{"id":"de5599ea-8064-4fd1-9a02-cb1056c3d681","arxiv_id":"1908.03891","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Dudek proposes setting the weights and biases of random hidden-layer sigmoids using local linear fits to the target function, and shows improved regression accuracy on a synthetic function and two real datasets.","lead":"Randomized neural networks get hidden units whose slopes are tuned to the local shape of the target function. The method places each sigmoid at a random training point, matches its steepness to a local straight-line fit, and then learns the output weights by least squares.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Kin8nm result uses k'=60, outside the declared D-DM search grid {5,10,...,50}; the flagship real-data comparison is not reproducible as written.","rationale":"D-DM's local-slope construction (Eqs. 23-24) is mathematically coherent and the paper's noise sweep is a useful sanity check. The reader's weakest assumption about slope instability in high dimensions is plausible but speculative; the more immediate, falsifiable problem is the mismatch between the declared D-DM hyperparameter grid and the Table 1 Kin8nm entry. This mismatch directly affects the strongest real-data comparison and can be settled by rerunning with the stated grid. Since the paper is already conditional and our finding strengthens the need for a reproducible rerun rather than overturning the method, I leave the verdict unchanged.","tokens_in":10841,"tokens_out":10355,"duration_ms":114722,"concrete_test":"Re-run the Section 4 Kin8nm experiment exactly under the stated protocol, with D-DM's k' restricted to {5,10,...,50}; record the best CV-selected k' and test RMSE. If the best RMSE under this grid is >= 0.0523 or is not below RARSM's 0.0618, the headline Kin8nm improvement is unsupported as reported. As a secondary check, rerun Stock and Kin8nm with k' >= n+1 to verify that no reported optimum ever uses an underdetermined hyperplane fit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that D-DM beats the RARSM and rsM baselines, especially on Kin8nm (Table 1: 0.0523 vs 0.0618). That result depends on k' = 60, but Section 4 fixes the D-DM neighborhood-size grid as k' in {5,10,...,50}. The same section defines k' = k+1, and Algorithm 1 requires k >= n, i.e. at least n+1 points to fit the hyperplane. The stated grid also contains k' = 5 for Stock (n=9) and Kin8nm (n=8), values below n+1 that cannot support the local hyperplane. So the table and the protocol are mutually inconsistent. If k' = 60 was actually used, D-DM received extra hyperparameter freedom beyond the declared grid, and the Kin8nm margin may not survive a rerun with k' capped at 50; if k' = 60 is a typo, the paper must say what the true optimum was. This is the load-bearing weak point because it sits directly on the strongest evidence for the headline claim. The derivation in Eqs. (19)-(24) itself is sound; the issue is whether the reported comparison was generated under the stated rules.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven method for generating hidden-node weights and biases in single-hidden-layer feedforward neural networks trained by randomized learning. For each hidden neuron, the method randomly selects a training point, finds its k nearest neighbors, fits a hyperplane to that neighborhood, and sets the sigmoid weights to a_j = 4 a'_j and the bias to b = -a^T x^* (Eqs. 19-24). The output weights are then obtained by least squares. The paper argues that this local slope adjustment places sigmoids so that their steep fragments match local target-function fluctuations, improving approximation accuracy compared with fixed-interval random parameters and earlier interval- or angle-tuned methods. Experiments are reported on one synthetic two-variable function and two real datasets (Stock and Kin8nm).","tokens_in":11130,"tokens_out":5018,"duration_ms":58052,"significance":"If the reported results are reproducible, the method is a simple and potentially useful contribution to randomized neural-network learning. The derivation of a = 4a' from the sigmoid derivative (Eqs. 14-17 and 20-23) is correct, and the algorithm is clearly specified. The idea of adjusting random hidden-node parameters to local data geometry is intuitive and goes beyond the usual fixed-interval generation. The paper also provides extensive simulations with 100 trials and 10-fold cross-validation, which is a reasonable experimental protocol. The main limitation is that the empirical support for the headline claim rests on one synthetic function and two real datasets, one of which (Stock) shows only a marginal improvement over the best baseline; the strongest real-data result (Kin8nm) is currently affected by a reporting inconsistency in the hyperparameter grid, as detailed below. The contribution is significant conditional on correcting that inconsistency.","major_comments":[{"comment":"The declared D-DM neighborhood-size grid is k' in {5, 10, ..., 50}, but Table 1 reports k' = 60 for Kin8nm. This is outside the declared search space. Since the Kin8nm result is the strongest evidence for D-DM (0.0523 vs. 0.0618 for RARSM), the headline comparison is not reproducible as written. Please clarify whether k' = 60 was actually used (and correct the grid description) or rerun the experiment with k' restricted to the declared grid. Additionally, the declared grid contains k' = 5 for Stock (n = 9) and Kin8nm (n = 8), but Algorithm 1 requires k >= n, i.e., k' = k + 1 >= n + 1; those grid points are infeasible and should be removed or the feasibility condition should be restated.","section":"Section 4 vs. Table 1"},{"comment":"The hidden-node candidate set in Section 4 is declared as {50, 100, ..., 1000}, but Table 1 reports #nodes = 1300 for FIM, OIM, rsM, and RARSM on Kin8nm. This is outside the declared search space, so the comparison protocol is not consistently described. The authors should correct the grid description or the table and confirm that all methods were optimized under the same rules; otherwise the baseline results are not reproducible.","section":"Table 1, Kin8nm rows"},{"comment":"The abstract and conclusions claim that the method 'brings very good results' compared with other methods, but the empirical support is limited to one synthetic function and two real datasets, and on Stock the D-DM improvement over RARSM is small (0.0277 vs. 0.0285 with overlapping standard deviations). If the Kin8nm result is corrected, the strength of the general claim should be moderated, or additional datasets should be added to support the claimed advantage.","section":"Section 5 / Abstract"}],"minor_comments":[{"comment":"The text states that 'the weights and biases of the sigmoids are no longer random' (page 7), which is in tension with the paper's use of the term 'randomized learning' and with the abstract's wording about 'random parameters.' Please clarify the terminology: the hidden parameters are data-dependent in this method, not drawn from a fixed distribution, and this departure from the standard randomized-learning setting should be made explicit.","section":"Section 3 (intro paragraph)"},{"comment":"No preprocessing is described for the Stock and Kin8nm datasets. Since the method uses k-nearest-neighbor searches and hyperplane fits, both of which are sensitive to input scaling, the paper should report whether and how the input variables were standardized or normalized.","section":"Section 4"},{"comment":"The notation for the noise bound c in the synthetic experiments conflicts with the random sign variable c in Eq. (10). Please use distinct symbols to avoid confusion.","section":"Section 4"},{"comment":"There is a typo in Fig. 6's y-axis label ('RSME' instead of 'RMSE'), and in the Stock dataset description 'aproximate' should be 'approximate.' These should be corrected in the final version.","section":"Figures 5-6 and text"},{"comment":"Reference [9] is cited as 'in print'; if the venue and year are now available, the citation should be updated. Also consider citing any publicly available code or data-reproduction details for the proposed method, as the manuscript currently provides no implementation details beyond the algorithm.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a sensible incremental advance on the author's earlier rsM/RARSM work, and the core derivation is correct. But the strongest real-data result (Kin8nm) is reported with k'=60, outside the D-DM search grid the paper itself specifies, and the grid contains values that Algorithm 1 rules out. That needs fixing before I'd trust the headline numbers.\n\nWhat's actually new: instead of drawing sigmoid slopes from a global interval, D-DM picks a random training point, fits a hyperplane to its k nearest neighbors, and sets the neuron weights to 4 times the hyperplane slopes, with the bias placing the inflection point at that training point. The sigmoid derivative identity (a/4 at the inflection point) is correct, and the idea of per-neuron slope adaptation to local target curvature is a genuine departure from the earlier interval-based schemes. On the synthetic function and on Stock/Kin8nm, D-DM beats FIM, OIM, rsM, and RARSM, and it uses fewer hidden nodes on the synthetic example. That is a real, if modest, contribution. The paper also clearly explains the bias-variance tradeoff via the neighborhood size k.\n\nThe soft spots. First, the protocol mismatch: Section 4 fixes the D-DM k' grid at {5,10,...,50}, but Table 1 shows k'=60 for Kin8nm. For Stock (n=9) and Kin8nm (n=8), the grid values include k'=5, which corresponds to k=4, below the k>=n requirement in Algorithm 1. So the stated hyperparameter search is internally inconsistent, and the one dataset where D-DM shows a decisive improvement is exactly the one whose reported optimum lies outside the declared grid. The authors need to explain whether k'=60 was actually used, and if so, why it is not in the search set; if it is a typo, the real optimum should be reported. Without that, the Kin8nm claim is not reproducible as written.\n\nSecond, the evaluation is thin: one synthetic 2-D function and two real datasets, one of which (Stock) shows only a marginal improvement. There is no code, no sensitivity analysis for k on the real data, and no discussion of behavior when the local linearity assumption is violated (e.g., high-dimensional or noisy inputs). These are moderate issues, not fatal. The derivation and the method are sound; the evidence base just needs to be larger and cleaner.\n\nBottom line: this deserves a serious referee, but it needs a corrected protocol and a broader empirical test. I would conditionally accept if the authors fix the inconsistency and supply code. If you want, I can dig into the equations with you.","headline":"A simple, plausible local-slope trick for randomized FNNs, undermined by a reproducibility gap in its headline Kin8nm result.","tokens_in":11613,"tokens_out":3462,"would_cite":false,"duration_ms":33708,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A randomized single-hidden-layer network can approximate complex functions by placing each sigmoid at a random training point and setting its weight to four times the local slope fitted to that point's nearest neighbors.","keywords":["data-driven randomized learning","feedforward neural networks","neural networks with random hidden nodes","randomized learning algorithms","sigmoid slope adjustment","local hyperplane fitting","k nearest neighbors","regression approximation"],"falsifier":"Compare D-DM and RARSM on the synthetic function $g(x_1,x_2)=\\sin(20\\exp(x_1))x_1^2+\\sin(20\\exp(x_2))x_2^2$ with noise $c=0.2$ at the same number of hidden nodes, say $m=300$; if D-DM's test RMSE is not below RARSM's, the claim that local slope matching gives an advantage at matched capacity fails. A more direct check is to compare the hyperplane coefficients used by D-DM with the true partial derivatives of the target at the same anchors: the coefficients should approximate those derivatives after the factor 4.","tokens_in":10646,"feed_emoji":"📈","tokens_out":8235,"duration_ms":76299,"temperature":0.7,"pith_summary":"This paper proposes a data-driven rule for generating the hidden-neuron parameters of a randomized feedforward network: pick a training point at random, fit a hyperplane to it and its $k$ nearest neighbors, then set the sigmoid weights to four times the hyperplane slopes and the bias so the sigmoid's inflection point sits at the chosen point. The claim is that this places sigmoids where the target function changes and gives each one the local slope of the data, so a linear combination of such sigmoids can approximate strongly nonlinear functions much better than parameters drawn from a fixed interval. On a noisy two-variable target function and on two regression benchmarks, the proposed D-DM method reaches lower test RMSE than the fixed-interval, optimized-interval, and two recent slope-adapting methods, often with fewer hidden nodes. If correct, randomized hidden layers no longer need a hand-tuned parameter interval because their basis functions are locally fitted to the data.","feed_headline":"Hidden neurons tuned to local slopes beat fixed intervals","feed_subtitle":"D-DM fits a hyperplane to each anchor's neighbors, scales sigmoids to local slopes, cuts node counts.","key_machinery":"The load-bearing object is the local hyperplane fit to a random training point's $k$-nearest-neighbor neighborhood, together with the derivative-matching identity $a_j = 4a'_j$. For a logistic sigmoid $h(\\mathbf{x}) = (1+\\exp(-(\\mathbf{a}^\\top \\mathbf{x}+b)))^{-1}$, the derivative at the inflection point in direction $j$ is $a_j h(1-h)=a_j/4$, so multiplying a fitted slope $a'_j$ by four makes the sigmoid's local directional slope equal the target's local slope as estimated by the hyperplane. The bias formula $b = -\\mathbf{a}^\\top \\mathbf{x}_*$ then anchors each sigmoid's most sensitive part at a real data point. This turns the randomized hidden layer from a random projection into an ensemble of locally tangent basis functions while keeping the output layer a linear least-squares problem.","core_discovery":"For each hidden node, D-DM draws one training sample $\\mathbf{x}_*$, collects the $k$ nearest neighboring training points, and fits a hyperplane $y = a'_1 x_1 + \\cdots + a'_n x_n + b'$ to this neighborhood. Because a sigmoid's slope at its inflection point is a quarter of its weight, setting $a_j = 4a'_j$ makes the sigmoid's directional slopes match the hyperplane's local estimates of the target slope; the bias $b = -\\mathbf{a}^\\top \\mathbf{x}_*$ fixes the inflection point exactly at $\\mathbf{x}_*$. The fitted curve is then the usual linear combination of these sigmoids with output weights from the pseudoinverse. The paper's key numerical finding is that this construction approximates strongly nonlinear targets with lower RMSE than the baselines: on the Kin8nm benchmark the D-DM test RMSE is 0.0523 versus 0.0618 for the best slope-angle baseline, and on the synthetic function it uses 300 nodes versus RARSM's 350 at a lower error. The intended upshot is that data-anchored, locally scaled sigmoids form a better projection space for randomized networks than interval-sampled parameters.","pith_inferences":["The same derivative-matching recipe could be extended to other locally fitted basis families, such as radial basis functions centered at random data points, where the local scale would be set by nearest-neighbor distances instead of hyperplane slopes.","In high-dimensional inputs, $k$-nearest-neighbor neighborhoods become less local, so the local hyperplane slope estimates would degrade; a testable fix is to fit hyperplanes in random low-dimensional projections and combine the resulting coordinates.","For classification, anchors chosen near decision-boundary training points rather than uniformly at random should concentrate nonlinear flexibility where class boundaries curve; the paper only demonstrates regression.","If local slope estimates are unstable, averaging hyperplanes over several random anchors per hidden node, or ridge-regressing the neighborhood, might recover accuracy; this is an observable modification of Algorithm 1."],"forward_implications":["Strongly nonlinear regression targets can be approximated accurately without searching for a good weight interval; $k$ and the number of hidden nodes $m$ are the only tunable hyperparameters.","Fewer hidden nodes are needed: on the synthetic function D-DM reaches lower RMSE with 300 nodes than RARSM's 350, and on Kin8nm with 900 nodes versus RARSM's 1300.","The hidden layer can be built in one pass plus $k$-nearest-neighbor searches, and the output weights come from the Moore-Penrose pseudoinverse, so no gradient computation enters the learning loop.","The neighborhood size $k$ controls the bias-variance tradeoff: low-noise data favors small neighborhoods ($k'=20$) and high-noise data favors larger ones ($k'=30$), giving a concrete tuning path."],"supporting_citations":[{"why":"supplies the universal-approximation result for random parameters, establishing the feasibility that D-DM tries to improve.","marker":"[2]"},{"why":"documents why the fixed [−1, 1] interval fails on nonlinear maps, the failure mode D-DM targets.","marker":"[5]"},{"why":"predecessor method rsM that generates data-dependent weights and biases; it is one of the baselines D-DM is compared with.","marker":"[8]"},{"why":"introduces RARSM, the random slope-angle baseline whose local performance D-DM improves on in the experiments.","marker":"[9]"},{"why":"supplies the convex least-squares solution for the output weights, the training rule that combines the generated sigmoids.","marker":"[1]"}],"fun_headline_variants":["Local slope fitting sharpens randomized neural learning","Data-aware sigmoid placement surpasses fixed intervals","Sigmoid slopes from neighbor fits reduce test error","Target-adaptive sigmoids beat fixed-range random weights","Neighbor-anchored sigmoid scaling cuts neural net RMSE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction rests on the assumption that the hyperplane fitted to a random training point and its $k$ nearest neighbors gives trustworthy estimates of the target function's local slopes in every input direction; if noise or high dimensionality makes these slope estimates unreliable, the sigmoids are placed with wrong slopes and the advantage over interval-tuned baselines shrinks.","fun_headline_variants_meta":{"raw":{"variants":["Local slope fitting sharpens randomized neural learning","Data-aware sigmoid placement surpasses fixed intervals","Sigmoid slopes from neighbor fits reduce test error","Target-adaptive sigmoids beat fixed-range random weights","Neighbor-anchored sigmoid scaling cuts neural net RMSE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3359,"prompt_tokens":937,"completion_tokens":2422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":2345}},"tokens_in":553,"tokens_out":2422,"duration_ms":19178,"temperature":1.0,"reasoning_tokens":2345,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:58:53.621769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare D-DM and RARSM on the synthetic function $g(x_1,x_2)=\\sin(20\\exp(x_1))x_1^2+\\sin(20\\exp(x_2))x_2^2$ with noise $c=0.2$ at the same number of hidden nodes, say $m=300$; if D-DM's test RMSE is not below RARSM's, the claim that local slope matching gives an advantage at matched capacity fails. A more direct check is to compare the hyperplane coefficients used by D-DM with the true partial derivatives of the target at the same anchors: the coefficients should approximate those derivatives after the factor 4.","supporting_citations":[{"cited_title":"In: Neural Networks for Conditional Probability Estimation: Forecasting Beyond Point Predictions, chap- ter 6, 87–97, Springer-Verlag London (1999)","cited_arxiv_id":null,"evidence_quote":"supplies the universal-approximation result for random parameters, establishing the feasibility that D-DM tries to improve."},{"cited_title":"Information Sciences 481, 33–56 (2019)","cited_arxiv_id":null,"evidence_quote":"predecessor method rsM that generates data-dependent weights and biases; it is one of the baselines D-DM is compared with."},{"cited_title":"In: 15th Intern ational Work-Conference on Artificial Neural Networks, (2019) (in print)","cited_arxiv_id":null,"evidence_quote":"introduces RARSM, the random slope-angle baseline whose local performance D-DM improves on in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the convex least-squares solution for the output weights, the training rule that combines the generated sigmoids."}],"review_version":1}