 is captioned 'Ablation and perturbation Strategies'; §4.3 refers to Figure 3A for perturbation results, but Figure 3 is captioned as overlap results; §4.4 refers to Figure 4A for ablation-retraining, but Figure 4 is the perturbation figure and Figure 5 shows the removal/retraining results. Please renumber or correct the cross-references.
  2. [Algorithm 1, line 5] With 1-based indexing, 'sorted indices[1:n]' omits the largest-magnitude index. This should be 'sorted indices[0:n]' (or '[:n]') if the intent is to ablate the top-n neurons/weights.
  3. [Throughout] There are numerous typos and stylistic errors: 'T able 2' in the caption, 'Multipy' for 'Multiply', 'CIF AR-10' with an inserted space, 'Neuroncomputing' for 'Neurocomputing', and inconsistent use of curved quotes/apostrophes. The paper would benefit from a careful proofreading pass.
  4. [Introduction, reference [12]] The text attributes 'DeepMind' to reference [12], but reference [12] is Morcos et al., 'On the importance of single directions for generalization' (arXiv:1803.06959), which is not a DeepMind paper. The attribution should be corrected or removed.
  5. [Appendix A] The financial-data 'generalization experiment' is disconnected from the rest of the paper: the OHLC chart dataset is not a natural-image classification task, no training protocol is given for the LeNet model on this dataset, and the appendix is never cited in the abstract, introduction, or conclusion. It appears to be a separate study and should either be integrated with full experimental details or removed.

Circularity Check

0 steps flagged

No load-bearing circularity: the experiments are direct measurements on standard benchmarks; the only substantive issue is the paper's equation of a scalar weight with a 'neuron,' which is a construct-validity concern, not a circular derivation.

full rationale

The paper's central claim is an empirical observation about measured quantities. The three experiments (overlap, perturbation, and ablation-retraining) are all direct measurements on standard models and datasets (CIFAR-10, Mini-ImageNet, ResNet-18/50, LeNet). No parameter is fitted and then renamed as a prediction; no load-bearing self-citation appears in the reference list; and no uniqueness theorem or prior result by the same authors is invoked to force the conclusion. Equations (5)-(6) and (8) define normalized 'effect' and 'irreplaceability' indices relative to the best-performing interval, but these are indexing choices for reporting measured accuracy values, not assumptions of the conclusion. The claimed nonlinearity (e.g., top-10% overlap only ~25%, low-weight intervals showing 10-17% degradation) follows from the measured data rather than from the definitions. The paper does, however, call each scalar weight a 'neuron' (e.g., 'flatten the pretrained neuron weight matrices... to a one-dimensional vector' and 'setting the weights of individual neurons to zero'). That is a definitional/construct-validity choice: the experiments concern individual scalar weights, not convolutional filters/channels, so the generalization to 'neurons' is not automatically justified. But this is not circular reasoning—it is a potential mismatch between the operational definition and the broader term. Since no step in the derivation reduces to its own input by construction, the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim is empirical, so there are no fitted physical constants; however, the quantitative conclusions are contingent on many hand-chosen analysis choices (bins, perturbation sizes, retraining budget) and on the premise that scalar weights can stand in for neurons.

free parameters (4)
  • Weight-interval boundaries = 10%, 20%, ..., 100%; top-10% subdivided into 0.01-0.1% etc.
    Hand-chosen partitions; all reported overlap/perturbation/retraining percentages are defined relative to these bins.
  • Perturbation magnitudes = e.g., fixed 0.04, 5%, proportional 0.5
    Hand-chosen per experiment; no sensitivity analysis across magnitudes, and the exact values used for Tables 1-2 are not specified.
  • Epsilon in Eqs. 5 and 8 = arbitrary small positive
    Added to avoid a zero baseline; shifts the relative effect scores, though not the ordering of intervals.
  • Retraining budget = 20 epochs, LR 0.0001
    Fixed for all ablation-retraining runs; no evidence that 20 epochs is sufficient for recovery or that results are stable across seeds.
axioms (5)
  • ad hoc to paper Each scalar connection weight in the flattened weight matrices is treated as a 'neuron' whose ablation/perturbation measures neuron importance.
    Flattening in §3.2/Figure 1A equates a weight with a neuron, which is not the standard CNN definition of a neuron/filter.
  • domain assumption Individual ablation (setting one weight to zero at a time) identifies the accuracy-impacting neurons.
    Defines the ground-truth 'important' set in §3.2; ignores interactions and may be noisy per-parameter.
  • domain assumption Accuracy after 20 epochs of masked retraining measures neuron 'irreplaceability'.
    Algorithm 1; no convergence criterion; other weights may compensate over more epochs.
  • domain assumption The best-performing interval after perturbation is a valid zero-effect baseline.
    Eqs. 5-6 normalize perturbation effects relative to max-interval accuracy; this is the index later read as absolute degradation.
  • domain assumption All layers, biases, and model types can be pooled into one sorted list.
    The method flattens all trainable parameters into a single list (§3.1), ignoring layer structure and positional effects, though the Discussion admits position matters.

pith-pipeline@v1.3.0-alltime-deepseek · 13654 in / 16814 out tokens · 156800 ms · 2026-08-01T02:09:18.262810+00:00 · methodology

0 comments
read the original abstract

As neural network models for image classification advance, neurons play critical roles in pruning, backdoor defense, and interpretability. Yet existing work lacks clarity on the weight-importance relationship. We address this with a neuron importance assessment method using three experiments: quantifying overlap between high-weight and accuracy-impacting neurons, analyzing high-weight neuron perturbation effects, and testing post-retraining accuracy after high-weight neuron ablation. Experiments on CIFAR-10 and Mini-ImageNet reveal key patterns. Overlap analysis shows top 10\% high-weight neurons overlap with important ones by only about 25\% at maximum, dropping further in subsequent intervals. Perturbation tests find top 10\% high-weight neurons cause 45-80\% accuracy degradation under certain operations compared to 3-7\% for random perturbations, but a third of them show minimal impact. Ablation-retraining results show removing top 10\% high-weight neurons leaves accuracy 10-20\% below baseline with no recovery, while ablating top 0.1\% allows near-full recovery. Notably, some low-weight intervals show 10-17\% degradation when perturbed, comparable to mid-range high-weight neurons. These results confirm not all high-weight neurons are important: their importance is nonlinear. Low-weight neurons also contribute significantly. This challenges weight-importance equivalence, offering refined neuron role insights. It supports applications like encryption prioritizing critical high-weight neurons and pruning removing non-critical ones, advancing neural network analysis.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 13 linked inside Pith

  1. [1]

    Learning efficient convolutional networks through network slimming

    Liu Z, Li J, Shen Z, et al. Learning efficient convolutional networks through network slimming. Proceedings of the IEEE international conference on computer vision, 2017: 2736-2744

    2017

  2. [2]

    Lightweight Neural Networks for Adversarial Defense: A Novel NTK-Guided Pruning Approach

    Yadulla, Reddy A, Bhargavi K, et al. Lightweight Neural Networks for Adversarial Defense: A Novel NTK-Guided Pruning Approach. 2025 37th Conference of Open Innovations Association (FRUCT), 2025: 331-337

    2025

  3. [3]

    Pruning filter in filter

    Meng F, Cheng H, Li K, et al. Pruning filter in filter. Advances in Neural Information Processing Systems, 2020, 33: 17629-17640

    2020

  4. [4]

    Compression of neural machine translation models via pruning

    See A, Luong M-T, Manning C D. Compression of neural machine translation models via pruning. arXiv preprint arXiv:1606.09274, 2016. 17

    Pith/arXiv arXiv 2016

  5. [5]

    Towards Understanding Multi-Task Learning (Generalization) of LLMs via Detecting and Exploring Task-Specific Neurons

    Leng Y, Xiong D. Towards Understanding Multi-Task Learning (Generalization) of LLMs via Detecting and Exploring Task-Specific Neurons. arXiv preprint arXiv:2407.06488, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    ResPrune: An energy-efficient restorative filter pruning method using stochastic optimization for accelerating CNN

    Anusha J, P P. ResPrune: An energy-efficient restorative filter pruning method using stochastic optimization for accelerating CNN. Pattern Recognition, 2024, 155: 110671

    2024

  7. [7]

    Nemotron-h: A family of accurate and efficient hybrid mamba-transformer models

    Blakeman A, Basant A, Khattar A, et al. Nemotron-h: A family of accurate and efficient hybrid mamba-transformer models. arXiv preprint arXiv:2504.03624, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Plug & play generative networks: Con- ditional iterative generation of images in latent space

    Nguyen A, Yosinski J, Bengio Y, et al. Plug & play generative networks: Con- ditional iterative generation of images in latent space. CoRR, abs/1612.00005, 2016

    Pith/arXiv arXiv 2016

  9. [9]

    Network dissection: Quantifying interpretability of deep visual representations

    Bau D, Zhou B, Khosla A, et al. Network dissection: Quantifying interpretability of deep visual representations. Proceedings of the IEEE conference on computer vision and pattern recognition, 2017: 6541-6549

    2017

  10. [10]

    Gan dissection: Visualizing and under- standing generative adversarial networks

    Bau D, Zhu J Y, Strobelt H, et al. Gan dissection: Visualizing and under- standing generative adversarial networks. International Conference on Learning Representations, 2018

    2018

  11. [11]

    Understanding the role of individual units in a deep neural network

    Bau D, Zhu J Y, Strobelt H, et al. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 2020, 117(48): 30071-30078

    2020

  12. [12]

    On the importance of single directions for generalization

    Morcos A S, Bahri D G T, Cho N, et al. On the importance of single directions for generalization. arXiv preprint arXiv:1803.06959, 2018

    Pith/arXiv arXiv 2018

  13. [13]

    Revisiting the importance of individual units in cnns via ablation

    Zhou B, Sun Y, Bau D, et al. Revisiting the importance of individual units in cnns via ablation. arXiv preprint arXiv:1806.02891, 2018

    Pith/arXiv arXiv 2018

  14. [14]

    Feature visualization

    Olah C, Mordvintsev A, Schubert L. Feature visualization. Distill, 2017, 2(11): e7

    2017

  15. [15]

    Understanding CNN Hidden Neuron Activations Using Structured Background Knowledge and Deductive Reasoning

    Dalal A, Sarker M K, Barua A, et al. Understanding CNN Hidden Neuron Activations Using Structured Background Knowledge and Deductive Reasoning. Kansas State University, 2024

    2024

  16. [16]

    MULTIFLOW: Shifting Towards Task- Agnostic Vision-Language Pruning

    Farina M, Mancini M, Cunegatti E, et al. MULTIFLOW: Shifting Towards Task- Agnostic Vision-Language Pruning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024: 16185-16195

    2024

  17. [17]

    Pruning filter in filter

    Fanxu M, Hao C, Ke L, et al. Pruning filter in filter. Advances in Neural Information Processing Systems, 2020, 33: 17629-17640. 18

    2020

  18. [18]

    Comprehensive Survey of Model Compression and Speed up for Vision Transformers

    Feiyang C, Ziqian L, Lisang Z, et al. Comprehensive Survey of Model Compression and Speed up for Vision Transformers. arXiv preprint arXiv:2404.10407, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Sparsegpt: Massive language models can be accurately pruned in one-shot

    Frantar E, Alistarh D. Sparsegpt: Massive language models can be accurately pruned in one-shot. International Conference on Machine Learning, 2023: 10323- 10337

    2023

  20. [20]

    DepGraph: Towards Any Structural Prun- ing

    Gongfan F, Xinyin M, Mingli S, et al. DepGraph: Towards Any Structural Prun- ing. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023: 16091-16101

    2023

  21. [21]

    Lightweight spiking neural network training based on spike timing dependent backpropagation

    Gong Y, Chen T, Wang S, et al. Lightweight spiking neural network training based on spike timing dependent backpropagation. Neuroncomputing, 2024: 127059

    2024

  22. [22]

    A novel and efficient model pruning method for deep convolutional neural networks by evaluating the direct and indirect effects of filters

    Zheng Y, Sun, P, Ren Q, et al. A novel and efficient model pruning method for deep convolutional neural networks by evaluating the direct and indirect effects of filters. Neuroncomputing, 2024: 127124

    2024

  23. [23]

    Network Trimming: A Data-Driven Neuron Pruning Approach towards Efficient Deep Architectures

    Hu H, Peng R, Tai Y W, et al. Network Trimming: A Data-Driven Neuron Pruning Approach towards Efficient Deep Architectures. arXiv preprint arXiv:1607.03250, 2016

    Pith/arXiv arXiv 2016

  24. [24]

    Learning both weights and connections for efficient neural networks

    Han S, Pool J, Tran J, et al. Learning both weights and connections for efficient neural networks. Advances in neural information processing systems, 2015, 28

    2015

  25. [25]

    Deep residual learning for image recognition

    He K, Zhang X, Ren S, et al. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, 2016: 770-778

    2016

  26. [26]

    Adversarial attacks and defenses in explainable artificial intelligence: A survey

    Hubert B, Przemyslaw B. Adversarial attacks and defenses in explainable artificial intelligence: A survey. Information Fusion, 2024: 102303

    2024

  27. [27]

    Adversarial Training and Provable Robustness: A Tale of Two Objectives

    Fan J, Li W. Adversarial Training and Provable Robustness: A Tale of Two Objectives. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(8): 7367-7376

    2021

  28. [28]

    Learning multiple layers of features from tiny images

    Krizhevsky A, Hinton G. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009

    2009

  29. [29]

    Gradient-based learning applied to document recognition

    LeCun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998, 86(11): 2278-2324

    1998

  30. [30]

    Pruning filters for efficient ConvNets

    Li H, Kadav A, Durdanovic I, et al. Pruning filters for efficient ConvNets. Interna- tional Conference on Learning Representations, Toulon, France: OpenReview.net, 2017

    2017

  31. [31]

    LLM-Pruner: On the structural pruning of large language models

    Ma X, Fang G, Wang X. LLM-Pruner: On the structural pruning of large language models. Advances in neural information processing systems, 2023, 36: 21702-21720. 19

    2023

  32. [32]

    Pruning based on separable convolutions for object tracking

    Mao Y H, He Z Z, Liu L L. Pruning based on separable convolutions for object tracking. Journal of Xi’an Jiaotong University, 2021, 55(1): 52-59

    2021

  33. [33]

    Matching Networks for One Shot Learning

    Oriol V, Charles B, Timothy L, et al. Matching Networks for One Shot Learning. Advances in neural information processing systems, 2016, 29

    2016

  34. [34]

    Does localization inform editing? surprising differences in causality-based localization vs

    Hase P, Bansal M, Kim B. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models. Advances in Neural Information Processing Systems, 2023, 36: 17643-17668

    2023

  35. [35]

    Interspace Pruning: Using Adaptive Filter Representations to Improve Training of Sparse CNNs

    Wimmer P, Mehnert J, Condurache A. Interspace Pruning: Using Adaptive Filter Representations to Improve Training of Sparse CNNs. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022: 12527-12537

    2022

  36. [36]

    DLP: Dynamic Layerwise Pruning in Large Language Models

    Chen Y, Cheng B, Han J, et al. DLP: Dynamic Layerwise Pruning in Large Language Models. arxiv preprint arxiv:2505.23807, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Deep Model Compression based on the Training History

    Shabbeer Basha S H, Mohammad F, Viswanath P, et al. Deep Model Compression based on the Training History. Neurocomputing, 2024, 573: 127257

    2024

  38. [38]

    Quantifying local specialization in deep neural networks

    Hod S, Casper S, Filan D, et al. Quantifying local specialization in deep neural networks. arXiv preprint arXiv:2110.08058, 2021

    Pith/arXiv arXiv 2021

  39. [39]

    Natural and Imperceptible Backdoor Attack against Deep Neural Networks

    Ni S, Wang X, Shang Y, et al. Natural and Imperceptible Backdoor Attack against Deep Neural Networks. 2023 4th International Conference on Electronic Communication and Artificial Intelligence (ICECAI), 2023: 377-381

    2023

  40. [40]

    A simple and effective pruning approach for large language models

    Sun M, Liu Z, Bair A, et al. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023

    Pith/arXiv arXiv 2023

  41. [41]

    A novel and efficient model pruning method for deep convolutional neural networks by evaluating the direct and indirect effects of filters

    Yongbin Z, Peng S, Qiang R, et al. A novel and efficient model pruning method for deep convolutional neural networks by evaluating the direct and indirect effects of filters. Neurocomputing, 2024, 569: 127124

    2024

  42. [42]

    Understanding neural networks through deep visualization

    Yosinski J, Clune J, Nguyen A, et al. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015

    Pith/arXiv arXiv 2015

  43. [43]

    MTS2Graph: Interpretable multivariate time series classification with temporal evolving graphs

    Younis R, Hakmeh A, Ahmadi Z. MTS2Graph: Interpretable multivariate time series classification with temporal evolving graphs. Pattern Recognition, 2024, 152: 110486

    2024

  44. [44]

    Adversarial Attacks and Defenses in Machine Learning-Empowered Communication Systems and Networks: A Con- temporary Survey

    Yulong W, Tong S, Shenghong L, et al. Adversarial Attacks and Defenses in Machine Learning-Empowered Communication Systems and Networks: A Con- temporary Survey. IEEE Communications Surveys and Tutorials, 2023, 25: 2245-2298. 20

    2023

  45. [45]

    A survey of model compression strategies for object detection

    Zonglei L, Tong Y, Fuxi P, et al. A survey of model compression strategies for object detection. Multimed Tools Applications, 2024, 83: 48165-48236

    2024

  46. [46]

    Ablation studies in artificial neural networks

    Meyes R, Lu M, de Puiseau C W, et al. Ablation studies in artificial neural networks. arXiv preprint arXiv:1901.08644, 2019

    Pith/arXiv arXiv 1901

  47. [47]

    Recursive least squares method for training and pruning convolutional neural networks

    Yu T, Zhang C, Ma M, et al. Recursive least squares method for training and pruning convolutional neural networks. Applied Intelligence, 2023, 53(20): 24603- 24618. Appendix A Generalization Experiment Fig. A1: 20-day interval futures images. OHLC bar charts are used as the core representation, where the highest and lowest prices are clearly marked by the...

    2023