Pith. sign in

REVIEW 2 major objections 4 minor 24 references

A Structural Interpretation of GELU and Threshold-Transmission Activations via the First-Order Loss Function

T0 review · 2 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read GELU(z)=zΦ(z) is the expected output of a hard linear gate whose Gaussian threshold the input must clear.

desk verdict Correct restatement with a useful unifying table and honest but modest experiments — send to review. read the letter →

arxiv 2607.03664 v3 pith:6HJ33J3Q submitted 2026-07-04 cs.LG math.OCstat.ML

Roberto Rossi This is my paper
classification cs.LGmath.OCstat.ML
keywords GELUactivationfunctionsfirst-orderlossfunctionthreshold-transmissionhardswishuniformthresholdUELU/TUELUGaussianrandom
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 argues that GELU is best understood not just as an average over input-dependent Bernoulli dropouts but as the expected output of a hard linear gate with a Gaussian random threshold: a neuron transmits z exactly when z clears a latent threshold T~N(0,1), so GELU(z)=E[z 1{T≤z}] = zΦ(z). This makes GELU the signal-transmission component of the Gaussian complementary first-order loss function — an expected-surplus quantity from inventory theory — once the truncated-moment correction φ(z) is removed. The same threshold-transmission logic zF(z) yields ReLU, SiLU/Swish, and hard swish as special cases for different threshold laws, and a uniform threshold gives UELU, a piecewise-quadratic gate with an explicit boundary width β. A GELU-matched width β=√(π/2) and a learned-width variant stay competitive with standard activations on compact vision and language models, while the conventional hard-swish width β=3 leaves almost all preactivations inside the transition region.

What carries the argument

The Gaussian complementary first-order loss function bL(z)=E[(z−T)_+] for T∼N(0,1), which decomposes as bL(z)=φ(z)+zΦ(z). The paper treats zΦ(z) as the boundary-gated signal-transmission term and φ(z) as a truncated-moment loss-accounting correction; discarding the correction yields GELU, and replacing T's law with other distributions yields the family zF(z), with F the threshold CDF. For a uniform threshold on [−β,β], this gives UELUβ(z)=z·clip((z+β)/(2β),0,1), the piecewise-quadratic hard-swish-like gate whose half-width β parametrises the uncertain boundary layer.

What would settle it

Run the paper's best-performing UELU/TUELU configurations on a large-scale task (e.g., ImageNet or GPT-class language modelling) against GELU; if TUELU falls behind or if per-layer learned widths beat a shared β by a large margin, the single-width parametrisation and its GELU-matched calibration are not doing the claimed work.

Watch

Extended reading notes

Core claim

The paper's central claim is that GELU(z)=zΦ(z) equals E[z 1{T≤z}] for T∼N(0,1), so the activation is the expected transmission of a hard gate opened when the input clears a latent Gaussian threshold. This is not a new function but a new generative reading: the gate-open probability Φ(z) is the probability of threshold clearance, and the Bernoulli-dropout view becomes a consequence rather than a definition. The claim is derived from the identity bL(z)=E[(z−T)_+]=φ(z)+zΦ(z) for the Gaussian complementary first-order loss, so GELU = bL − φ; dropping φ converts an expected-surplus measure into an origin-preserving signal-transmission operator. Generalising the threshold law produces the activat

Load-bearing premise

The practical claims rest on assuming that a single global width β — initialised to match GELU's quadratic coefficient at zero — is an adequate parametrisation of threshold uncertainty across layers and tasks, and that the compact five-seed experimental settings are representative enough to judge it.

Editorial extensions

If this is right

  • GELU's negative dip near zero is explained as the boundary term zφ(z) in its derivative d/dz GELU(z)=Φ(z)+zφ(z), a signature of transmitting across an uncertain boundary rather than smoothing ReLU.
  • Hard swish is exactly UELU with β=3; calibrating β=√(π/2) matches GELU's local quadratic coefficient at the origin, giving a principled width instead of the ad-hoc '3'.
  • Learned shared width TUELU converges to architecture-dependent values (≈1.25 Mixer, 1.14 ViT, 0.70 char-GPT, 1.02 TinyStories, 0.89 WikiText-2), all far below β=3, and fixed/learned UELU matches or slightly beats GELU on these compact benchmarks.
  • Region-occupancy statistics show hard-swish's β=3 confines nearly all preactivations to the transition region, while the GELU-matched width creates nonzero closed/open regions, so the width parameter has operational meaning.
  • Retaining a centred remnant of the loss correction (DGELU) consistently underperforms, supporting the claim that the correction term should be discarded for activations.

Reading between the lines

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

  • The threshold-transmission construction suggests a principled recipe for converting any smooth gated activation into a piecewise-polynomial counterpart: pick the threshold law, match its CDF's local Taylor coefficients at the origin to set the width. This could be applied to logistic or heavy-tailed thresholds to generate new cheap activations with explicit transition widths.
  • Since learned widths vary by architecture, the shared-scalar β is likely a compromise; per-layer or per-channel threshold widths might push the family further, and the paper's own experiments leave that door open.
  • If the interpretation is taken literally, the Gaussian threshold is the maximum-entropy choice under mean-variance information; under bounded-support information the uniform threshold is the maximum-entropy choice, which gives a principled basis for choosing between GELU and UELU depending on what priors a network actually has about its preactivations.
  • The structural identity is parameter-free and can be tested at scale: if UELU's competitiveness persists on ImageNet-sized models, the single-width calibration is more than a small-scale artifact.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper develops a structural interpretation of GELU(z)=zΦ(z) as the expected output of a hard linear gate z·1{T≤z} with T∼N(0,1), and places this in the context of the complementary first-order loss function bL(z)=φ(z)+zΦ(z). It shows that GELU is the signal-transmission component after discarding the truncated-moment correction φ(z). This construction is extended to a family of threshold-transmission activations zF(z), including ReLU, GELU, SiLU/Swish, UELU (uniform threshold, recovering hard swish for β=3), and ExpELU. The paper introduces UELU with a GELU-calibrated width β=√(π/2) and TUELU with a learned shared width, and reports controlled experiments on compact vision and language models comparing these with GELU, ReLU, SiLU, and a DGELU variant. It also provides a universal approximation argument for the new activations.

Significance. The algebraic identities used here are standard partial-expectation results; the paper acknowledges that the threshold view does not change the expectation (Sec. 4). The value of the paper is in the unifying probabilistic framework, the explicit connection between GELU and inventory-theoretic loss functions, and the derivation of a hard-swish-like activation with an interpretable width. The paper is unusually transparent: it reports per-seed means and standard deviations, provides a code link, documents failures (DGELU and β=3), and explicitly identifies TUELU as a restricted subfamily of AQuLU. The universal approximation argument in Appendix A.4 is correct. If accepted as an interpretive and exploratory empirical contribution, the main claims are defensible. The principal weaknesses are the limited statistical power of the empirical comparisons and the risk that the generative threshold interpretation is over-read as empirically validated.

major comments (2)
  1. [Sec. 7, Tables 3–7] The text uses 'improve' for results whose confidence intervals overlap substantially. For example, Table 3 gives UELU 51.28±0.52 vs GELU 50.89±0.66, a difference well within one standard deviation; Table 7 shows UELU and TUELU essentially tied with GELU. The abstract's 'consistently competitive' is acceptable, but section-level wording (Secs. 7.1, 7.2, 7.4) overstates what the reported precision supports. Please report paired per-seed differences or state that mean differences are not statistically significant, and adjust the wording accordingly.
  2. [Sec. 8 / Abstract] The paper presents the empirical results as evidence for the threshold-transmission framework, but because TUELU is a one-parameter restriction of AQuLU, the experiments cannot distinguish the generative mechanism from a constrained piecewise-quadratic activation. The paper itself notes this in Sec. 6, but the abstract and conclusion should be qualified so that the competitiveness claim is attributed to the restricted family rather than to the latent-threshold interpretation. This is not an error in the mathematics, but it is important for setting the reader's expectation about what the experiments establish.
minor comments (4)
  1. [Sec. 7] The dataset name 'CIF AR-100' appears throughout Sec. 7; it should be 'CIFAR-100'.
  2. [Table 3 caption] The 'Regions C/T/O' column is not defined in the caption. Please spell out 'closed/transition/open' in the caption or in a footnote.
  3. [Sec. 4] The sentence 'The threshold view does not change the expectation' is an important caveat. It would help to explicitly add that the 'generative' story is a reparametrization of the existing Bernoulli gate, not a new stochastic process, so readers do not misattribute the empirical results to a novel mechanism.
  4. [Appendix A.2] The decomposition for ExpELU is algebraically correct, but the phrase 'signal transmission minus truncated-moment accounting' is slightly confusing because the bracket is positive while the correction as defined elsewhere is negative. Consider aligning notation with the Gaussian case where the correction is explicitly -E[T 1{T≤z}].

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivations are self-contained algebraic identities; the threshold view is an openly acknowledged equivalent restatement, and the empirical width calibration is a disclosed fit, not a disguised prediction.

full rationale

The paper's mathematical core is self-contained and correctly derived. The complementary loss identity bL(z) = E[(z-T)_+] = zPhi(z) + phi(z) is obtained by direct partial-moment computation, and GELU(z) = bL(z) - phi(z) follows algebraically. The threshold statement GELU(z) = E[z 1{T<=z}] with T ~ N(0,1) is true by construction because P(T<=z) = Phi(z); the paper itself says 'The threshold view does not change the expectation' (Sec. 4), so it is presented as an interpretation, not as an independent prediction. No load-bearing self-citation occurs: the only self-citation (ref. [17], Rossi et al.) merely names the 'complementary function' and is not used to establish the paper's claims; the closed forms are proven in the text. The empirical parameter beta = sqrt(pi/2) is transparently calibrated by matching GELU's quadratic coefficient at the origin (Sec. 7), and the paper reports actual experiments rather than deriving competitiveness from that fit. It also discloses its limitations, including per-layer or per-channel widths and full AQuLU as future work. Accordingly, there is no circular step: the equivalences are acknowledged restatements, and the empirical claims are tested, not assumed.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

The central claim rests only on elementary calculus identities for the Gaussian distribution and standard known results in stochastic inventory theory. The empirical portion adds a fitted width parameter β (calibrated to GELU) and a modeling assumption that activations are threshold-transmission zF(z) with a shared latent distribution. No independent evidence for the latent threshold exists beyond the activation shapes it reproduces.

free parameters (2)
  • β (uniform-threshold half-width of UELU/TUELU) = 1.25 fixed (β=√(π/2)); learned in TUELU to values ≈1.25, 1.14, 0.70, 1.02, 0.89 across tasks
    Chosen to match the quadratic coefficient of GELU at the origin (Sec. 7); learned by gradient descent in TUELU with the same initialization.
  • λ (rate of ExpELU) = not fixed empirically (used only as contrast case, Fig. 2, λ=1)
    A one-sided activation suggested via the exponential threshold; not part of the empirical comparisons.
assumptions (5)
  • standard math Gaussian complementary first-order loss identity: bL(z)=E[(z-Z)_+]=φ(z)+zΦ(z) for Z~N(0,1).
    Used in Sec. 2 to identify GELU = bL - φ; standard result (Zipkin).
  • standard math Truncated moment integral ∫_{-∞}^z tφ(t)dt = -φ(z).
    Used to expand bL(z) in Sec. 2.
  • standard math Maximum-entropy characterization of normal/uniform/exponential laws under moment-or-support constraints.
    Motivates Gaussian and uniform threshold laws in Secs. 4–5; standard information-theoretic results (Jaynes).
  • standard math Universal approximation holds for non-polynomial, locally bounded, piecewise continuous activations (Leshno et al.).
    Applied in Appendix A.4 to show UELU/TUELU retain universal approximation.
  • ad hoc to paper The activation is modeled as the transmitted signal of a latent threshold process, a_F(z)=zF(z), with the threshold distribution shared across units/layers.
    This is the paper's construction; it is a definitional modeling choice rather than an empirical finding, and it is what conflates GELU with a random-threshold gate.
invented entities (1)
  • latent threshold T with distribution F (Gaussian, logistic, uniform, exponential)
    purpose: Introduces a generative mechanism that explains the Bernoulli gate as threshold clearance and generates activation functions through zF(z).
    The threshold is not observed; the same activation functions arise from the original Bernoulli-gate definition without any latent variable. No falsifiable handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Structural Interpretation of GELU and Threshold-Transmission Activations via the First-Order Loss Function." pith.science (2026). https://pith.science/paper/6HJ33J3Q

@misc{pith2026260703664,
  author       = {Pith},
  title        = {Pith review of: A Structural Interpretation of GELU and Threshold-Transmission Activations via the First-Order Loss Function},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6HJ33J3Q}},
  note         = {Machine review of arXiv:2607.03664}
}
read the original abstract

The Gaussian Error Linear Unit is usually motivated as the expected output of an input-dependent Bernoulli gate. This work gives an alternative interpretation: GELU is the expected output of a hard linear gate with a Gaussian random threshold. This view provides a generative interpretation for the Bernoulli gate: the gate opens once the input clears a latent Gaussian threshold. This interpretation stems from a decomposition based on well-known results in stochastic inventory theory and leads to a threshold-transmission family that includes ReLU, GELU, SiLU/Swish, and hard swish as special cases. By considering a latent uniform threshold, we recover a hard-swish-like piecewise-polynomial gate whose nonlinear transition is confined to a finite interval, yielding fixed- and learned-width variants. Controlled experiments on compact vision and language models show that calibrated or learned uniform-threshold gates are consistently competitive with GELU, ReLU, and SiLU/Swish, display architecture-dependent learned widths, and use the finite transition region nontrivially.

Figures

Figures reproduced from arXiv: 2607.03664 by the authors.

Figure 1
Figure 1. The complementary loss is positive at the origin, the centred complementary loss has [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Two threshold-transmission examples. Left: ExpELU for [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 2
Figure 2. Two threshold-transmission examples. Left: ExpELU for [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figures from the paper (11 more)
Figure 3
Figure 3. Figure 3: MLP-Mixer results. Left: validation accuracy over training. Right: learned shared [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 3
Figure 3. Figure 3: MLP-Mixer results. Left: validation accuracy over training. Right: learned shared [PITH_FULL_IMAGE:figures/full_fig_p009_3.png]
Figure 4
Figure 4. Figure 4: Compact Vision Transformer results. Left: validation accuracy over training. Right: [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 4
Figure 4. Figure 4: Compact Vision Transformer results. Left: validation accuracy over training. Right: [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Tiny GPT results. Left: validation perplexity over training. Right: learned shared [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 5
Figure 5. Figure 5: Tiny GPT results. Left: validation perplexity over training. Right: learned shared [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: TinyStories GPT results. Left: validation perplexity over training. Right: learned [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: WikiText-2 GPT results. Left: validation perplexity over training. Right: learned [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 7
Figure 7. Figure 7: WikiText-2 GPT results. Left: validation perplexity over training. Right: learned [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Closed, transition, and open region occupancy for fixed-width UELU and TUELU [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 8
Figure 8. Figure 8: Closed, transition, and open region occupancy for fixed-width UELU and TUELU [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Review history (3 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 3 linked inside Pith

  1. [1]

    Lei Jimmy Ba and Brendan J. Frey. Adaptive dropout for training deep neural networks. In Christopher J. C. Burges, L´ eon Bottou, Zoubin Ghahramani, and Kilian Q. Weinberger, editors,Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings of a meeting held December 5-8, 2013, ...

    2013

  2. [2]

    Birge and Fran¸ cois Louveaux.Introduction to Stochastic Programming

    John R. Birge and Fran¸ cois Louveaux.Introduction to Stochastic Programming. Springer New York, 2011. 16 Phrase Mathematical object Meaning Input signalzThe preactivation value to be transmitted or suppressed. Gaussian random thresholdT∼N(0,1) The uncertain boundary that the input must exceed. Hard threshold event{T≤z}The event that the input clears the ...

    2011

  3. [3]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In9th International Conference on Learning Representations, ICLR 2021, V...

    2021

  4. [4]

    Activation functions in deep learning: A comprehensive survey and benchmark.Neurocomputing, 503:92–108, September 2022

    Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark.Neurocomputing, 503:92–108, September 2022

    2022

  5. [5]

    Tinystories: How small can language models be and still speak coherent english?CoRR, abs/2305.07759, 2023

    Ronen Eldan and Yuanzhi Li. Tinystories: How small can language models be and still speak coherent english?CoRR, abs/2305.07759, 2023

    arXiv 2023

  6. [6]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, November 2018

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, November 2018

    2018

  7. [7]

    Whitin.Analysis of Inventory Systems

    George Hadley and Thomson M. Whitin.Analysis of Inventory Systems. Prentice-Hall, Englewood Cliffs, NJ, 1963

    1963

  8. [8]

    Gaussian Error Linear Units (GELUs).CoRR, abs/1606.08415, 2016

    Dan Hendrycks and Kevin Gimpel. Gaussian Error Linear Units (GELUs).CoRR, abs/1606.08415, 2016

    arXiv 2016

Show all 24 references
  1. [9]

    Searching for MobileNetV3

    Andrew Howard, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingxing Tan, Grace Chu, Vijay Vasudevan, Yukun Zhu, Ruoming Pang, Hartwig Adam, and Quoc Le. Searching for MobileNetV3. In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1314–1324. IE...

    2019

  2. [10]

    Jaynes.Probability theory: The logic of science

    Edwin T. Jaynes.Probability theory: The logic of science. Cambridge University Press, Cambridge, 2003

    2003

  3. [11]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, Toronto, Ontario, 2009

    2009

  4. [12]

    Three decades of activations: A comprehensive survey of 400 activation functions for neural networks.CoRR, abs/2402.09092, 2024

    Vladim ´ ır Kunc and Jir ´ ı Kl´ ema. Three decades of activations: A comprehensive survey of 400 activation functions for neural networks.CoRR, abs/2402.09092, 2024

    2024 arXiv

  5. [13]

    Lin, Allan Pinkus, and Shimon Schocken

    Moshe Leshno, Vladimir Ya. Lin, Allan Pinkus, and Shimon Schocken. Multilayer feedfor- ward networks with a nonpolynomial activation function can approximate any function. Neural Networks, 6(6):861–867, January 1993. 17

    1993

  6. [14]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In7th Interna- tional Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019

    2019

  7. [15]

    Pointer sentinel mixture models

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017

    2017

  8. [16]

    Vinod Nair and Geoffrey E. Hinton. Rectified linear units improve restricted boltzmann machines. In Johannes F¨ urnkranz and Thorsten Joachims, editors,Proceedings of the 27th International Conference on Machine Learning (ICML-10), June 21-24, 2010, Haifa, Israel, pages 807–81...

    2010

  9. [17]

    Armagan Tarim, Steven Prestwich, and Brahim Hnich

    Roberto Rossi, S. Armagan Tarim, Steven Prestwich, and Brahim Hnich. Piecewise linear lower and upper bounds for the standard normal first order loss function.Applied Mathematics and Computation, 231:489–502, March 2014

    2014

  10. [18]

    Neural machine translation of rare words with subword units

    Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In Katrin Erk and Noah A. Smith, editors,Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715–1...

    2016

  11. [19]

    Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958, 2014

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhut- dinov. Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958, 2014

    1929

  12. [20]

    Rethinking the Inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the Inception architecture for computer vision. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818–2826. IEEE, June 2016

    2016

  13. [21]

    Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-Mixer: An all-MLP architecture for vision. In Marc’Aurelio Ranzato, A...

    2021

  14. [22]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, ...

    2017

  15. [23]

    Z. Wu, H. Yu, L. Zhang, and Y. Sui. The Adaptive Quadratic Linear Unit (AQuLU): Adaptive non monotonic piecewise activation function.Tehnicki vjesnik - Technical Gazette, 30(5):1469–1485, 2023

    2023

  16. [24]

    Irwin Professional Publishing, Maiden- head, England, April 2000

    Paul Zipkin.Foundations of inventory management. Irwin Professional Publishing, Maiden- head, England, April 2000. 18

    2000

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.