{"id":"db117120-b157-47a3-b1bf-5ef7a3af002b","arxiv_id":"2501.14000","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Local Control Networks put a separate learnable B-spline activation on every neuron and report small accuracy gains over MLPs and KANs on benchmark tasks.","lead":"This paper introduces Local Control Networks, where each neuron gets its own learnable B-spline activation curve instead of a shared ReLU or sigmoid. On standard benchmarks the networks show small accuracy gains over plain MLPs and sometimes over KANs, but the reported advantages are modest and the theoretical derivation contains errors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.3's backpropagation derivation uses the output-layer error for hidden layers, so the claimed localized/sparse gradient updates and efficiency advantage over KAN are not established; the empirical accuracy claims remain testable but the theoretical efficiency argument is unsupported.","rationale":"I read the paper as an empirical proposal for per-node B-spline activations, with a mathematical analysis intended to justify local-update efficiency claims. The central claim is twofold: accuracy at least comparable to MLP/KAN, and lower computational cost. The B-spline local-support property itself is real, and the architecture is simple enough to reimplement, but no code, seeds, or error bars are provided, and no machine-checked proof is offered. The most load-bearing weakness is the incorrect gradient derivation in Section 3.3 and Appendix A.2.4. Equation (13) applies the output-layer error 2/m(ŷ_i − y_i) to every hidden-layer activation, and Equation (16) treats ∂z_i^(l)/∂x_d as W_id^(1) for every layer. Correct backpropagation requires a recursive sum over downstream neurons for hidden layers, so Equations (6) and (8) are not true gradients. This directly undermines the paper's efficiency rationale: B-spline local support only makes individual neuron terms vanish, not whole downstream sums, so the claimed sparse or localized updates and natural dropout do not follow. The empirical accuracy comparisons are not logically destroyed by this error, but the theoretical support for the efficiency advantage is invalid, and the paper's own Section 5.2.5 admits that gains are marginal on the simple datasets used. I agree with the reader that this is the weakest assumption, and the recommended conditional verdict is unchanged.","tokens_in":10492,"tokens_out":5380,"duration_ms":51526,"concrete_test":"Implement a two-hidden-layer LCN in an autodiff framework with random inputs, targets, and B-spline coefficients; compute ∂L/∂W_{ij}^{(2)} via automatic differentiation and compare it to Equation (8). If the values differ, or if the correct hidden-layer gradient contains the summation over downstream units Σ_k ..., then the derivation is incorrect. As a second check, compute the sparsity fraction of hidden-layer weight gradients at initialization: if it is not substantially sparser than the gradient of a ReLU MLP of comparable size, the localized-update and sparse-efficiency claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing flaw is in Section 3.3 and Appendix A.2.4. Equation (13) sets ∂L/∂h_i^(l) = 2/m(ŷ_i − y_i) for every hidden layer l, but this form holds only at the output layer. For l < L, the correct chain rule introduces a sum over all downstream units: ∂L/∂h_i^(l) = Σ_k ∂L/∂h_k^(l+1) · ∂h_k^(l+1)/∂z_k^(l+1) · W_{ki}^{(l+1)}. Equations (6) and (8) therefore omit the backpropagated error through the downstream weight matrices, and Equation (16) incorrectly treats ∂z_i^(l)/∂x_d as W_id^(1) for every layer. Consequently, the claimed 'localized weight updates' and 'sparse updates for efficiency' do not follow from B-spline local support: the B-spline factor only zeroes the update for a single neuron when its pre-activation lies outside that neuron's support, while the hidden-layer weight gradient still couples every downstream neuron through the weight matrices. The theoretical efficiency argument in Section 4.2 and the 'effective dropout' claim in Section 3.3.2 are therefore unsupported. The empirical accuracy results could survive this error, but the central claim that LCNs are 'more computationally efficient than KANs' loses its stated mathematical basis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Local Control Networks (LCNs), in which each neuron has its own learnable B-spline activation function, as a middle ground between fixed-activation MLPs and Kolmogorov-Arnold Networks. The authors derive gradient formulas, argue that B-spline local support yields localized and sparse updates and hence efficiency advantages over KANs, and report experiments on four tabular benchmarks, MNIST/FMNIST, and symbolic regression tasks. The central thesis is that per-node B-spline activations achieve comparable or better accuracy than MLPs and KANs at lower computational cost.","tokens_in":10791,"tokens_out":4736,"duration_ms":40096,"significance":"If the efficiency claims were established, LCNs would be a useful contribution: the architecture is simple, retains the standard network structure, and B-spline local support is a principled way to obtain node-specific activations. The paper also addresses a timely comparison to KANs and evaluates on standard benchmarks, including symbolic regression. However, the theoretical efficiency argument, which is a central contribution, currently rests on an incorrect backpropagation derivation, and the empirical comparisons lack error bars and exact configuration details. With corrected derivations and stronger empirical reporting, the idea could be valuable, but as written the main novelty is not substantiated.","major_comments":[{"comment":"Equation (13) asserts that ∂L/∂h_i^(l) = 2/m(ŷ_i - y_i) for every hidden layer l. This identity holds only for the output layer when the loss is MSE. For l < L the correct recursion is ∂L/∂h_i^(l) = Σ_k ∂L/∂h_k^(l+1) · φ'_k(z_k^(l+1)) · W_{ki}^{(l+1)}, which couples all downstream neurons through the weight matrices. Because Eqs. (6) and (8) drop this coupling, the claimed localized weight updates and sparse updates in §3.3.3 and the efficiency comparison in §4.2 do not follow from B-spline local support. The B-spline factor only limits updates for an individual neuron whose pre-activation is outside its support; it does not decouple layers. This is the load-bearing error for the paper's central efficiency claim.","section":"Section 3.3.2, Eqs. (5)-(8), (13), (17)"},{"comment":"Equation (16) states that ∂z_i^(l)/∂x_d = W_id^(1) for every layer l, but this is only valid for l = 1. For l > 1 the pre-activation depends on x_d through h^(l-1), so the correct expression is ∂z_i^(l)/∂x_d = Σ_j W_ij^(l) ∂h_j^(l-1)/∂x_d. This compounds the error in Eq. (17), so the input-gradient formula is not a correct chain-rule application.","section":"Appendix A.2.4, Eq. (16)"},{"comment":"The empirical claims are not supported with uncertainty estimates: Figures 3-5 show single accuracy curves without error bars, and the text reports improvements of 'approximately 5%' (MNIST/FMNIST) and '1%'/'0.6%' without standard deviations or significance tests. Without repeated seeds and variance reporting, these margins may be within noise. In addition, the claim that parameter counts are 'standardized' (§5.1.2) is not backed by a table of model configurations or a description of the grid search, so the fairness of the MLP/KAN/LCN comparison is hard to verify.","section":"Section 5, Figures 3-5"}],"minor_comments":[{"comment":"The notation in Eq. (3) is unclear: the B-spline basis B_{N_l,p_l,n} depends on a knot sequence, but no per-neuron knot sequence is defined in Section 3.2.2. Please specify the knots for each neuron or state explicitly that knots are shared and learned.","section":"Section 3.2.2, Eq. (3)"},{"comment":"The input layer assumes x ∈ [0,1]^D, but the experiments on Bank, Bean, Spam, Telescope, MNIST, and FMNIST presumably require rescaling; please describe the preprocessing applied to each dataset.","section":"Section 3.2.1"},{"comment":"The text says LCN showed faster learning and convergence, but no learning curves, epoch counts, or convergence thresholds are reported; please either add the curves or temper the claim.","section":"Section 5.2.4"},{"comment":"Appendix A.3 defines FLOPs as 'floating-point operations per second', which is a rate, while the figures plot accuracy versus FLOPs; please clarify whether FLOPs means total floating-point operations.","section":"Appendix A.3"},{"comment":"The claim that LCNs are 'more computationally efficient than KANs' conflates theoretical sparsity with measured runtime; the figures in Appendix A.3 show accuracy versus FLOPs but do not include training time or memory measurements.","section":"Section 4.2 and Appendix A.3"},{"comment":"Appendix A.1 on the Consciousness Prior appears unrelated to LCNs and is not referenced anywhere in the main derivation or experiments; it should be removed or integrated into the related-work discussion.","section":"Appendix A.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a timely and interesting architecture, but the main theoretical justification for its efficiency advantage is currently invalid because of chain-rule errors in Section 3.3. I would ask the authors to re-derive the backpropagation, revise the theoretical efficiency claims accordingly, and add uncertainty measures to the experiments. Given that the empirical results are reported without reproducibility details, I would also encourage the editor to request code or detailed hyperparameters upon resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThis paper proposes Local Control Networks (LCNs), where each neuron has its own learnable B-spline activation function. The core idea is not new: the authors cite Bohra et al. (2020) on learning spline activations and Dushkoff & Ptucha (2016) on per-neuron activation selection, but do not compare against them. Equation (3) is a standard sum of B-spline basis functions, so LCNs are a modest variant of existing work. The empirical gains over MLPs and KANs are small (1–5%) and reported without error bars, seeds, or code, so the central claim is not convincingly supported.\n\nWhat the paper does well: the architecture is simple and clearly described, and the local-support property of B-splines is a legitimate reason to expect sparse, localized updates. The FLOPs comparisons in the appendix are a reasonable start, even if they are self-assessed.\n\nThe soft spots are serious. Section 3.3 and Appendix A.2.4 contain a clear chain-rule error: Equation (13) treats the output-layer error 2/m(ŷ_i − y_i) as the hidden-layer error for every layer, and Equation (16) uses first-layer input weights for all layers. Consequently, the derived gradients in Eqs. (6) and (8) are not correct backpropagation. This invalidates the theoretical claims of localized weight updates and sparse efficiency that are central to the paper's argument. The B-spline local support does create sparsity, but the error propagation through downstream weight matrices is missing, so the efficiency argument does not follow from the mathematics as written.\n\nThe lack of comparison to prior spline-activation work also undercuts the novelty claim. The paper cites the right references but does not benchmark against them, so we cannot tell what LCNs add beyond what is already known.\n\nOverall, this is a plausible idea with a flawed theoretical foundation and weak empirical reporting. It is not a takedown—the architecture could work in practice—but the paper, as submitted, does not support its claims. I would recommend against sending it to peer review in its current form. If the authors correct the derivation, add the missing baselines, and report proper experimental details, a revised version might be worth a serious look.","headline":"A plausible but not novel architecture is undermined by an incorrect backprop derivation and under-reported experiments.","tokens_in":11393,"tokens_out":6174,"would_cite":false,"duration_ms":51908,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes Local Control Networks, which give each neuron its own learnable B-spline activation curve, and argues this matches or beats both fixed-activation MLPs and the more complex Kolmogorov-Arnold Networks while using less…","keywords":["Local Control Networks","B-spline activations","learnable activation functions","activation diversity","Kolmogorov-Arnold Networks","sparse gradient updates","local support","neural network efficiency"],"falsifier":"Train an LCN with standard backpropagation and count, for each minibatch, what fraction of B-spline coefficients receive an exactly nonzero gradient. If the fraction is close to 1, or if exact backpropagation gives hidden-layer gradients different from the paper's formula, the central efficiency claim fails.","tokens_in":10261,"feed_emoji":"🧠","tokens_out":6940,"duration_ms":58540,"temperature":0.7,"pith_summary":"The paper argues that the standard practice of using one fixed activation function at every neuron is unnecessarily restrictive. It proposes Local Control Networks (LCNs), which give each neuron its own learnable B-spline activation curve, allowing different regions of the network to respond to data with different shapes. The authors claim that on standard benchmarks LCNs match or slightly beat MLPs, outperform KANs by about 5% on vision tasks and about 0.6% on basic tasks, and are cheaper to run than KANs because the local support of B-splines makes gradient updates sparse and localized. If correct, this means the extra architectural machinery of KANs is not always necessary and a standard feed-forward network with per-node spline activations can capture the same patterns at lower cost.","feed_headline":"Per-node B-spline activations beat MLPs and KANs on benchmarks","feed_subtitle":"Give each neuron its own learnable activation curve: better accuracy, fewer FLOPs, and no KAN complexity.","key_machinery":"The machinery is the B-spline basis function, a piecewise polynomial that is nonzero only on a bounded knot interval. LCNs build each neuron's activation as $h_i^{(l)} = \\sum_{n} w_{l,i,n} B_{N_l,p_l,n}(z_i^{(l)})$: the same standard linear transformation $z$ followed by a spline curve whose shape is set by learnable coefficients $w$. The load-bearing property is local support: for any input only a few basis functions are active, which the paper uses to justify sparse gradients, selective updates, and a built-in dropout effect, while the smoothness of B-splines is invoked for stable optimization.","core_discovery":"The paper's central claim is that replacing a shared fixed activation with a per-node learnable B-spline activation, while keeping the ordinary MLP feed-forward structure, gives a network that captures data patterns more flexibly and updates more locally. Concretely, each neuron's activation is a linear combination of B-spline basis functions with trainable coefficients, and because each basis function is nonzero only on a bounded interval, each neuron responds only to localized input regions. The paper reports that this design improves accuracy over MLPs on all tested benchmarks, outperforms KANs by about 5% on MNIST and Fashion-MNIST and by about 0.6% on basic tabular tasks, matches KANs on symbolic regression, and does so with fewer FLOPs than KANs.","pith_inferences":["If the hidden-layer gradient error is corrected, the qualitative claim of local updates may still hold, but the specific proof of sparsity would need to be re-derived from the exact chain rule.","A sharper test of the per-node-diversity hypothesis would be an ablation that keeps the architecture fixed and only toggles shared versus per-node spline parameters; the paper's MLP and KAN comparisons mix architectural and activation differences.","The fixed knot grid and spline degree are untested design choices; making them learnable or adaptive could enlarge the accuracy gap on the high-dimensional tasks where the paper expects LCNs to excel.","The claimed margin over KANs may be dataset-sensitive, so retesting on larger image sets such as CIFAR-10 would show whether the advantage scales."],"forward_implications":["On MNIST and Fashion-MNIST, LCNs outperform KANs by about 5% and slightly beat MLPs, so per-node B-spline activations can replace KAN-style edge functions in these vision tasks.","On the basic tabular benchmarks, LCNs improve accuracy over MLPs by about 1% and over KANs by about 0.6%, while converging faster in the first epochs.","Because B-splines have local support, LCN updates are sparse and localized, which the paper argues gives faster convergence, lower compute, and a natural regularization effect.","On symbolic regression, LCNs match KANs and both beat MLPs, suggesting edge-wise nonlinearity is not required for these function-representation tasks.","Each neuron's learned spline shape can be inspected, which the paper claims makes LCNs more interpretable than fixed-activation networks."],"supporting_citations":[{"why":"Defines the KAN architecture that LCNs claim to beat on efficiency and accuracy; supplies the main comparison baseline.","marker":"Liu et al., 2024"},{"why":"Provides the benchmark standards and parameter-matching protocol used for fair MLP, KAN, and LCN comparison.","marker":"Yu et al., 2024"},{"why":"Supplies the foundational definition and properties of B-splines on which LCN's activation curves are built.","marker":"De Boor, 1978"},{"why":"Provides the spline theory background for the local support and partition-of-unity properties used throughout the paper.","marker":"Lyche et al., 2018"},{"why":"Shows that spline activations can be optimized during training, the line of work LCN extends.","marker":"Bohra et al., 2020"},{"why":"Introduces an interpretable B-spline-based network, supporting LCN's interpretability claims.","marker":"Fakhoury et al., 2022"},{"why":"Demonstrates that mixing activation functions per node can outperform homogeneous networks, the core hypothesis of LCNs.","marker":"Hagg et al., 2017"},{"why":"Surveys activation functions and their limitations, supporting the motivation against a single fixed activation.","marker":"Dubey et al., 2022"}],"fun_headline_variants":["Local Control Networks: Per-neuron B-splines beat KANs by 5%","Give each neuron its own B-spline curve for better accuracy","B-spline per-neuron activations: simpler than KAN, more accurate","Localized B-spline activations outperform KANs on vision tasks","Replace fixed ReLU with per-node B-splines: gains on benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of sparse, localized updates assumes that for every neuron the error signal is simply the difference between prediction and true value, but that is only true for output neurons; for hidden neurons the error must be propagated backwards, so the claimed sparsity is not established.","fun_headline_variants_meta":{"raw":{"variants":["Local Control Networks: Per-neuron B-splines beat KANs by 5%","Give each neuron its own B-spline curve for better accuracy","B-spline per-neuron activations: simpler than KAN, more accurate","Localized B-spline activations outperform KANs on vision tasks","Replace fixed ReLU with per-node B-splines: gains on benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1801,"prompt_tokens":979,"completion_tokens":822,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":720}},"tokens_in":595,"tokens_out":822,"duration_ms":6391,"temperature":1.0,"reasoning_tokens":720,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:48:13.393221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train an LCN with standard backpropagation and count, for each minibatch, what fraction of B-spline coefficients receive an exactly nonzero gradient. If the fraction is close to 1, or if exact backpropagation gives hidden-layer gradients different from the paper's formula, the central efficiency claim fails.","supporting_citations":[{"cited_title":"A practical guide to splines","cited_arxiv_id":null,"evidence_quote":"Supplies the foundational definition and properties of B-splines on which LCN's activation curves are built."},{"cited_title":"Foundations of Spline Theory: B-Splines, Spline Approximation, and Hierarchical Refinement, pp.\\ 1--76","cited_arxiv_id":null,"evidence_quote":"Provides the spline theory background for the local support and partition-of-unity properties used throughout the paper."},{"cited_title":"Learning activation functions in deep (spline) neural networks","cited_arxiv_id":null,"evidence_quote":"Shows that spline activations can be optimized during training, the line of work LCN extends."},{"cited_title":"Exsplinet: An interpretable and expressive spline-based neural network","cited_arxiv_id":null,"evidence_quote":"Introduces an interpretable B-spline-based network, supporting LCN's interpretability claims."},{"cited_title":"Activation functions in deep learning: A comprehensive survey and benchmark","cited_arxiv_id":null,"evidence_quote":"Surveys activation functions and their limitations, supporting the motivation against a single fixed activation."}],"review_version":1}