Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Training CNNs with Hebbian plasticity, neuromodulation, and metric fine-tuning adds 2 to 10 points of Top-1 accuracy.

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 ↗

NM-Hebb combines a Hebbian activation-weight alignment regulariser, a loss-gated neuromodulator, and a metric fine-tuning phase to improve CNN accuracy and embedding compactness.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible combination of Hebbian, neuromodulation, and metric losses, but the headline accuracy gains are confounded by an unequal training budget (100 vs 50 epochs) and several reporting gaps; worth refereeing if the authors fix the baseline. the 5 major comments →

arxiv 2508.19896 v1 pith:UURY23CB submitted 2025-08-27 cs.LG cs.CV

NM-Hebb: Coupling Local Hebbian Plasticity with Metric Learning for More Accurate and Interpretable CNNs

Davorin Mili\v{c}evi\'c , Ratko Grbi\'c This is my paper
classification cs.LG cs.CV
keywords Hebbian learningneuromodulationmetric learningconvolutional neural networksinterpretabilityembedding compactnessCIFAR-10TinyImageNet
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.

The reading

The paper sets out to show that standard supervised CNN training leaves accuracy and interpretability on the table, and that a two-phase training schedule can recover both. Phase 1 couples cross-entropy training with a Hebbian regulariser that aligns each filter's mean activation with its mean weight, plus a small learned neuromodulator that strengthens or relaxes the regulariser depending on the current classification loss. Phase 2 fine-tunes the network on image pairs with a Euclidean-margin loss that compresses same-class embeddings and separates different-class ones, while a gated consolidation penalty protects Phase 1 parameters. Across CIFAR-10, CIFAR-100, and TinyImageNet, and across five CNN backbones, the paper reports Top-1 gains of +2 to +10 percentage points and NMI gains up to +0.15, with more selective and more texture-aligned filters. If true, this means accuracy and interpretability can be improved without changing the architecture or using pretrained weights, which matters for small models and safety-critical deployments.

Core claim

NM-Hebb is a two-phase recipe. Phase 1 minimises cross-entropy plus a neuromodulated Hebbian regulariser $R_{\text{Hebb}}$, a per-channel penalty on the squared difference between mean activation and mean filter weight applied to one early-to-mid convolutional layer, with a two-layer MLP mapping the current cross-entropy loss to a gate in $(0,1)$, so Hebbian consolidation is strong at high error and relaxes as training stabilises. Phase 2 freezes a copy of the Phase 1 weights and fine-tunes the model on same/different-class pairs, optimising cross-entropy on each image plus a Euclidean-margin pair loss, plus a neuromodulator-gated consolidation penalty and continued Hebbian regularisation. The central re

What carries the argument

The load-bearing object is the Hebbian regulariser $R_{\text{Hebb}}$, defined as $R_{\text{Hebb}}(x) = \frac{1}{C_{\text{out}}} \sum_{f} (\bar{a}_f - \bar{w}_f)^2$, with $\bar{a}_f$ the spatial mean of activations in channel $f$ and $\bar{w}_f$ the mean of the corresponding filter weights. It is local, parameter-free, and applied to one chosen convolutional layer per backbone, keeping filter statistics aligned with what the filters themselves look like. Around it sit two gates: the neuromodulator $\nu_{\phi}$, a two-layer MLP mapping the scalar cross-entropy loss to a strength in $(0,1)$, and the Phase-2 Euclidean-margin pair loss that reshapes the embedding space. The consolidation term $||\theta - \theta_{-}||^2$ gated by $\nu_{\phi}$ plays an elastic-weight-

Load-bearing premise

The load-bearing premise is that the gains come from the proposed Hebbian, neuromodulatory, and metric mechanisms rather than from extra training time: NM-Hebb runs 100 epochs (120 on TinyImageNet) while the baseline runs 50 (60), and no equal-budget baseline is reported.

What would settle it

Train every backbone with plain cross-entropy for exactly the same epoch count, batch size, augmentation, and schedule as NM-Hebb (100 epochs on CIFAR, 120 on TinyImageNet, including Phase 2's lower learning rate). If that matched baseline matches or beats NM-Hebb on Top-1 accuracy or NMI, the mechanism claim is falsified. As a second check, run NM-Hebb with the Hebbian regulariser removed and again with the Phase-2 metric loss removed; if accuracy and NMI do not drop, those components are not doing the claimed work.

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

If this is right

  • If the reported gains hold, supervised CNN training can be upgraded by a drop-in two-phase procedure that changes no architecture and requires no pretrained weights.
  • The gain pattern—largest on the smallest backbone and on datasets with more classes—points to the method being most valuable for lightweight models and fine-grained tasks.
  • The reported NMI and filter-selectivity improvements imply tighter class clusters and more structured low-level features, which should make misclassifications easier to localize and explain in safety-critical settings.
  • Because Phase 2 supplies most of the clustering gain at roughly 20% extra wall-clock time, the metric fine-tuning step could be applied on top of many already-trained models as a post-processing stage.

Where Pith is reading between the lines

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

  • Editorial inference: an equal-budget comparison is the decisive experiment; if a 100-epoch plain cross-entropy run matches NM-Hebb, the proposed mechanisms may only be acting as an implicit regularisation schedule.
  • Editorial inference: the choice of a single hand-picked Hebbian layer per backbone is untested; sweeping the regulariser over early, middle, and late layers would show whether the 'early-to-mid stabilisation' story is required or incidental.
  • Editorial inference: the loss-gated neuromodulator makes the combined penalty adaptive in a way a fixed coefficient cannot; ablating the gate and keeping nu fixed would quantify how much of the gain is due to adaptive gating rather than to Hebbian alignment or metric learning alone.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 6 minor

Summary. The paper proposes NM-Hebb, a two-phase training framework for CNNs that combines (i) a Hebbian regularizer aligning spatial activation means with filter-weight means, (ii) a learnable neuromodulator MLP that gates Hebbian and consolidation terms, and (iii) a pairwise metric-learning fine-tuning phase that enforces intra-class compactness and inter-class separation. Experiments on CIFAR-10, CIFAR-100, and TinyImageNet across five backbones (VGG-11, ResNet-18, MobileNet-v2, EfficientNet-V2, DenseNet-121) report consistent Top-1 accuracy gains over standard supervised training and several comparative methods, together with improved NMI, filter selectivity, and texture-selective units in a Network Dissection probe. The central claim is that the gains are caused by the proposed mechanisms, yielding both higher accuracy and greater interpretability.

Significance. The proposed combination is timely and potentially useful: it is architecturally simple, adds very few parameters (the neuromodulator has <100 parameters), and the evaluation covers a reasonably wide range of backbones and dataset scales. If the accuracy improvements survive a matched training-budget control, the method would be a practical contribution to embedded and safety-critical vision. The interpretability analysis, especially the filter-morphology statistics, is a nice attempt to connect the Hebbian prior to measurable properties of learned filters. However, the current evidence is weakened by a confounded training-budget comparison, unreported hyperparameters, and missing variance information; these issues affect the load-bearing claim that the gains are caused by the proposed mechanisms rather than by additional training.

major comments (5)
  1. [§4.4 and §4.1] The evaluation is confounded by unequal training budgets. Phase 1 runs 50 epochs (60 for TinyImageNet) and Phase 2 another 50 (60), while the baseline is 'standard supervised CNN training' and the baseline numbers in Tables 3-5 appear to be 50/60-epoch results. No baseline trained for 100/120 epochs or with matched compute is reported. Consequently, the +2-10 pp accuracy gains could be due simply to additional optimization. The statement in §5.1 that Phase 2 adds 'only ~20% extra wall-clock time' is also inconsistent with a 50-epoch phase that processes pairs of images (roughly double per-step cost) and is not substantiated by wall-clock measurements. An equal-budget or equal-compute baseline is essential before attributing the gains to NM-Hebb's mechanisms.
  2. [§3.2 and §3.3] Key hyperparameters are not reported. The text says λ_hebb^(1) is set via validation search, and §3.3 introduces λ_metric, λ_cons, λ_hebb^(2), and margin m, but no values or search ranges are given anywhere in the manuscript. Table 2 fixes a single Hebbian-regularized layer per backbone without sensitivity analysis. These choices are load-bearing for the method's claimed robustness; without them the experiments are not reproducible, and the claim that NM-Hebb is 'architecture-agnostic' is untested.
  3. [Abstract and Tables 3-5] The reported numbers do not match. The abstract claims '+2.0-10.0 pp' on CIFAR-10, but Table 3 shows a maximum gain of +6.6 pp (VGG-11). Section 1 states NMI improvements of +0.07-0.15, while Table 3 shows +0.20 for MobileNet-v2 on CIFAR-10. The abstract says TinyImageNet gains are 'up to +4.3-8.9 pp', but §5.1 reports a +9.0 pp gain for ResNet-18. These discrepancies must be corrected, and the abstract should reflect the actual experimental range.
  4. [§4.5 and Tables 3-5] The paper states that results are averaged over five random initializations, but Tables 3-5 report only point estimates. Several gains are small (e.g., +2.0 pp for DenseNet-121 on CIFAR-10; NMI changes of +0.01-0.02 on TinyImageNet). Without error bars or significance tests, the claim that NM-Hebb 'consistently' outperforms all baselines cannot be reliably assessed.
  5. [§5.3 and §3.3] The interpretability claims are partly circular. NMI is computed on the embedding space that Phase 2's metric loss directly optimizes for intra-class compactness (Eq. 6), and the HAF and texture-selectivity measurements reflect the same activation-weight alignment that R_Hebb explicitly enforces (Eq. 4). The paper presents these as emergent benefits, but they are largely by construction. The authors should reframe these results as expected consequences of the objective and provide at least one interpretability measure that is not directly aligned with the training losses.
minor comments (6)
  1. [§5.1] 'ResNet-18 + BP' is undefined; presumably backpropagation-trained ResNet-18. Please define the acronym at first use.
  2. [Figure 2 caption] The caption in the text says 'after the complete two-phase NM-Hebb pipeline', but the subfigure label says 'Phase 2'. Use consistent terminology.
  3. [Table 2] EfficientNet-V2 has multiple variants; specify the exact variant (e.g., EfficientNetV2-S or -M) since the parameter count depends on it.
  4. [§4.4] The augmentation description leaves some details implicit (e.g., padding/crop size for CIFAR random cropping). Also clarify the interaction between SWA and early stopping, since SWA is said to 'often trigger training termination earlier'.
  5. [References] References [28] and [29] use arXiv numbers '2403.00001' and '2402.00001'; these look like placeholder identifiers. Please verify.
  6. [Code availability] The code is only 'available upon acceptance'; for a method with many unstated hyperparameters, actual release at submission time is strongly preferable.

Circularity Check

0 steps flagged

No significant circularity: NM-Hebb's claims are empirical evaluations, not derivations that reduce to their own inputs.

full rationale

The paper's central chain is an empirical comparison: Phase 1 optimizes cross-entropy plus a Hebbian regularizer (Eq. 1), and Phase 2 optimizes cross-entropy plus a pairwise metric loss plus consolidation and continued Hebbian terms (Eq. 6). The reported Top-1 accuracy and NMI are test-set measurements after training, not algebraic consequences of the loss definitions. While the Phase 2 metric loss directly encourages intra-class compactness and inter-class separation, NMI is an independent clustering metric computed on test embeddings; it is not identical to the Euclidean-margin loss, so the NMI gains are an empirical outcome rather than a tautology. Similarly, the Hebbian regularizer aligns activation means with filter-weight means, but the reported interpretability metrics (HAF, speckle rate, Network Dissection texture alignment) are not defined in terms of that alignment, so those results are not forced by construction. There are no load-bearing self-citations: the cited Hebbian, neuromodulation, metric-learning, and distillation works are external, and the comparative baselines are standard references. The unequal training budget (100 epochs for NM-Hebb vs. 50 for the baseline) is a legitimate experimental confound that threatens the accuracy attribution, but it is not a circularity: the accuracy numbers are not fitted from the method's equations and the paper does not derive them from its own premises. No quoted step reduces to its own input, so no circular step can be exhibited under the required standard.

Axiom & Free-Parameter Ledger

6 free parameters · 3 axioms · 1 invented entities

The method relies on several hyperparameters that are tuned on validation but never disclosed, plus design choices such as the Hebbian layer index. These are free parameters of the central claim. The invented entity is just the small neuromodulator network.

free parameters (6)
  • lambda_hebb (Phase 1) = not reported
    Weighting of Hebbian term in Eq. 1, set via validation search per Section 3.2.
  • lambda_metric = not reported
    Weight of pairwise metric loss in Eq. 2/6, not specified in text.
  • lambda_cons = not reported
    Weight of consolidation penalty in Eq. 2, not specified.
  • lambda_hebb (Phase 2) = not reported
    Weight of Hebbian term in Phase 2, not specified.
  • margin m = not reported
    Margin in pairwise metric loss, not specified.
  • Hebbian layer index = per backbone
    A single layer per backbone is chosen 'in the early-to-mid hierarchy' (Table 2); no sensitivity study.
axioms (3)
  • domain assumption Aligning activation means with filter weight means is a beneficial regulariser
    The paper assumes this penalty stabilises early features; no theoretical or biological justification beyond analogy.
  • domain assumption A loss-gated neuromodulator improves plasticity-stability trade-off
    The use of the MLP gate is heuristic; no proof that it beats fixed weights.
  • domain assumption A simple random-pair Euclidean margin loss is sufficient for embedding shaping
    The pairwise loss uses equal-probability same/different pairs; no analysis of mining strategies.
invented entities (1)
  • Neuromodulator MLP no independent evidence
    purpose: Gates the Hebbian and consolidation losses based on current cross-entropy loss
    A tiny two-layer MLP with 8 ReLU units and a sigmoid output; it is a standard neural component, not an empirically validated new mechanism beyond being part of the proposed method.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of NM-Hebb: Coupling Local Hebbian Plasticity with Metric Learning for More Accurate and Interpretable CNNs." pith.science (2026). https://pith.science/paper/UURY23CB

@misc{pith2026250819896,
  author       = {Pith},
  title        = {Pith review of: NM-Hebb: Coupling Local Hebbian Plasticity with Metric Learning for More Accurate and Interpretable CNNs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UURY23CB}},
  note         = {Machine review of arXiv:2508.19896}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep Convolutional Neural Networks (CNNs) achieve high accuracy but often rely on purely global, gradient-based optimisation, which can lead to overfitting, redundant filters, and reduced interpretability. To address these limitations, we propose NM-Hebb, a two-phase training framework that integrates neuro-inspired local plasticity with distance-aware supervision. Phase 1 extends standard supervised training by jointly optimising a cross-entropy objective with two biologically inspired mechanisms: (i) a Hebbian regulariser that aligns the spatial mean of activations with the mean of the corresponding convolutional filter weights, encouraging structured, reusable primitives; and (ii) a learnable neuromodulator that gates an elastic-weight-style consolidation loss, preserving beneficial parameters without freezing the network. Phase 2 fine-tunes the backbone with a pairwise metric-learning loss, explicitly compressing intra-class distances and enlarging inter-class margins in the embedding space. Evaluated on CIFAR-10, CIFAR-100, and TinyImageNet across five backbones (ResNet-18, VGG-11, MobileNet-v2, EfficientNet-V2, DenseNet-121), NM-Hebb achieves consistent gains over baseline and other methods: Top-1 accuracy improves by +2.0-10.0 pp (CIFAR-10), +2.0-9.0 pp (CIFAR-100), and up to +4.3-8.9 pp (TinyImageNet), with Normalised Mutual Information (NMI) increased by up to +0.15. Qualitative visualisations and filter-level analyses further confirm that NM-Hebb produces more structured and selective features, yielding tighter and more interpretable class clusters. Overall, coupling local Hebbian plasticity with metric-based fine-tuning yields CNNs that are not only more accurate but also more interpretable, offering practical benefits for resource-constrained and safety-critical AI deployments.

Figures

Figures reproduced from arXiv: 2508.19896 by Davorin Mili\v{c}evi\'c, Ratko Grbi\'c.

Figure 1
Figure 1. Figure 1: t-SNE visualisations of ResNet-18 embeddings on CIFAR-10 after Phase 1 (a) and after the full two-phase NM-Hebb pipeline (b). (a) CIFAR-100, Phase 1 (b) CIFAR-100, Phase 2 [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE visualisations of ResNet-18 embeddings on CIFAR-100 after Phase 1 (a) and after the full two-phase NM-Hebb pipeline (b). 10 [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Constrained Hebbian Learning Supports Efficient Representational Allocation under Structural Constraints

    cs.LG 2026-07 conditional novelty 4.0

    A constrained Hebbian rule produces audiovisual representations with lower task-information cost (retained input information per unit of task-relevant information) than sparse backpropagation and DDTP at comparable ac...

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    Simonyan, A

    K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, in: International Conference on Learning Representations (ICLR), 2015

    work page 2015

  2. [2]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    work page 2016

  3. [3]

    Sandler, A

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L.-C. Chen, Mobilenetv2: Inverted residuals and linear bottlenecks, in: CVPR, 2018, pp. 4510–4520

    work page 2018

  4. [4]

    M. Tan, Q. V . Le, Efficientnet: Rethinking model scaling for convolutional neural networks, in: ICML, 2019, pp. 6105–6114

    work page 2019

  5. [5]

    X. Ding, X. Zhang, Y . Zhou, J. Han, G. Ding, J. Sun, Scaling up your kernels to 31 ×31: Revisiting large kernel design in cnns, in: CVPR, 2022

    work page 2022

  6. [6]

    M. Tan, Q. V . Le, Efficientnetv2: Smaller models and faster training, in: ICML, 2021

    work page 2021

  7. [7]

    Woo, et al., Convnext v2: Co-designing and scaling convnets with masked auto-encoders, in: CVPR, 2023

    S. Woo, et al., Convnext v2: Co-designing and scaling convnets with masked auto-encoders, in: CVPR, 2023

    work page 2023

  8. [8]

    F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, K. Keutzer, Squeezenet: Alexnet-level accuracy with 50x fewer parameters and<0.5mb model size, arXiv preprint arXiv:1602.07360 (2016)

    Pith/arXiv arXiv 2016

  9. [9]

    K. Sohn, D. Berthelot, C.-L. Li, Z. Zhang, N. Carlini, E. D. Cubuk, A. Kurakin, H. Zhang, C. Ra ffel, Fixmatch: Simplifying semi-supervised learning with consistency and confidence, in: Advances in Neural Information Processing Systems (NeurIPS), 2020

    work page 2020

  10. [10]

    Hebbian Semi-Supervised Learning in a Sample Efficiency Setting

    G. Lagani, F. Falchi, C. Gennaro, G. Amato, Hebbian semi-supervised learning in a sample efficiency setting, arXiv preprint arXiv:2103.09002 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  11. [11]

    Hinton, O

    G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, arXiv preprint arXiv:1503.02531 (2015)

    Pith/arXiv arXiv 2015

  12. [12]

    Touvron, M

    H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, H. J ´egou, Training data-efficient image transformers and distillation through attention, in: ICML, 2021, pp. 10347–10357

    work page 2021

  13. [13]

    Y . Liu, Z. Yang, Y . Hu, Repeated self-distillation, in: NeurIPS, 2024. 12

    work page 2024

  14. [14]

    Khosla, et al., Supervised contrastive learning, in: NeurIPS, V ol

    P. Khosla, et al., Supervised contrastive learning, in: NeurIPS, V ol. 33, 2020, pp. 18661–18673

    work page 2020

  15. [15]

    S. Kim, D. Kim, M. Cho, S. Kwak, Proxy-anchor loss for deep metric learning, in: CVPR, 2020

    work page 2020

  16. [16]

    Amato, F

    G. Amato, F. Carrara, F. Falchi, C. Gennaro, G. Lagani, Hebbian learning meets deep convolutional neural networks, in: ICIAP, 2019, pp. 324–334

    work page 2019

  17. [17]

    Journ ´e, H

    A. Journ ´e, H. Sprekeler, G. Bellec, Hebbian deep learning without feedback, in: ICLR, 2023

    work page 2023

  18. [18]

    Miconi, A

    T. Miconi, A. Rawal, J. Clune, K. O. Stanley, Backpropamine: Training self-modifying neural networks with di fferentiable neuromodulated plasticity, in: ICLR, 2020

    work page 2020

  19. [19]

    FastHebb: Scaling Hebbian Training of Deep Neural Networks to ImageNet Level

    G. Lagani, C. Gennaro, H. Fassold, G. Amato, Fasthebb: Scaling hebbian training of deep neural networks to imagenet level, arXiv preprint arXiv:2207.03172 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    Jim ´enez Nimmo, E

    J. Jim ´enez Nimmo, E. Mondrag ´on, Advancing the biological plausibility and e fficacy of hebbian convolutional neural networks, Neural Networks 190 (2025) 107628. doi:10.1016/j.neunet.2025.107628

    work page arXiv 2025

  21. [21]

    Miconi, J

    T. Miconi, J. Clune, K. O. Stanley, Di fferentiable plasticity: Training plastic neural networks with backpropagation, in: ICML, 2018

    work page 2018

  22. [22]

    Liu, et al., Sstdp: Supervised spike-timing-dependent plasticity for e fficient spiking neural network training, Frontiers in Neuroscience 15 (2021) 756876

    F. Liu, et al., Sstdp: Supervised spike-timing-dependent plasticity for e fficient spiking neural network training, Frontiers in Neuroscience 15 (2021) 756876

    work page 2021

  23. [23]

    Kirkpatrick, et al., Overcoming catastrophic forgetting in neural networks, PNAS 114 (13) (2017) 3521–3526

    J. Kirkpatrick, et al., Overcoming catastrophic forgetting in neural networks, PNAS 114 (13) (2017) 3521–3526

    work page 2017

  24. [24]

    Zenke, B

    F. Zenke, B. Poole, S. Ganguli, Continual learning through synaptic intelligence, in: ICML, 2017, pp. 3987–3995

    work page 2017

  25. [25]

    Foret, A

    P. Foret, A. Kleiner, H. Mobahi, B. Neyshabur, Sharpness-aware minimization for efficiently improving generalization, in: ICLR, 2021

    work page 2021

  26. [26]

    Deng, et al., Arcface: Additive angular margin loss for deep face recognition, in: CVPR, 2019

    J. Deng, et al., Arcface: Additive angular margin loss for deep face recognition, in: CVPR, 2019

    work page 2019

  27. [27]

    Qian, et al., Softtriple loss: Deep metric learning without triplet sampling, in: ICCV , 2019

    Q. Qian, et al., Softtriple loss: Deep metric learning without triplet sampling, in: ICCV , 2019

    work page 2019

  28. [28]

    Metabolic scaling in small life forms

    D. Geißler, K.-R. M ¨uller, W. Samek, Enhancing interpretability through loss-defined classification objective in structured latent spaces, arXiv preprint arXiv:2403.00001 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  29. [29]

    Opitz, T

    M. Opitz, T. Ropinski, Comparison of embedded spaces for deep learning classification, arXiv preprint arXiv:2402.00001 (2024)

    Pith/arXiv arXiv 2024

  30. [30]

    Huang, Z

    G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, Densely connected convolutional networks, in: CVPR, 2017, pp. 4700–4708

    work page 2017

  31. [31]

    Krizhevsky, Cifar-10 dataset, dataset page, accessed 2025-08-11 (2009)

    A. Krizhevsky, Cifar-10 dataset, dataset page, accessed 2025-08-11 (2009). URL https://www.cs.toronto.edu/~kriz/cifar.html

    work page 2025

  32. [32]

    Krizhevsky, Cifar-100 dataset, dataset page, accessed 2025-08-11 (2009)

    A. Krizhevsky, Cifar-100 dataset, dataset page, accessed 2025-08-11 (2009). URL https://www.cs.toronto.edu/~kriz/cifar.html

    work page 2025

  33. [33]

    Y . Le, X. Yang, Tiny imagenet (tiny-imagenet-200), stanford CS231n dataset file, accessed 2025-08-11 (2015). URL http://cs231n.stanford.edu/tiny-imagenet-200.zip

    work page 2025

  34. [34]

    Slegroux, Tiny imagenet-200 (clean) with test labels, Available at https://huggingface.co/datasets/slegroux/ tiny-imagenet-200-clean , accessed: 2025-08-16 (2023)

    work page 2025

  35. [35]

    Willmore, D

    B. Willmore, D. J. Tolhurst, Characterizing the sparseness of neural codes, Network: Computation in Neural Systems 12 (3) (2001) 255–270. doi:10.1080/net.12.3.255.270

    work page doi:10.1080/net.12.3.255.270 2001

  36. [36]

    response probability

    B. Haider, et al., Synaptic and network mechanisms of sparse and reliable visual cortical activity, Neuron 63 (4) (2010) 506–519, see supplemental for the “response probability” sparseness measure. doi:10.1016/j.neuron.2009.07.025

    work page doi:10.1016/j.neuron.2009.07.025 2010

  37. [37]

    D. J. Field, Relations between the statistics of natural images and the response properties of cortical cells, Journal of the Optical Society of America A 4 (12) (1987) 2379–2394

    work page 1987

  38. [38]

    Portilla, E

    J. Portilla, E. P. Simoncelli, A parametric texture model based on joint statistics of complex wavelet coe fficients, International Journal of Computer Vision 40 (1) (2000) 49–70

    work page 2000

  39. [39]

    Bruna, S

    J. Bruna, S. Mallat, Invariant scattering convolution networks, IEEE TPAMI 35 (8) (2013) 1872–1886

    work page 2013

  40. [40]

    Geirhos, J.-H

    R. Geirhos, J.-H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge, F. A. Wichmann, Shortcut learning in deep neural networks, Nature Machine Intelligence 2 (11) (2020) 665–673

    2020

  41. [41]

    D. Yin, R. Gontijo Lopes, J. Shlens, E. D. Cubuk, J. Gilmer, A fourier perspective on model robustness in computer vision, in: NeurIPS, V ol. 32, 2019

    work page 2019

  42. [42]

    D. Bau, B. Zhou, A. Khosla, A. Oliva, A. Torralba, Network dissection: Quantifying interpretability of deep visual representations, in: CVPR, 2017, pp. 6541–6549

    work page 2017

  43. [43]

    C. Olah, A. Satyanarayan, I. Johnson, S. Carter, L. Schubert, K. Ye, A. Mordvintsev, The building blocks of interpretability, Distill 3 (3) (2018) e10. doi:10.23915/distill.00010. URL https://distill.pub/2018/building-blocks/

    doi:10.23915/distill.00010 2018

  44. [44]

    Zhang, R

    Q. Zhang, R. Cao, F. Shi, Y . N. Wu, S.-C. Zhu, Interpretable convolutional neural networks, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 8827–8836

    work page 2018

  45. [45]

    Neyshabur, Towards learning convolutions from scratch, in: Advances in Neural Information Processing Systems, V ol

    B. Neyshabur, Towards learning convolutions from scratch, in: Advances in Neural Information Processing Systems, V ol. 33, 2020, pp. 13890–13902. 13

    work page 2020

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.