{"id":"76063e5c-c25f-44a5-bd3a-26b0693bcf71","arxiv_id":"2501.15223","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Lehmer activation units, a trainable weighted-mean activation with a complex-exponent extension, report 94-99% accuracy on Iris, Wine, WBC and MNIST, but with no baseline comparisons.","lead":"The paper proposes a new neural-network building block, the Lehmer activation unit, a trainable weighted mean that can also use complex exponents to add oscillatory behavior. The authors report good accuracy on small benchmarks and claim it is more interpretable and faster, but they provide no baseline comparisons or code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"A 3-neuron LAU dense layer cannot produce 10-class MNIST logits, and the added conv layers contradict the claimed single-layer model; Table 1 cannot support the central claim.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern I would put first: the described architecture cannot instantiate the claimed model. A 3-neuron LAU dense layer feeding a softmax cannot produce 10 MNIST logits, and the separate mention of convolutional layers contradicts the 'single layer' wording in the abstract. This is an internal-consistency problem, not a disagreement with external consensus, and it directly blocks reproduction of the paper's strongest empirical result. I do not see a fatal flaw in the mathematical definitions themselves: the weighted Lehmer transform and its complex extension are coherent, and the derivative identities in Section 2 are standard. But the empirical section is the only support for the headline claims, and it fails at the level of basic architecture specification. The absence of baselines, runtime measurements, and code further weakens 'state-of-the-art' and 'significantly improved computational efficiency,' but the architecture gap is decisive. Therefore the reader's REJECT verdict should stand unchanged.","tokens_in":9299,"tokens_out":5737,"duration_ms":53428,"concrete_test":"Request the exact model definition or code for the MNIST experiment, then re-run the network literally described in Section 4: convolutional feature extraction followed by a dense layer of 3 LAU neurons and a softmax output. Verify where the 10 class logits are produced. If the softmax has only 3 units, the model cannot classify MNIST at all; if a separate 10-unit output layer exists, disclose it and re-evaluate whether the 'single dense layer' claim still holds. Also compare the re-run accuracy against Table 1 after including all unstated layers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single layer of Lehmer Activation Units delivers state-of-the-art accuracy. The only evidence offered for this is Table 1, and its MNIST entries are not reproducible from the architecture described in Section 4. The text states: 'For all datasets, the architecture with real-valued or complex-valued LAUs employed a single dense layer with 3 neurons, followed by a softmax layer to produce the final output.' MNIST has 10 classes, yet a softmax over 3 logits cannot assign probabilities to 10 classes. The next sentence adds that 'For the MNIST dataset, convolutional layers (Conv2D, Batch Normalization, and MaxPooling) were incorporated to extract spatial features, with LAUs serving as the dense layers,' which both contradicts the 'single layer' claim and still does not explain how 3 LAU neurons yield 10 outputs. A 10-unit output layer would be required but is not mentioned. The same issue affects WBC, which is binary, so 3 logits cannot match its 2 classes. Because the reported MNIST accuracies of 97% and 98% are the main evidence for the abstract's 'state-of-the-art' and 'significantly improved computational efficiency' statements, the headline result is unverifiable and internally inconsistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":9558,"tokens_out":6757,"duration_ms":58831,"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":[{"comment":"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.","section":"Section 4, Table 1"},{"comment":"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":"Table 1"},{"comment":"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.","section":"Section 2.3, Eq. (14)"},{"comment":"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.","section":"Sections 3 and 4"}],"minor_comments":[{"comment":"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":"Section 2.1"},{"comment":"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":"Section 2.3"},{"comment":"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.","section":"Section 3"},{"comment":"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.","section":"Table 1"},{"comment":"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.","section":"Abstract and Section 4"}],"recommendation":"reject","confidential_remarks":"The paper has a sound but elementary mathematical core; the empirical evaluation, which is the basis for the state-of-the-art claim, is not reproducible as described and lacks baselines entirely. In my view, repairing this would require substantially reworking the experiments and correcting the complex-differentiability error, which goes beyond a routine revision. The authors may wish to resubmit after a more careful empirical study and with more modest claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe 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.\n\nWhat'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.\n\nWhere 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.\n\nFor 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.","headline":"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.","tokens_in":10119,"tokens_out":3587,"would_cite":false,"duration_ms":30174,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","26E60"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that one layer of Lehmer-mean activation units can match or exceed standard deep models while staying interpretable, the paper argues.","keywords":["Lehmer Transform","Neural Network Interpretability","Complex-Valued Neural Networks","Nonlinear Activation Functions","Weighted Lehmer Transform","Feature Selection","Phase-Sensitive Aggregation","Single-Layer Architecture"],"falsifier":"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.","tokens_in":9043,"feed_emoji":"🧠","tokens_out":12260,"duration_ms":105048,"temperature":0.7,"pith_summary":"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.","feed_headline":"One layer of Lehmer means matches deep nets","feed_subtitle":"Trainable weighted means give transparent, low-cost networks; the paper reports 95–99% on benchmarks.","key_machinery":"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).","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"introduces the Lehmer transform and its mathematical properties, which the activation unit is built on","marker":"[2]"},{"why":"describes the recently proposed alternative architecture that motivates a fixed, single-layer activation design","marker":"[1]"},{"why":"supplies the source for Schur convexity and mean inequalities used in the paper's analysis","marker":"[24]"},{"why":"provides the Wirtinger calculus used to establish complex differentiability of the transform","marker":"[19]"},{"why":"gives the complex-valued network framework the paper extends with its phase-sensitive activation","marker":"[17]"},{"why":"supplies background on complex-valued neural networks for oscillatory and phase-sensitive data","marker":"[16]"}],"fun_headline_variants":["One Lehmer layer rivals deep net accuracy","Trainable weighted mean: a powerful single-layer activation","Complex Lehmer transform for interpretable neural nets","Single-layer Lehmer networks: efficient and transparent","Lehmer means replace deep layers with one unit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["One Lehmer layer rivals deep net accuracy","Trainable weighted mean: a powerful single-layer activation","Complex Lehmer transform for interpretable neural nets","Single-layer Lehmer networks: efficient and transparent","Lehmer means replace deep layers with one unit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2285,"prompt_tokens":897,"completion_tokens":1388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1316}},"tokens_in":513,"tokens_out":1388,"duration_ms":10345,"temperature":1.0,"reasoning_tokens":1316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:29:41.808338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Scientific Reports 12(1), 3663 (2022)","cited_arxiv_id":null,"evidence_quote":"introduces the Lehmer transform and its mathematical properties, which the activation unit is built on"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the source for Schur convexity and mean inequalities used in the paper's analysis"},{"cited_title":": Complex-valued Neural Networks and Wirtinger Calculus","cited_arxiv_id":null,"evidence_quote":"provides the Wirtinger calculus used to establish complex differentiability of the transform"},{"cited_title":"arXiv preprint arXiv:1 705.09792 (2018)","cited_arxiv_id":null,"evidence_quote":"gives the complex-valued network framework the paper extends with its phase-sensitive activation"},{"cited_title":"John Wiley & Sons, ??? (2012)","cited_arxiv_id":null,"evidence_quote":"supplies background on complex-valued neural networks for oscillatory and phase-sensitive data"}],"review_version":1}