Softlog-Softmax Layers and Divergences Contribute to a Computationally Dependable Ensemble Learning

Abdourrahmane Mahamane Atto (LISTIC)

REVIEW 4 major objections 5 minor 22 references

Softlog-Softmax Layers and Divergences Contribute to a Computationally Dependable Ensemble Learning

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 Inserting a bounded log transform between softmax layers prevents cross-entropy divergence and yields a normalized ensemble-agreement score.

desk verdict A modest bounded-log trick with a plausible but unproved divergence bound, wrapped in an ensemble architecture that is never tested against a control. read the letter →

arxiv 2506.04297 v1 pith:T57TATWD submitted 2025-06-04 cs.LG

Abdourrahmane Mahamane Atto (LISTIC) This is my paper · ORCID

classification cs.LG

keywords softlogensemblelearningcross-entropystabilityboundeddivergenceperformancetensorconvolutionalfrustumsoftmaxcascadegeometrons

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 tries to establish that a simple bounded log transform, the softlog, removes a numerical failure mode that appears when ensemble decision layers are stacked: without it, successive softmax operations exponentiate already-exponential values and the cross-entropy objective regularly diverges. With the softlog inserted between layers, the same cascade stays stable and sub-model decisions can be fused directly at the probability level instead of on hidden features. The paper also derives a softlog-based divergence that it claims always lies in a normalized $[0,1]$ interval, giving a readable agreement score between any branch and the ensemble. It further proposes a performance tensor that reports how accuracy varies across random initializations rather than only the best run. If these claims hold, the practical gain is that dependable ensemble evaluation and interpretation need not rely on expensive class activation maps or max-accuracy reporting.

What carries the argument

The load-bearing object is the softlog transform $L_e(x)=\log((e-e^{-1})x+e^{-1})$, a bounded logarithm that maps probability values in $[0,1]$ to $[-1,1]$. Inserted between softmax layers in a cascade, it prevents the repeated exponentiation that makes cross-entropy diverge, and it also defines a softlog entropy/cross-entropy and the symmetric SoftLog Divergence $D_e$, which the paper uses as a normalized agreement measure between classifiers. The supporting machinery is the frustum-structured ensemble DRAGONFLY, whose twelve sub-networks are integrated through softlog-softmax communities, and the performance tensor $\langle \min, (\text{mean}, \text{median}), \max\rangle$ that reports stochastic training behavior.

What would settle it

Take two sets of class probabilities in which one set is certain of a class that the other essentially never predicts, and evaluate the symmetric softlog divergence from Eq. (9). The paper's claimed bound says the result must lie between 0 and 1, so any value outside that interval refutes the normalized-divergence claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a bounded logarithmic transform, the softlog $L_e(x)=\log((e-e^{-1})x+e^{-1})$, makes it safe to compose several softmax decision layers in an ensemble. Without such a transform, applying softmax twice in succession exponentiates exponentially distributed values and drives the cross-entropy objective to diverge; with a softlog inserted between layers, the values stay in $[-1,1]$ and training remains stable. The same transform yields a softlog entropy and a symmetric softlog divergence, $D_e$, that the paper asserts always lies in $[0,1]$ for categorical distributions, turning divergence into a normalized agreement score. On the DRAGONFLY ensemble, built from twelve convolutional frustum networks integrated through softlog-softmax communities, the paper shows that sub-model and sub-community contributions to the final decision can be read from divergence histograms without class activation maps. It also introduces a performance tensor based on Min, Mean/Median, and Max to report training outcomes across random initializations, arguing that reporting only the maximum accuracy hides a lack of dependability.

Load-bearing premise

The interpretation of divergence values as normalized agreement scores depends on the paper's assertion, stated without proof, that the symmetric softlog divergence always falls between 0 and 1 for every pair of class-probability distributions.

Editorial extensions

If this is right

Decision fusion can be performed directly on softmax probability outputs rather than on hidden features, because the softlog between softmax layers removes the repeated-exponentiation failure mode.
The symmetric softlog divergence gives a bounded $[0,1]$ agreement score between any two classifiers, so contributions of individual branches and sub-communities can be compared in absolute terms without class activation maps.
Reporting Min, Mean/Median, and Max plus the Ability scalar exposes initialization sensitivity that a max-accuracy report hides, as shown by the low minimum performance on the more difficult experiments.
On more difficult recognition problems, diversity among the canonical frustum shapes becomes more valuable, so the ensemble design principle scales with problem complexity.

Reading between the lines

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

The paper leaves open whether the softlog's stabilization argument transfers to other settings that compose softmax with logarithms, such as large-vocabulary classifiers; a direct test would be to replace every log-of-softmax term in a large-scale model with the softlog and measure training stability.
If the claimed $[0,1]$ bound on the softlog divergence can be proven, the divergence histogram could serve as an online monitoring signal for ensemble health, flagging branches whose decisions drift away from the community during training.
The performance tensor is stated for ensemble models but applies to any stochastic training run; a natural extension is to report $\langle \min, (\text{mean}, \text{median}), \max\rangle$ across random seeds as a general reproducibility standard.

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

The headline is this: the paper contains one small, real idea—a bounded log transform that maps [0,1] to [-1,1] and avoids log(0) when composed with softmax—and one large unsubstantiated claim that this trick improves ensemble consistency and dependability. The math of the softlog is straightforward and correct; the problem is that the empirical section never removes it.

What is actually new: the softlog operator Le(x)=log((e-e^{-1})x+e^{-1}) is a neat parameter-free solution to log(softmax) numerical blowup. The chosen constants come from boundary conditions, not fitting. Entropy variants with finite bounds are natural, and the performance tensor with ability A is a sensible way to report stochastic training results. The DRAGONFLY architecture is a specific instantiation, not a general principle. The paper is readable, if occasionally overwrought.

The soft spots are load-bearing. Section 3.4 asserts that without softlog, two successive softmaxes cause cross-entropy to regularly diverge, but no experiment trains a control ensemble without softlog, or compares DRAGONFLY against plain averaging or standard log-softmax. Table 2 reports only the softlog version. So the central causal claim is untested. The bound 0≤D_e≤1 in Eq. (11) is asserted without proof; the 1/2 factor in Eq. (9) is also unexplained. A referee could probably fill these gaps, but as written they are holes. There are also small inconsistencies (N17's ability is 89.7% in Table 3 versus 88.7% in Table 2). The experiments are three small letter-recognition tasks from the author's own dataset, with no task-complexity ablation and no code or seed artifacts, so the general claims about next-generation ensemble systems are not supported.

Who this is for: someone working on softmax stabilization or ensemble fusion might read it for the softlog trick, but should not cite it as evidence without re-running the baselines. The paper deserves a serious referee because the mathematical claims are checkable and the core idea, while modest, is not wrong—it just lacks proof and controls. I would send it out for review with a request for a proof of Eq. (11) and a comparison against softlog-free ensembles.

Referee Report

4 major / 5 minor

Summary. The paper introduces a bounded logarithmic transform, the softlog L_e(x)=log((e-e^{-1})x+e^{-1}), and proposes cascading softlog-softmax operations inside ensemble decision layers to avoid numerical divergence of cross-entropy. It also defines softlog-based entropy and a symmetric softlog divergence (SLD) claimed to lie in [0,1], and proposes a performance-tensor framework (min, mean, median, max, and a weighted Ability metric) for evaluating stochastic training runs. The method is demonstrated on three small letter-recognition tasks from the author's own geometrons dataset, with SLD used to rank sub-model contributions to the ensemble.

Significance. The softlog construction is elegant and partly parameter-free: its scale and shift are fixed by the boundary conditions L_e(0)=-1 and L_e(1)=1 rather than fitted to results, and it avoids the numerical zero-log problem of standard log-softmax combinations. If the claimed divergence bound and the empirical stability benefit were rigorously established, the softlog-softmax cascade could be a useful, inexpensive recipe for ensemble aggregation, and the normalized SLD could provide an interpretable agreement score. However, the current manuscript does not provide the control experiments or proof needed to support the central 'improve consistency and dependability' claim, so the significance is potential rather than demonstrated.

major comments (4)

[Abstract and Section 3.4] The central causal claim that softlog-softmax cascades 'improve both consistency and dependability' is not tested against any baseline without the softlog. Section 3.4 asserts that without the softlog, '2 successive softmax ... and the cross entropy regularly diverges,' but no loss curves, accuracy values, or convergence statistics are reported for a softmax-softmax control, a standard log-softmax aggregation, or a plain averaging ensemble. Every experiment in Tables 2 and 3 uses DRAGONFLY with softlog, so the divergence-prevention mechanism is never empirically isolated. This is load-bearing: without such a control, the abstract's improvement claim is unsupported.
[Section 4.1, Eq. (11)] The interval property 0 ≤ D_e(Y_ris,Y_rjs) ≤ 1 for the symmetric softlog divergence is asserted without proof or citation. Section 6 relies on this property to interpret SLD values as normalized agreement scores, so the bound is essential. Please provide a complete proof, including how the 1/2 factor in Eq. (9) and the additional 1/2 symmetrization in Eq. (10) combine, and state the conditions under which equality at 0 and 1 is attained.
[Section 2.1 and Tables 2-3] The paper generalizes from three small letter-recognition tasks on a single self-created dataset to 'next generation ensemble learning systems,' but it provides no task-complexity ablation, no scaling analysis, and no independent dataset. The abstract claims 'the more complex the problem, the more useful this ensemble diversity is,' yet only three binary-or-four-class tasks with very similar structure are reported. This is a load-bearing generalization gap; a systematic complexity sweep or a second dataset is needed to support the claim.
[Section 6, Table 3] The claim that SLD 'allows identifying the most significant contributory elements with remarkably reduced effort' is supported only by descriptive histograms and a few anecdotal failure examples. There is no quantitative validation against established attribution or interpretability baselines (e.g., CAM-based methods, gradient saliency, or leave-one-out ablation), and no measure of whether the SLD rankings are stable across random seeds. The qualitative conclusions about sub-model roles are therefore not yet substantiated.

minor comments (5)

[Section 3.3, Eq. (3)] The notation 'e = e1' in Eq. (3) is confusing; please simply state that e denotes Euler's number and avoid the 'e1' notation.
[Section 3.4] The phrase '2 successive softmax' should be reworded to 'two successive softmax operations' for clarity.
[Section 5.2, Eq. (13)] The Ability weights α=1/2, β=γ=1/4 are stated without sensitivity analysis; since the performance-tensor conclusions in Section 5.3 are drawn from Ability values, a brief robustness check (e.g., a small grid of weight combinations) would strengthen the claim.
[Section 2.1 and Table 2] The experimental setup omits details needed for reproducibility: the encryption generation procedure for the SV A level, the optimization solver version, learning-rate schedule, weight decay, and random seed policy are not specified.
[Section 6, Figure 3] The histograms of SLD values lack information about the number of bins, the range shown, and whether the test set is the same as that used for Table 3; adding these details would make the figure self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the softlog and SLD are constructed from boundary conditions and definitions, and the self-citations are supporting artifacts rather than load-bearing evidence.

full rationale

The paper's central constructions are not circular. The softlog in Eq. (3) is derived from the explicit boundary conditions L(0)=-1 and L(1)=1, not fitted to any target result. The softlog entropy and SLD in Section 4 are defined by substituting the softlog for the standard logarithm; the claimed bounds in Eqs. (5), (7), and (11) are mathematical assertions about those definitions, not conclusions equivalent to the definitions by construction. The performance tensor of Eq. (12) and Ability of Eq. (13) are similarly definitions with a stated default weight choice, not fitted predictions. The self-citations to the author's geometrons dataset [21] and the earlier performance-tensor chapter [22] provide experimental material and a prior definition, respectively; neither is invoked as a uniqueness theorem or as the sole justification for a central claim. Section 3.4's assertion that softmax-softmax cascades 'regularly diverge' without the softlog is not accompanied by a softlog-free control experiment, but that is an empirical-evidence weakness, not a circularity: the claim is falsifiable and is not made true by construction. No step in the derivation reduces a predicted quantity to its own input, so the circularity burden is not met.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the probability simplex assumption, the representativeness of the geometrons dataset, the transferability of the performance tensor from simple to complex tasks, and the unproved [0,1] bound of the SLD divergence.

free parameters (3)

Softlog affine parameters = α = e - e^{-1}, β = e^{-1}
Chosen in Section 3.3 to enforce L(0)=-1 and L(1)=1, giving a bounded log. This normalization defines the operator and is not derived from data.
SLD half-scale factor = 1/2
Introduced in Eq. (9) as part of the relative softlog entropy definition to scale the maximum divergence to 1. It is a hand-selected normalization.
Ability combination weights = α=1/2, β=γ=1/4
Set in Section 5.2 with the comment that the community should later decide on a consensual combination. They are chosen, not fitted.

assumptions (4)

standard math Class probabilities lie in the unit simplex, summing to 1.
Used throughout for softmax outputs and divergence definitions; standard probability assumption.
domain assumption The geometrons dataset and the three tasks are representative of ensemble learning problems.
All experiments use the author's own dataset; no external benchmark is used, and the paper generalizes from these small tasks.
domain assumption The performance tensor computed on simple problems provides a template that transfers to more complex problems.
Section 5.1 states models should be characterized on simple problems before claiming performance on complex ones; this transferability is assumed rather than tested.
ad hoc to paper The symmetric softlog divergence is bounded between 0 and 1 as stated.
Eq. (11) asserts 0≤D_e≤1 without proof; Section 6 interprets SLD values as normalized agreement on the basis of this claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Softlog-Softmax Layers and Divergences Contribute to a Computationally Dependable Ensemble Learning." pith.science (2026). https://pith.science/paper/T57TATWD

@misc{pith2026250604297,
  author       = {Pith},
  title        = {Pith review of: Softlog-Softmax Layers and Divergences Contribute to a Computationally Dependable Ensemble Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T57TATWD}},
  note         = {Machine review of arXiv:2506.04297}
}

read the original abstract

The paper proposes a 4-step process for highlighting that softlog-softmax cascades can improve both consistency and dependability of the next generation ensemble learning systems. The first process is anatomical in nature: the target ensemble model under consideration is composed by canonical elements relating to the definition of a convolutional frustum. No a priori is considered in the choice of canonical forms. Diversity is the main criterion for selecting these forms. It is shown that the more complex the problem, the more useful this ensemble diversity is. The second process is physiological and relates to neural engineering: a softlog is derived to both make weak logarithmic operations consistent and lead, through multiple softlog-softmax layers, to intermediate decisions in the sense of respecting the same class logic as that faced by the output layer. The third process concerns neural information theory: softlog-based entropy and divergence are proposed for the sake of constructing information measures yielding consistent values on closed intervals. These information measures are used to determine the relationships between individual and sub-community decisions in frustum diversitybased ensemble learning. The concluding process addresses the derivation of an informative performance tensor for the purpose of a reliable ensemble evaluation.

Figures

Figures reproduced from arXiv: 2506.04297 by the authors.

**Figure 2.** DRAGONFLY N17 frustum conovolution structuring: Elements T1,1 and T1,2 (parent’s community); T2,1 and T2,2(children community); T3 (multigenerational). Canonical network Nk has convolution frustum size Mk with: M1 “ M7 “ t128u: cylindric-wide frustum; M2 “ M8 “ t32, 32, 32u: cylindric-elongated frustum; M3 “ M9 “ t96, 64, 32u: contractive frustum; M4 “ M10 “ t80, 48, 80u: hyperbolic frustum; M5 “ M11 “ t48, 80, 48u… view at source ↗

**Figure 3.** Histograms of SLD ` p ris , p r17s ˘ of Eq. (14) for Experiment #3. ppF “ “O”q N17 0.293 N6 1 N9 0.997 N10 0.999 N11 0.997 ppF “ “O”q N17 0.463 N4 0.695 N9 0.864 N12 0.948 N16 0.998 [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Illustrative DRAGONFLY N17 failures: it predicts “I” whereas the true label is “O”. The majority of sub-models are wrong, excepted, per image, those highlighted at paq - pbq. Finally, [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

Lessons from a dragon fly’s brain: Evolution built a small, fast, efficient neural network in a dragonfly. why not copy it for missile defense?

F. Chance, “Lessons from a dragon fly’s brain: Evolution built a small, fast, efficient neural network in a dragonfly. why not copy it for missile defense?” IEEE Spectrum, vol. 58, no. 8, pp. 28–33, 2021. 1

work page 2021
[2]

Redefining neural architecture search of hetero- geneous multinetwork models by characterizing variation operators and model components,

U. Garciarena, R. Santana, and A. Mendiburu, “Redefining neural architecture search of hetero- geneous multinetwork models by characterizing variation operators and model components,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 8, pp. 10 561–10 575,

work page
[3]

Zero-shot neural architecture search: Challenges, solutions, and opportunities,

G. Li, D. Hoang, K. Bhardwaj, M. Lin, Z. Wang, and R. Marculescu, “Zero-shot neural architecture search: Challenges, solutions, and opportunities,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 7618–7635, 2024. 1

work page 2024
[4]

Weakly supervised text attention network for generating text proposals in scene images,

L. Rong, E. MengYi, L. JianQiang, and Z. HaiBin, “Weakly supervised text attention network for generating text proposals in scene images,” in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR) , vol. 01, 2017, pp. 324–330. 2

work page 2017
[5]

Deep features class activation map for thermal face detection and tracking,

A. Kwa´sniewska, J. Rumi´nski, and P. Rad, “Deep features class activation map for thermal face detection and tracking,” in 2017 10th International Conference on Human System Interactions (HSI), 2017, pp. 41–47. 2

work page 2017
[6]

Learning of counting crowded birds of various scales via novel density activation maps,

S. Kim and M. Kim, “Learning of counting crowded birds of various scales via novel density activation maps,” IEEE Access, vol. 8, pp. 155 296–155 305, 2020. 2

work page 2020
[7]

Task-aware weakly supervised object localization with transformer,

M. Meng, T. Zhang, Z. Zhang, Y . Zhang, and F. Wu, “Task-aware weakly supervised object localization with transformer,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 7, pp. 9109–9121, 2023. 2

work page 2023
[8]

Saliency as pseudo-pixel supervision for weakly and semi-supervised semantic segmentation,

M. Lee, S. Lee, J. Lee, and H. Shim, “Saliency as pseudo-pixel supervision for weakly and semi-supervised semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 10, pp. 12 341–12 357, 2023. 2

work page 2023

Show all 22 references

[9]

Memory-based cross-image contexts for weakly supervised semantic segmentation,

J. Fan and Z. Zhang, “Memory-based cross-image contexts for weakly supervised semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 5, pp. 6006–6020, 2023. 2

2023
[10]

Background- aware classification activation map for weakly supervised object localization,

L. Zhu, Q. She, Q. Chen, X. Meng, M. Geng, L. Jin, Y . Zhang, Q. Ren, and Y . Lu, “Background- aware classification activation map for weakly supervised object localization,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence , vol. 45, no. 12, pp. 14 175–14 191, 2023. 2

2023
[11]

Learning to recognize actions on objects in egocen- tric video with attention dictionaries,

S. Sudhakaran, S. Escalera, and O. Lanz, “Learning to recognize actions on objects in egocen- tric video with attention dictionaries,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 6, pp. 6674–6687, 2023. 2

2023
[12]

Ts-cam: Token semantic coupled attention map for weakly supervised object localization,

Y . Yao, F. Wan, W. Gao, X. Pan, Z. Peng, Q. Tian, and Q. Ye, “Ts-cam: Token semantic coupled attention map for weakly supervised object localization,” IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 7, pp. 9109–9121, 2024. 2

2024
[13]

Regularized loss with hyperparameter estimation for weakly supervised single class segmentation,

Z. Ji and O. Veksler, “Regularized loss with hyperparameter estimation for weakly supervised single class segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 5, pp. 3923–3937, 2024. 2

2024
[14]

Translation-classification loss for sar image understanding with deep learning,

A. Bralet, A. M. Atto, J. Chanussot, and E. Trouvé, “Translation-classification loss for sar image understanding with deep learning,” Computer Vision and Image Understanding , vol. 257, p. 104374, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S1...

2025
[15]

A class activation mapping guided adversarial training method for land-use classification and object detection,

R. Yang, X. Xu, Z. Xu, C. Ding, and F. Pu, “A class activation mapping guided adversarial training method for land-use classification and object detection,” in IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium , 2019, pp. 9474–9477. 2

2019
[16]

Class activation map-based data augmentation for satellite smoke scene detection,

J. Xie, F. Yu, H. Wang, and H. Zheng, “Class activation map-based data augmentation for satellite smoke scene detection,” IEEE Geoscience and Remote Sensing Letters , vol. 19, pp. 1–5, 2022. 2

2022
[17]

C. M. Bishop, Pattern Recognition and Machine Learning (Information Science and Statistics) , 1st ed. Springer, 2007. 2 11

2007
[18]

K. P. Murphy, Machine learning : a probabilistic perspec- tive. Cambridge, Mass. [u.a.]: MIT Press, 2013. [Online]. Avail- able: https://www.amazon.com/Machine-Learning-Probabilistic-Perspective-Computation/ dp/0262018020/ref=sr_1_2?ie=UTF8&qid=1336857747&sr=8-2 2

2013
[19]

On the difficulty of training recurrent neural networks,

R. Pascanu, T. Mikolov, and Y . Bengio, “On the difficulty of training recurrent neural networks,” in Proceedings of the 30th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, S. Dasgupta and D. McAllester, Eds., vol. 28, no. 3. Atla...

2013
[20]

Captcha: Using hard ai problems for security,

L. von Ahn, M. Blum, N. J. Hopper, and J. Langford, “Captcha: Using hard ai problems for security,” inAdvances in Cryptology — EUROCRYPT 2003, E. Biham, Ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 2003, pp. 294–311. 2

2003
[21]

Geometrons - ghost symbols - buried information,

A. M. Atto, “Geometrons - ghost symbols - buried information,” 2024. [Online]. Available: https://dx.doi.org/10.21227/5112-ez52 2, 3

2024 doi
[22]

A. M. Atto, H. Hadhri, F. Vernier, and E. Trouvé, Multiclass Multilabel Change of State Transfer Learning from Image Time Series. John Wiley & Sons, Ltd, 2021, ch. 6, pp. 223–245. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/9781119882299.ch6 9 12

2021 doi

Pith tools

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

[1] [1]

Lessons from a dragon fly’s brain: Evolution built a small, fast, efficient neural network in a dragonfly. why not copy it for missile defense?

F. Chance, “Lessons from a dragon fly’s brain: Evolution built a small, fast, efficient neural network in a dragonfly. why not copy it for missile defense?” IEEE Spectrum, vol. 58, no. 8, pp. 28–33, 2021. 1

work page 2021

[2] [2]

Redefining neural architecture search of hetero- geneous multinetwork models by characterizing variation operators and model components,

U. Garciarena, R. Santana, and A. Mendiburu, “Redefining neural architecture search of hetero- geneous multinetwork models by characterizing variation operators and model components,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 8, pp. 10 561–10 575,

work page

[3] [3]

Zero-shot neural architecture search: Challenges, solutions, and opportunities,

G. Li, D. Hoang, K. Bhardwaj, M. Lin, Z. Wang, and R. Marculescu, “Zero-shot neural architecture search: Challenges, solutions, and opportunities,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 7618–7635, 2024. 1

work page 2024

[4] [4]

Weakly supervised text attention network for generating text proposals in scene images,

L. Rong, E. MengYi, L. JianQiang, and Z. HaiBin, “Weakly supervised text attention network for generating text proposals in scene images,” in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR) , vol. 01, 2017, pp. 324–330. 2

work page 2017

[5] [5]

Deep features class activation map for thermal face detection and tracking,

A. Kwa´sniewska, J. Rumi´nski, and P. Rad, “Deep features class activation map for thermal face detection and tracking,” in 2017 10th International Conference on Human System Interactions (HSI), 2017, pp. 41–47. 2

work page 2017

[6] [6]

Learning of counting crowded birds of various scales via novel density activation maps,

S. Kim and M. Kim, “Learning of counting crowded birds of various scales via novel density activation maps,” IEEE Access, vol. 8, pp. 155 296–155 305, 2020. 2

work page 2020

[7] [7]

Task-aware weakly supervised object localization with transformer,

M. Meng, T. Zhang, Z. Zhang, Y . Zhang, and F. Wu, “Task-aware weakly supervised object localization with transformer,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 7, pp. 9109–9121, 2023. 2

work page 2023

[8] [8]

Saliency as pseudo-pixel supervision for weakly and semi-supervised semantic segmentation,

M. Lee, S. Lee, J. Lee, and H. Shim, “Saliency as pseudo-pixel supervision for weakly and semi-supervised semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 10, pp. 12 341–12 357, 2023. 2

work page 2023