Efficient and Interpretable Neural Networks Using Complex Lehmer Transform

Masoud Ataei; Xiaogang Wang

REVIEW 4 major objections 5 minor 24 references

Efficient and Interpretable Neural Networks Using Complex Lehmer Transform

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that one layer of Lehmer-mean activation units can match or exceed standard deep models while staying interpretable, the paper argues.

desk verdict The mathematical framing of a classical weighted mean is fine, but the experimental section is internally inconsistent and cannot support the state-of-the-art claims. read the letter →

arxiv 2501.15223 v1 pith:YJTBWML7 submitted 2025-01-25 cs.LG cs.AI

Masoud Ataei , Xiaogang Wang This is my paper · ORCID

classification cs.LGcs.AI MSC 68T0726E60

keywords LehmerTransformNeuralNetworkInterpretabilityComplex-ValuedNetworksNonlinearActivationFunctionsWeightedFeatureSelectionPhase-SensitiveAggregationSingle-LayerArchitecture

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes making the Lehmer transform, a parameterized family of means, the activation function of a neural network. The authors claim that a single layer of these Lehmer activation units (LAUs) reaches competitive or better accuracy than standard models on Iris, Wine, Wisconsin Breast Cancer, and MNIST, while needing far fewer parameters and remaining transparent about how features are combined. A weighted form supplies trainable feature importance, and a complex-valued form adds oscillatory, phase-sensitive behavior that the paper says helps on structured or high-dimensional data. If the central claim holds, cheap and interpretable networks would not require deep stacks to be competitive on common benchmarks.

What carries the argument

The central object is the weighted Lehmer transform $L(s;x,w)=\sum_i w_i x_i^s / \sum_i w_i x_i^{s-1}$, used as an activation unit. The trainable exponent $s$ is called the suddency moment, and it lets the unit decide whether to emphasize small or large input values; the complex extension $s=a+bi$ adds oscillatory factors $e^{bi\ln x_i}$, which the paper interprets as phase-sensitive aggregation. Positivity of the weights is maintained by $w_i=\ln(1+e^{v_i})$, and the complex output is converted to a real number by a trainable affine combination of its real and imaginary parts. This combination is what carries both the efficiency claim (one layer, few parameters) and the interpretability claim (each unit is literally a weighted mean of its inputs).

What would settle it

On the standard MNIST train/test split, run the exact architecture the paper states — one dense layer of 3 LAU neurons followed by a softmax layer, with no convolutional layers — and check whether accuracy lands near 97–98%; because 3 LAU neurons cannot produce the 10 scores a softmax needs, the run cannot even be set up as described, and any reported accuracy must come from a different, currently unspecified configuration. Identifying that true configuration and rerunning it would settle whether the central empirical claim is reproducible.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a trainable weighted mean is enough to serve as a powerful nonlinear activation: a single dense layer of real-valued or complex-valued LAUs is reported to reach 95% on Iris, 95–99% on Wine, 94% on WBC, and 97–98% on MNIST. The real-valued LAU interpolates between weighted harmonic, arithmetic, and contra-harmonic means through the trainable exponent $s$, and the complex-valued variant lets $s=a+bi$, so the imaginary part produces phase-like oscillations through $\cos(b\ln x_i)$ and $\sin(b\ln x_i)$ before an affine combination of real and imaginary parts converts the output back to a real number. The paper also establishes structural properties of the transform — homogeneity of degree one, invariance to proportional weight scaling, permutation invariance, Schur convexity, and differentiability — which make the unit a principled aggregator rather than an ad-hoc nonlinearity. The empirical section claims that with a single dense layer (plus convolutional feature extraction for MNIST) these units achieve the reported accuracies, with the complex version doing best on MNIST.

Load-bearing premise

The load-bearing premise is that the reported benchmark numbers come from a definite, reproducible architecture; the paper says every dataset used a single dense layer of 3 neurons before a softmax, yet MNIST has 10 classes, so the description cannot be literally true and the accuracy claims rest on an unspecified model.

Editorial extensions

If this is right

If the central claim holds, tabular benchmarks like Iris, Wine, and WBC would need only a single dense layer of LAUs, cutting parameter counts by orders of magnitude compared with multi-layer networks.
Complex-valued LAUs would give a built-in mechanism for periodic and phase-sensitive patterns, so time-series and signal-processing tasks could use the same transparent architecture without extra feature engineering.
Because each unit is a weighted mean, trained weights and the suddency moment directly show which inputs dominate and whether large or small values are emphasized, making the model's decisions inspectable.
The transform's homogeneity, invariance, and Schur convexity imply predictable behavior under rescaling and reordering of inputs, which simplifies deployment and robustness checks.
On MNIST, a single LAU dense layer after convolutional feature extraction would show that deep stacks are not required for high-dimensional image classification if the reported 97–98% accuracy is reproducible.

Reading between the lines

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

The paper leaves implicit that a linear baseline would sharpen the claim; a single weighted mean with $s$ near 1 is close to linear, so comparing against logistic regression would isolate the nonlinear contribution.
A test the authors do not run is a dataset with periodic or rotated class structure; that would show whether the complex oscillatory term is essential for the reported phase-sensitive advantage.
The 3-neuron MNIST description is internally inconsistent with a 10-class softmax; readers should treat the MNIST numbers as conditional on an unspecified architecture until the true layer width is stated.

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

Colleague,

The honest headline: this is a classical mean (weighted Lehmer) with a complex-exponent twist and a trainable activation built around it. The formal parts are mostly fine; the empirical core is not. The reader's take is close to mine, and the stress-test concern is real: a 3-neuron dense layer cannot produce 10 logits for MNIST or 2 for WBC, and the added conv layers for MNIST contradict the "single layer" claim. Table 1 does not support the abstract's "state-of-the-art" headline.

What's actually new: using the weighted Lehmer mean as a trainable activation, and the complex suddency moment s = a + bi with its oscillatory terms. That is a legitimate incremental idea, and the mathematical properties—homogeneity, permutation invariance, Schur convexity—are stated correctly and follow from standard mean theory (Bullen). The positivity reparameterization w_i = log(1 + e^{v_i}) is sensible. The self-citation to [2] is fair; they did introduce the Lehmer transform in an EEG context, so building on it is not a problem.

Where it falls apart: (1) The experimental section has no baselines, no runtime measurements, no hyperparameters, no error bars for MNIST, and no code. The claim "significantly improved computational efficiency" is never measured. (2) The architecture is internally inconsistent, as the stress-test notes. (3) Eq (14) misstates the Wirtinger derivative: ∂L/∂s is not ∂L/∂a + i ∂L/∂b; that combination equals 2∂L/∂\bar{s} (for holomorphic L, zero). This is a math error, though it may not invalidate the training scheme if they use real-valued backprop on a and b. (4) The "interpretability" claims are asserted, not demonstrated—no analysis of learned weights or suddency parameters on any dataset.

For whom: someone working on alternative activations for small tabular data might find this an interesting seed, but the paper needs a major rewrite of the experiments and a fix of Eq (14) before it is publishable. I would not cite it as-is. A serious referee could ask for baselines, architecture clarity, and an honest single-layer comparison; I'd send it out rather than desk reject, but with a clear request for reanalysis.

Referee Report

4 major / 5 minor

Summary. The manuscript introduces the weighted Lehmer transform as a neural-network activation function, including a complex-valued extension with a complex suddency moment s = a + bi. It derives algebraic properties such as homogeneity, weight-scale invariance, permutation invariance, Schur convexity, and limiting behavior, then defines Lehmer Activation Units (LAUs) with trainable weights and an affine post-processing step for the complex output. Section 4 reports accuracies on Iris, Wine, WBC, and MNIST, and the abstract and conclusion claim that a single layer of LAUs delivers state-of-the-art accuracy while improving interpretability and computational efficiency.

Significance. If the empirical and theoretical claims were fully supported, a trainable mean-like activation with an oscillatory complex extension could be a convenient building block for compact networks. I verified several of the algebraic identities in Section 2 (Eqs. (2), (9), (11)) and they are correct; the mathematical sections are clearly written. However, the significance of the paper as submitted is sharply limited: the central empirical claim of state-of-the-art performance is not backed by any comparisons, the MNIST and WBC architecture descriptions are internally inconsistent, the efficiency claim has no runtime measurements, and the complex-differentiability statement in Eq. (14) is incorrect. No code or reproducibility artifact is provided, so the reported numbers cannot be checked independently.

major comments (4)

[Section 4, Table 1] The architecture description is internally inconsistent and cannot reproduce the reported results. The text states that 'for all datasets' the architecture used 'a single dense layer with 3 neurons, followed by a softmax layer,' but MNIST has 10 classes and WBC has 2 classes, so a softmax over 3 logits cannot produce the required outputs. The immediately following sentence adds convolutional layers for MNIST, which contradicts the 'single layer' claim and still does not explain how 3 LAU neurons yield 10 class scores. The 97% and 98% MNIST accuracies in Table 1 are therefore not reproducible from the described setup, and the central single-layer state-of-the-art claim is unverifiable as stated.
[Table 1] No baseline models are reported anywhere in Section 4. The abstract's claims of 'state-of-the-art performance' and 'competitive accuracy' cannot be assessed without comparisons to standard classifiers such as logistic regression, SVM, MLP, a standard CNN, or KAN. Likewise, the claim of 'significantly improved computational efficiency' is unsupported because no runtime, parameter count, or FLOP comparison is given. In addition, MNIST has no error bars or multiple-run statistics, so the difference between 97% and 98% may be within run-to-run variability.
[Section 2.3, Eq. (14)] Equation (14) states that the derivative with respect to s = a + bi is given by ∂L/∂s = ∂L/∂a + i∂L/∂b, but this is not the Wirtinger derivative. For a holomorphic function L(s), the correct relations are ∂L/∂s = ∂L/∂a = -i∂L/∂b, and the expression ∂L/∂a + i∂L/∂b vanishes identically when the Cauchy-Riemann equations hold. Since differentiability and gradient-based optimization are central to the complex LAU proposal, this statement needs to be corrected and justified.
[Sections 3 and 4] The claimed capabilities of 'adaptive feature selection,' 'hierarchical feature aggregation,' and 'phase-sensitive' modeling are asserted from the algebraic form of Eqs. (12) and (16) rather than demonstrated by experiments or defined operationally. The paper should provide a precise interpretation of what the learned weights and suddency parameters mean for a trained model, or include ablations and feature-importance analyses that support these claims, or explicitly temper these claims to what the mathematics alone establishes.

minor comments (5)

[Section 2.1] The term 'suddency moment' is used without a definition; it should be formally introduced when first used, and its connection to the parameter s in the Lehmer mean should be stated explicitly.
[Section 2.3] If Wirtinger calculus is intended, the operators ∂/∂s and ∂/∂s̄ should be defined explicitly, and Eq. (14) should be rewritten with the correct factors and conjugation conventions.
[Section 3] The paper says inputs are standardized to the range (e^{-1}, e), but the MNIST pipeline in Section 4 uses convolutional, batch-normalization, and max-pooling layers; the authors should specify how positivity of the inputs to the LAU is guaranteed after these operations.
[Table 1] Standard deviations are reported for Iris, Wine, and WBC but not for MNIST; the number of independent runs and the random seeds should be stated for all datasets.
[Abstract and Section 4] The abstract claims 'significantly improved computational efficiency,' but no runtime or complexity comparison appears in Section 4; either provide such measurements or remove the efficiency claim from the abstract.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central contribution is a new activation function. The math dependencies are mostly standard mean theory, but the empirical claims depend on trainable parameters and unstated experimental choices. No external benchmark or falsifiable prediction outside the reported accuracies is provided.

free parameters (5)

suddency moment s (real LAU) = trained (not reported)
The exponent in Eq (15) controls whether the unit behaves like harmonic, arithmetic, or contra-harmonic mean; its learned value determines the aggregation behavior.
complex suddency (a, b) = trained (not reported)
a and b in Eq (16) set magnitude aggregation and oscillation frequency; the claimed phase-sensitivity depends on fitted b.
input weights w_i = trained via wi = ln(1+exp(vi))
Positive weights determine feature importance; central to the weighted transform.
affine mapping parameters (alpha, beta, gamma) = trained
Eq (17) maps the complex LAU output to a real value; required for compatibility with softmax.
input standardization range = (e^{-1}, e)
Chosen by hand to keep powers and logarithms well-conditioned; not learned from data.

assumptions (4)

domain assumption Inputs and weights are strictly positive.
Eqs (4) and (12) require x_i > 0 and w_i > 0 for powers and logarithms; Section 3 says inputs are standardized to (e^{-1}, e) and weights are parameterized via wi = ln(1+e^{v_i}).
standard math The weighted Lehmer mean is monotone and Schur-convex.
Section 2 cites Bullen [24]; the paper relies on these properties for its interpretability and aggregation claims but does not prove them.
domain assumption The complex-valued transform is differentiable in a way compatible with gradient descent.
Section 2.3 claims differentiability and gives Eq (14) as the Wirtinger derivative, but Eq (14) is not the standard Wirtinger derivative, so the optimization guarantee is questionable.
ad hoc to paper A single dense layer with 3 neurons suffices for classification on the tested datasets.
Section 4 fixes 3 neurons for all datasets without justification; for 10-class MNIST the mapping from 3 neurons to 10 logits is unexplained.

invented entities (1)

Complex-valued Lehmer transform with suddency moment s = a + bi
purpose: Introduces oscillatory, phase-sensitive aggregation into the activation unit.
Defined in Eq (12) by substituting complex s into the weighted Lehmer mean; no independent benchmark or physical prediction validates the claimed 'phase-sensitive' benefit.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient and Interpretable Neural Networks Using Complex Lehmer Transform." pith.science (2026). https://pith.science/paper/YJTBWML7

@misc{pith2026250115223,
  author       = {Pith},
  title        = {Pith review of: Efficient and Interpretable Neural Networks Using Complex Lehmer Transform},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJTBWML7}},
  note         = {Machine review of arXiv:2501.15223}
}

read the original abstract

We propose an efficient and interpretable neural network with a novel activation function called the weighted Lehmer transform. This new activation function enables adaptive feature selection and extends to the complex domain, capturing phase-sensitive and hierarchical relationships within data. Notably, it provides greater interpretability and transparency compared to existing machine learning models, facilitating a deeper understanding of its functionality and decision-making processes. We analyze the mathematical properties of both real-valued and complex-valued Lehmer activation units and demonstrate their applications in modeling nonlinear interactions. Empirical evaluations demonstrate that our proposed neural network achieves competitive accuracy on benchmark datasets with significantly improved computational efficiency. A single layer of real-valued or complex-valued Lehmer activation units is shown to deliver state-of-the-art performance, balancing efficiency with interpretability.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages

[2]

Scientiﬁc Reports 12(1), 3663 (2022)

Ataei, M., Wang, X.: Theory of lehmer transform and its application s in identify- ing the electroencephalographic signature of major depressive dis order. Scientiﬁc Reports 12(1), 3663 (2022)

work page 2022
[1]

arXiv pr eprint arXiv:2404.19756 (2024)

Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljaˇ ci´ c, M., Hou, T.Y., Tegmark, M.: Kan: Kolmogorov-arnold networks. arXiv pr eprint arXiv:2404.19756 (2024)

arXiv 2024
[3]

In: International Conference on Learning Representations (ICLR) (2015)

Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by j ointly learning to align and translate. In: International Conference on Learning Representations (ICLR) (2015)

work page 2015
[4]

arXiv p reprint arXiv:1312.6034 (2013)

Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional n et- works: Visualising image classiﬁcation models and saliency maps. arXiv p reprint arXiv:1312.6034 (2013)

arXiv 2013
[5]

Advances in Neural Information Processing Systems 31 (2018)

Alvarez-Melis, D., Jaakkola, T.S.: Towards interpretable neural n etworks. Advances in Neural Information Processing Systems 31 (2018)

work page 2018
[6]

In: International Conference on Machine Learning (ICML) (2018)

Kim, B., Wattenberg, M., Gilmer, J., Cai, C., Wexler, J., Viegas, F., Say res, R.: Interpretability beyond feature attribution: Quantitative tes ting with con- cept activation vectors (tcav). In: International Conference on Machine Learning (ICML) (2018)

work page 2018
[7]

In: International Conference on Learning Represen tations (ICLR) (2017)

Kipf, T.N., Welling, M.: Semi-supervised classiﬁcation with graph conv olutional networks. In: International Conference on Learning Represen tations (ICLR) (2017)

work page 2017
[8]

Nature Machine Intelligence 1, 206–215 (2019)

Rudin, C.: Stop explaining black box models for high stakes decisions . Nature Machine Intelligence 1, 206–215 (2019)

work page 2019

Show all 24 references

[9]

arXiv preprint arXiv:1702.08608 (2017)

Doshi-Velez, F., Kim, B.: Towards a rigorous science of interpreta ble machine learning. arXiv preprint arXiv:1702.08608 (2017)

2017 arXiv
[10]

: Explaining nonlinear classiﬁers with layer-wise relevance propagation

Montavon, G., et al. : Explaining nonlinear classiﬁers with layer-wise relevance propagation. Digital Signal Processing 73, 1–15 (2017)

2017
[11]

: Why should i trust you? explaining predictions of machine learning models

Ribeiro, M.T., et al. : Why should i trust you? explaining predictions of machine learning models. In: Proceedings of KDD (2016)

2016
[12]

arXiv preprint arXiv:1 706.03762 (2017)

Vaswani, A., et al.: Attention is all you need. arXiv preprint arXiv:1 706.03762 (2017)

2017
[13]

: Explaining explanations: Interpretability of machine learning

Gilpin, L., et al. : Explaining explanations: Interpretability of machine learning. In: DSAA (2018) 13

2018
[14]

: Explainable machine learning for scientiﬁc insights and discoveries

Roscher, R., et al. : Explainable machine learning for scientiﬁc insights and discoveries. IEEE Access 8, 4227–4247 (2020)

2020
[15]

I EEE Trans- actions on Neural Networks and Learning Systems 32, 4793–4813 (2020)

Tjoa, E., Guan, C.: A survey on explainable artiﬁcial intelligence. I EEE Trans- actions on Neural Networks and Learning Systems 32, 4793–4813 (2020)

2020
[16]

John Wiley & Sons, ??? (2012)

Hirose, A., Yoshida, S.: Complex-valued Neural Networks: Adva nces and Appli- cations. John Wiley & Sons, ??? (2012)

2012
[17]

arXiv preprint arXiv:1 705.09792 (2018)

Trabelsi, C., et al.: Deep complex networks. arXiv preprint arXiv:1 705.09792 (2018)

2018
[18]

Biome dical Signal Processing and Control 85, 104862 (2023)

Du, H., Riddell, R.P., Wang, X.: A hybrid complex-valued neural netw ork framework with applications to electroencephalogram (eeg). Biome dical Signal Processing and Control 85, 104862 (2023)

2023
[19]

: Complex-valued Neural Networks and Wirtinger Calculus

Kreutz-Delgado, K., et al. : Complex-valued Neural Networks and Wirtinger Calculus. Springer, ??? (2009)

2009
[20]

Neural Computation 33, 900–923 (2021)

Zhang, L., Li, W.: Learning complex-valued representations for phase-sensitive data. Neural Computation 33, 900–923 (2021)

2021
[21]

Journal of Financial Data Science (2022)

Schober, A., et al.: Application of cvnns in ﬁnancial time series pre diction. Journal of Financial Data Science (2022)

2022
[22]

IEEE Transactions on Geoscience and Remote Sensing 58, 4237–4246 (2020)

Chen, J., Qian, Z.: Complex-valued neural networks for polsar im age classiﬁ- cation. IEEE Transactions on Geoscience and Remote Sensing 58, 4237–4246 (2020)

2020
[23]

IEEE Transactions on Signal Processing 69, 3567–3579 (2021)

Amin, F., Yadav, A.: Complex-valued neural networks for signal processing. IEEE Transactions on Signal Processing 69, 3567–3579 (2021)

2021
[24]

Bullen, P.S.: Handbook of Means and Their Inequalities vol. 560. Sp ringer, ??? (2013) 14

2013

Pith tools

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

[1] [2]

Scientiﬁc Reports 12(1), 3663 (2022)

Ataei, M., Wang, X.: Theory of lehmer transform and its application s in identify- ing the electroencephalographic signature of major depressive dis order. Scientiﬁc Reports 12(1), 3663 (2022)

work page 2022

[2] [1]

arXiv pr eprint arXiv:2404.19756 (2024)

Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljaˇ ci´ c, M., Hou, T.Y., Tegmark, M.: Kan: Kolmogorov-arnold networks. arXiv pr eprint arXiv:2404.19756 (2024)

arXiv 2024

[3] [3]

In: International Conference on Learning Representations (ICLR) (2015)

Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by j ointly learning to align and translate. In: International Conference on Learning Representations (ICLR) (2015)

work page 2015

[4] [4]

arXiv p reprint arXiv:1312.6034 (2013)

Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional n et- works: Visualising image classiﬁcation models and saliency maps. arXiv p reprint arXiv:1312.6034 (2013)

arXiv 2013

[5] [5]

Advances in Neural Information Processing Systems 31 (2018)

Alvarez-Melis, D., Jaakkola, T.S.: Towards interpretable neural n etworks. Advances in Neural Information Processing Systems 31 (2018)

work page 2018

[6] [6]

In: International Conference on Machine Learning (ICML) (2018)

Kim, B., Wattenberg, M., Gilmer, J., Cai, C., Wexler, J., Viegas, F., Say res, R.: Interpretability beyond feature attribution: Quantitative tes ting with con- cept activation vectors (tcav). In: International Conference on Machine Learning (ICML) (2018)

work page 2018

[7] [7]

In: International Conference on Learning Represen tations (ICLR) (2017)

Kipf, T.N., Welling, M.: Semi-supervised classiﬁcation with graph conv olutional networks. In: International Conference on Learning Represen tations (ICLR) (2017)

work page 2017

[8] [8]

Nature Machine Intelligence 1, 206–215 (2019)

Rudin, C.: Stop explaining black box models for high stakes decisions . Nature Machine Intelligence 1, 206–215 (2019)

work page 2019