{"id":"97bf4ae6-c4e8-4138-87d6-2353e6eefb4c","arxiv_id":"1908.05542","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A randomized-learning method that generates hidden-node weights and biases from slope angles and rotations, placing steep sigmoid fragments inside the input space, improves FNN approximation over fixed-interval random parameters.","lead":"This paper proposes a way to set the random internal parameters of a single-hidden-layer neural network: instead of drawing weights and biases from a fixed interval, it draws the slope angle of each sigmoid activation from a range tuned to the target function, then rotates and shifts the sigmoids so their steep parts land inside the input range.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported advantage rests on comparison against an untuned [-1,1] baseline, so the central improvement claim is not yet established.","rationale":"The reader's stated weakest assumption is that a uniform distribution of slope angles with a single alpha_min is an adequate prior for target complexity. That is a real but secondary issue: even a non-uniform angle distribution could be approximated by tuning alpha_min and alpha_max with enough nodes, so it does not by itself invalidate the method's usefulness. The more load-bearing concern is the empirical comparison. The paper's strongest experimental evidence is the enormous RMSE gap on TF (21) and TF (22), but the baseline is an untuned [-1,1] interval that Eq. (5) shows to correspond to very flat sigmoids, and the paper cites prior work showing such intervals can be badly misleading. Without a tuned-baseline comparison, the reported gap cannot be attributed to the angle parameterization; it may simply reflect that the baseline was prevented from using steep basis functions or placing transitions inside the input region. The reader's rationale does list the untuned baseline as a weakness, so there is partial agreement, but the formal weakest_assumption field points elsewhere. I therefore keep the reader's CONDITIONAL verdict: the algebraic derivation and algorithm are clear, but the central improvement claim needs additional experiments before it can be accepted as stated. The proposed concrete test would settle whether the concern lands by giving the standard method the same tuning opportunity and comparing against the most relevant prior method.","tokens_in":8959,"tokens_out":9510,"duration_ms":102120,"concrete_test":"Re-run all three benchmarks with the standard RVFL baseline under the same 10-fold cross-validation protocol, tuning weight and bias ranges separately over a grid of s values (for example s in {0.1, 0.5, 1, 2, 5, 10, 20, 50, 100}, using weight interval [-s_a, s_a] and bias interval [-s_b, s_b]), with the same m values and repeated trials; also include the generation method of [11]. Report best test RMSE for each method. If the tuned RVFL baseline or [11] closes most of the gap on TF (21) and TF (22), the central claim that the proposed method is much more accurate than existing randomized-learning approaches is not supported by the current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that generating hidden-node parameters by selecting slope angles uniformly in (alpha_min, alpha_max), rotating sigmoids, and shifting inflection points into the input hypercube yields substantially lower approximation error than standard randomized learning. The support is three simulations whose only comparator is the 'standard method' with both weights and biases drawn uniformly from [-1,1]. For the sigmoid in Eq. (2), weights in [-1,1] correspond by Eq. (5) to slope angles alpha in [-14 degrees, 14 degrees], i.e. deliberately flat basis functions, and biases in [-1,1] place many inflection points outside the input interval. This is exactly the regime that the paper's own references [5] and [6] identify as misleading, so the comparison does not isolate the proposed parameterization: it compares a tuned method against an untuned special case. A standard RVFL whose weight and bias ranges are themselves optimized by the same cross-validation budget could plausibly recover most of the advantage, since the key ingredients (steep sigmoids, inflection points or transition regions inside the input interval) are available to any RVFL with sufficiently large, well-chosen ranges. The absence of a comparison with the author's prior method [11], which already uses steepness-adjusted generation, further weakens the claim that the new angle-based rule is an improvement over existing randomized-learning practice. On the real CompActiv benchmark the advantage shrinks to 0.0335 versus 0.0358, which is consistent with the concern that the headline improvement is specific to carefully chosen oscillatory synthetic targets and an unfairly weak baseline. The uniform-angle prior is a separate untested assumption, but it is secondary to the baseline problem until a tuned-baseline comparison is performed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for generating the hidden-layer parameters of a single-hidden-layer feedforward network with random hidden nodes. Rather than drawing weights and biases from a fixed interval, it first samples a slope angle α for each sigmoid hidden node from an interval (α_min, α_max) chosen to match the target function's complexity, then computes the weight from a = 4 tan α in the one-dimensional case (or from a randomly rotated tangent hyperplane in the multidimensional case), and finally chooses the bias so that the sigmoid's inflection point lies at a randomly selected point inside the input hypercube. The method is illustrated on two synthetic regression functions and on the CompActiv data set; the reported test RMSEs are orders of magnitude smaller than those of the baseline that draws both weights and biases uniformly from [-1,1] for the synthetic functions, and slightly smaller on CompActiv.","tokens_in":9206,"tokens_out":11786,"duration_ms":112438,"significance":"The geometric rationale is clear and the algebraic derivation in Eqs. (3)-(8) and (11)-(19) is correct: the weight formula a = 4 tan α follows from the sigmoid derivative at the inflection point, and the bias formula b = -a^T x* correctly places the inflection point inside the input region. The procedure is interpretable and the hyperparameters are selected by standard cross-validation, with no circularity. If the empirical advantage were established against properly tuned baselines, the method would be a useful, practical alternative for randomized learning. However, the current evidence is not yet sufficient to support the central claim of substantial improvement, because the only comparator is an untuned fixed-interval baseline and no variability estimates are provided.","major_comments":[{"comment":"The central empirical claim is supported only by a comparison with the fixed-interval baseline a,b ∈ [-1,1]. As the paper itself notes in Section 1 (citing [5,6]), this interval is a known poor default; moreover, Eq. (5) shows that for the sigmoid in Eq. (2) weights in [-1,1] correspond to slope angles α ∈ [-14°,14°], i.e., deliberately flat basis functions, and biases in [-1,1] place many inflection points outside the input interval. The experiments therefore compare a cross-validated proposed method against an untuned special case rather than against standard randomized learning with a properly selected range. To support the abstract's claim, at least one baseline should tune the random-parameter range with the same cross-validation budget (e.g., the scaling factor s of [6] or the range optimization of [4]), and the comparison should include the author's earlier steepness-adjusted generation method [11].","section":"Section 3"},{"comment":"No variability estimates are reported for the synthetic experiments. Because the hidden parameters are random, every learning session yields a different model, and the single RMSE values (9.35e-7 for TF (21) and 0.0287 for TF (22)) may not be representative. Repeat the experiments over many independent random parameter generations for both the proposed and baseline methods and report means and standard deviations; for CompActiv, the small difference (0.0335 vs 0.0358) should be accompanied by a statistical test.","section":"Section 3, TFs (21) and (22)"},{"comment":"The selected optimum α_min = 85° lies at the edge of the searched grid shown in Fig. 3. Since the transformation a = 4 tan α diverges as α approaches 90°, a boundary optimum raises the concern that the headline result depends on an extreme hyperparameter value and that the grid search is not conclusive. Please extend the grid, constrain α_max, or otherwise demonstrate that the result is stable around the optimum.","section":"Section 3, TF (21)"}],"minor_comments":[{"comment":"The definition of α as the angle between the normal vector n and the y-axis unit vector u requires a'_0 to be positive for α ∈ (0°,90°), yet Algorithm 1 chooses the sign of a'_0 randomly. Because the tangent hyperplane is invariant to a global sign change of n, the random sign is redundant and the notation becomes confusing; please either fix a'_0 > 0 or define α through |a'_0|.","section":"Section 2.2, Eqs. (11)-(12), Algorithm 1"},{"comment":"The text reports the optimum α_min = 85°, but the left panel of Fig. 3 appears to show a grid search only up to α_min = 80°; please clarify the grid range and the coordinate labels.","section":"Section 3, Eq. (21)"},{"comment":"There are apparent typos in the data-generation description: 'uniform noise distributed in [0.2, 0.2]' should presumably read [-0.2, 0.2], and 'normalized in the range [1, 1]' should presumably read [-1, 1].","section":"Section 3, TF (22)"},{"comment":"For the standard method on CompActiv, the number of hidden nodes is not stated; please report it and indicate whether the same m is used for both methods.","section":"Section 3, CompActiv"},{"comment":"The paper does not provide code or a data repository, which would help reproducibility; please consider releasing the implementation and the exact hyperparameter grids.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The derivation is sound and the idea is publishable in principle, but the empirical section needs substantial strengthening before it can support the claimed improvement. The lack of tuned baselines and variance estimates is the main obstacle; I do not see any ethical or scope concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a clean parameterization idea with a weak empirical case. The slope-angle interval, rotation of sigmoids, and bias-from-inflection-point construction are genuinely new relative to the fixed-interval RVFL practice, and the derivation from tan(alpha) to a = 4 tan(alpha) and the multidimensional normal-vector formula check out. The paper also explains clearly why uniform weights in [-1,1] produce flat sigmoids, which is a real service.\n\nThe problem is the experiments. The only baseline is weights and biases uniformly in [-1,1], which the paper's own references [5] and [6] call misleading. That comparison is an untuned special case, not a fair standard. A standard RVFL with cross-validated ranges would likely capture much of the advantage, since steep sigmoids and inflection points in the input interval are available to any RVFL with well-chosen ranges. The author's prior method [11] is not compared at all, so we cannot tell whether the angle parameterization adds anything beyond that. On the CompActiv benchmark the gain is 0.0335 vs 0.0358, which is small and plausibly within noise; no repeated-trial variance is reported for the synthetic tasks, and the optimal alpha_min for TF (21) sits at the edge of the search grid. The uniform-angle prior is asserted, not tested.\n\nThat said, the soft spots are about support, not correctness. The algebra is sound, the method is described well, and the author honestly acknowledges the small real-data gain. With a tuned baseline, error bars, and a comparison to [11], this would be a solid contribution. Right now the central claim—much lower error—is conditional.\n\nWho is this for? People working on RVFL and randomized neural networks. It deserves a serious referee, but I would expect heavy revision.","headline":"A genuinely new angle-based parameterization for RVFL hidden nodes with clean algebra, but the experiments compare only against an untuned [-1,1] baseline, so the headline performance gain is not yet established.","tokens_in":9814,"tokens_out":1626,"would_cite":false,"duration_ms":16008,"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":"Randomized neural networks fit steep target functions when hidden sigmoids are generated by slope angle, rotation, and placement rather than by fixed-interval weights and biases.","keywords":["function approximation","feedforward neural networks","neural networks with random hidden nodes","randomized learning algorithms","sigmoid activation","slope-angle parameterization","regression"],"falsifier":"Fit a one-dimensional target with rapid oscillation confined to a small subinterval and near-flat elsewhere, e.g. $g(x)=\\sin(200x)$ on $[0,0.1]$ and $0$ on $(0.1,1]$, using the paper's Algorithm 1 with $\\alpha_{\\min}$ and $m$ tuned by cross-validation. If the error remains as low as on the paper's synthetic targets while the sigmoid inflection points are placed uniformly rather than concentrated in the oscillating region, the uniform-angle assumption holds; if not, that assumption is the part that fails.","tokens_in":8708,"feed_emoji":"📈","tokens_out":10759,"duration_ms":98964,"temperature":0.7,"pith_summary":"This paper addresses a gap in randomized learning of single-hidden-layer feedforward networks: how to choose the random weights and biases of hidden nodes. It argues that drawing both from a fixed interval such as [-1,1] creates flat, misplaced sigmoids and fails on target functions with strong fluctuations. The proposed alternative generates each hidden neuron from the geometry of the sigmoid: sample a slope angle at the inflection point from an interval tuned to the target, rotate the sigmoid randomly, then shift its inflection point into the input space. In simulations, the method reduces test root-mean-square error from above 0.1 to 9.35e-7 on a one-dimensional target and from 0.2326 to 0.0287 on a two-dimensional target. A reader would care because this turns the arbitrary weight range into an interpretable, tunable angle parameter and explains why the standard convention can fail.","feed_headline":"Slope-angle randomization cuts neural fit error to 9e-7","feed_subtitle":"Generating hidden neurons by angle and placement, not fixed [-1,1] weights, fits steep target functions.","key_machinery":"The load-bearing object is the slope angle $\\alpha$: the angle between the tangent line (or hyperplane) to the sigmoid at its inflection point and the input axes. Because the sigmoid derivative at the inflection point is one quarter of the weight, the weight is recovered by $a = 4\\tan\\alpha$ in one dimension, and in $n$ dimensions the normal vector to the tangent hyperplane supplies both orientation and slope. The construction randomly draws the first $n$ normal components, fixes the remaining component from $\\tan\\alpha$, derives hidden weights, and then determines each bias so that the inflection point coincides with a randomly chosen training point or cluster prototype. This single mechanism keeps activation functions in their non-saturated regime and aligns their steep parts with the input distribution; $\\alpha_{\\min}$ is the new scalar that encodes target complexity.","core_discovery":"The central claim is that weights and biases of hidden nodes should not be sampled independently from the same fixed interval; they should be constructed so that the steep fragment of every sigmoid lands inside the input hypercube. For a one-input sigmoid the paper sets the weight by sampling the slope angle $\\alpha$ at the inflection point and using $a = 4\\tan\\alpha$, then chooses the bias from $b_i = -a_i x^*_i$ so the inflection point sits at a randomly selected input point $x^*$. For multiple inputs, it builds a random tangent hyperplane to the sigmoid: the first $n$ components of the normal vector are drawn uniformly from $[-1,1]$, the last is fixed by $\\cos\\alpha = a'_0 / \\sqrt{\\sum a'_k{}^2}$, and the hidden weights follow $a_k = -4a'_k/a'_0$. This yields a set of constructional sigmoids that are steep, randomly oriented, and spread across the data. The paper reports test RMSE of $9.35\\times 10^{-7}$ versus above $0.1$ on target (21), and $0.0287$ versus $0.2326$ on target (22); on the 21-dimensional Compactiv dataset the advantage shrinks to $0.0335$ versus $0.0358$, which the paper attributes to the absence of strong fluctuations.","pith_inferences":["If the angle parameterization is the right prior, the fixed $[-1,1]$ weight interval may explain much of the excessive node counts and instability reported for randomized networks on nonlinear problems.","The construction can be adapted to nonstationary targets by letting $\\alpha_{\\min}$ vary across the input space, using local estimates of fluctuation strength; the paper does not test this.","The same geometry-based generation should carry over to other activation functions whose maximal derivative location is known, a direction the paper mentions only as future work.","A plausible practical extension is to draw the rotation components from a narrower or wider interval to control the diversity of orientations, even though the paper recommends $[-1,1]$."],"forward_implications":["Hidden-node weights and biases should be treated as coupled quantities derived from the activation function's geometry, not as independent draws from one interval.","A single cross-validated limit angle $\\alpha_{\\min}$ can replace a search over symmetric weight ranges, with $\\alpha_{\\max} = 90^\\circ$ as a default.","On strongly fluctuating target functions the method reaches test RMSE below $10^{-6}$ in the one-dimensional case and below $0.03$ in the two-dimensional case, while the fixed $[-1,1]$ baseline stays above $0.1$ and $0.23$.","The advantage shrinks on smooth, high-dimensional data, so the method's benefit is concentrated precisely where randomized networks usually struggle."],"supporting_citations":[{"why":"Formulates output-layer training as a linear least-squares problem solved by pseudo-inverse, the fitting procedure used throughout the study.","marker":"[1]"},{"why":"Proves universal approximation when random parameters come from a proper range, justifying the randomized setting.","marker":"[2]"},{"why":"Extends universal approximation to symmetric random-parameter intervals under a Lipschitz condition, the standard comparison setting.","marker":"[3]"},{"why":"Demonstrates empirically that the common [-1,1] interval fails on nonlinear maps, establishing the baseline problem the paper addresses.","marker":"[5]"},{"why":"Shows the random-parameter range controls the saturation versus discrimination tradeoff, motivating a slope-angle parameterization.","marker":"[6]"},{"why":"Argues random nodes should be placed according to input density, the rationale for shifting inflection points into the input hypercube.","marker":"[8]"},{"why":"Derives random-parameter formulas that place steep activation fragments in the input region and adjust slopes to target complexity; this paper proposes an angle-and-rotation alternative.","marker":"[11]"}],"fun_headline_variants":["Random neuron slopes beat fixed intervals for steep functions","Angle-based hidden neurons fit steep targets, RMSE 9e-7","Sigmoid slopes and rotation improve neural learning","Constructive sigmoids: better than fixed-range weights"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's edge depends on one global steepness interval, sampled uniformly, being able to supply every steepness the target function actually needs; if a target mixes very different local frequencies, the single limit angle $\\alpha_{\\min}$ can misplace or mis-size the steep sigmoids.","fun_headline_variants_meta":{"raw":{"variants":["Random neuron slopes beat fixed intervals for steep functions","Angle-based hidden neurons fit steep targets, RMSE 9e-7","Sigmoid slopes and rotation improve neural learning","Constructive sigmoids: better than fixed-range weights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1266,"prompt_tokens":930,"completion_tokens":336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":269}},"tokens_in":546,"tokens_out":336,"duration_ms":3835,"temperature":1.0,"reasoning_tokens":269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:10:34.160712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit a one-dimensional target with rapid oscillation confined to a small subinterval and near-flat elsewhere, e.g. $g(x)=\\sin(200x)$ on $[0,0.1]$ and $0$ on $(0.1,1]$, using the paper's Algorithm 1 with $\\alpha_{\\min}$ and $m$ tuned by cross-validation. If the error remains as low as on the paper's synthetic targets while the sigmoid inflection points are placed uniformly rather than concentrated in the oscillating region, the uniform-angle assumption holds; if not, that assumption is the part that fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formulates output-layer training as a linear least-squares problem solved by pseudo-inverse, the fitting procedure used throughout the study."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Proves universal approximation when random parameters come from a proper range, justifying the randomized setting."},{"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":"Extends universal approximation to symmetric random-parameter intervals under a Lipschitz condition, the standard comparison setting."},{"cited_title":"Information Sciences 382–383, 170–178 (2017)","cited_arxiv_id":null,"evidence_quote":"Demonstrates empirically that the common [-1,1] interval fails on nonlinear maps, establishing the baseline problem the paper addresses."},{"cited_title":"Information Sciences 367, 1094–1105 (2016)","cited_arxiv_id":null,"evidence_quote":"Shows the random-parameter range controls the saturation versus discrimination tradeoff, motivating a slope-angle parameterization."},{"cited_title":"Information Sciences 364, 146–155 (2016)","cited_arxiv_id":null,"evidence_quote":"Argues random nodes should be placed according to input density, the rationale for shifting inflection points into the input hypercube."},{"cited_title":"Information Sciences 481, 33–56 (2019)","cited_arxiv_id":null,"evidence_quote":"Derives random-parameter formulas that place steep activation fragments in the input region and adjust slopes to target complexity; this paper proposes an angle-and-rotation alternative."}],"review_version":1}