{"id":"a3362573-8d72-46f2-b958-46a5b59005e3","arxiv_id":"1908.01321","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A spatio-temporal RBF network, mathematically equivalent to a standard RBF with more hidden units, is reported to identify a nonlinear system with lower MSE, but the comparison uses unequal hyperparameters.","lead":"This paper adds an extra 'time' index to a radial basis function (RBF) neural network and reports lower error than two older RBF variants on one simulated nonlinear system identification task. The comparison is hard to trust because the new network uses a much larger learning rate, more hidden units, and leaves a key time parameter undefined.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported ST-RBF advantage is confounded with a 500x learning-rate increase and S×T hidden units; matching these in a conventional RBF would settle whether the architecture itself is responsible.","rationale":"The reader's weakest assumption pinpoints the same confound I find most load-bearing: the experiment changes learning rate, hidden-unit count, and leaves T and φ undefined. I read the paper's central claim as causal: the spatio-temporal indexing itself yields the large gains. The reported numbers are consistent with alternative explanations, so the correctness risk is high. The gradient-descent derivation itself is standard and likely correct, and the availability of MATLAB code is a point in the authors' favor, but the central empirical comparison is the basis for the paper's contribution and it is not controlled. The proposed matched-capacity ablation is feasible with the supplied code and would settle whether the confound cancels the claimed effect. Since the reader already rejected on this ground and I do not see that the concern misses, the verdict should remain unchanged.","tokens_in":5566,"tokens_out":4142,"duration_ms":45583,"concrete_test":"Run the MATLAB code (File Exchange #68415) on the same system in Eq. (14) with the ST-RBF update replaced by the conventional RBF update over the same S×T hidden units, same centers, σ=1, same initialization scheme, and η=1×10−2; keep T and φ(i,t) exactly as in the code. If the matched-capacity conventional RBF reaches test MSE near −19.67 dB, the claimed superiority of the spatio-temporal architecture is not supported. If it remains near −4.4 dB, the architecture effect is real. A complementary check is a learning-rate sweep for both methods at fixed capacity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that re-indexing RBF hidden units in the spatio-temporal architecture produces the large MSE gains. For that claim to be supported, the comparison in Section III must isolate the architecture. It does not: the conventional and fractional RBF baselines use η=2×10−5, while the proposed ST-RBF uses η=1×10−2 (a 500× increase, described as 'half' in the text); the proposed network sums over S×T hidden units in Eq. (5), whereas the baseline has S=6 hidden units; and T is never defined, nor are φ(i,t)(x,c(i,t)) or c(i,t) specified. A 500× larger learning rate alone can plausibly move the baseline from a poorly converged MSE around −1.7 dB to a much lower training and test MSE, independently of any architectural benefit. The increase from 6 to S×T hidden units also adds capacity and could explain a better fit of the nonlinear system in Eq. (14). No variance or confidence intervals are reported despite the claimed 1000 Monte Carlo rounds, so the −19.67 dB test figure cannot be separated from these confounds. The internal inconsistency in the learning-rate sentence reinforces that the hyperparameter change may be unintended.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a 'spatio-temporal RBF neural network' for nonlinear system identification. The network extends the standard RBF output sum in Eq. (1) to a double sum over spatial index i and temporal index t in Eq. (5), and the authors derive a gradient-descent update rule in Eqs. (8)-(13). The paper evaluates the method on a single simulated nonlinear system described by Eq. (14), comparing against a conventional RBF and a fractional-gradient RBF. Reported training MSEs are -15.1286 dB for the proposed method versus -1.6813 dB (RBF) and -1.7444 dB (FRBF); test MSEs are -19.67 dB versus -4.431 dB and -4.955 dB, respectively, averaged over 1000 Monte Carlo rounds.","tokens_in":5877,"tokens_out":3741,"duration_ms":41239,"significance":"If the reported gains were real, a simple re-indexing of RBF hidden units to incorporate temporal structure would be a practically useful contribution for nonlinear system identification. The paper also provides a link to the implementation, which is a positive reproducibility feature, and it fixes the nonlinear system and input signal explicitly. However, the current evidence does not support the central claim. The learning-rate update in Eq. (11) is the standard delta rule with a re-indexed weight, so the theoretical contribution is not new. The empirical comparison is confounded by a 500x difference in learning rate, an undefined temporal expansion with an increased number of hidden units, and a lack of error bars. The significance of the architecture itself is therefore not established in this manuscript.","major_comments":[{"comment":"The reported performance advantage is confounded with a 500x larger learning rate. The proposed ST-RBF uses eta = 1e-2, while both baselines use eta = eta_v = 2e-5. A larger step size alone can plausibly move a gradient-descent method from a poorly converged steady-state MSE near -1.7 dB to a much lower error, independent of any architectural benefit. The comparison must include baselines trained with the same learning rate (or a learning-rate sweep) and the same number of effective hidden units to isolate the effect of the spatio-temporal indexing. As written, the claim in the abstract that ST-RBF achieves 'significantly reduced estimation error' is not supported.","section":"Section III, simulation setup"},{"comment":"The proposed model is not fully specified. The basis functions phi_(i,t)(x, c_(i,t)) and the centers c_(i,t) are never defined; the only basis functions given in Eqs. (2)-(4) are functions of x and c_i without the temporal index. The truncation parameter T is also never assigned a numerical value or explained. Without these definitions, the architecture in Eq. (5) cannot be reproduced or tested, and the question of whether the double index (i,t) represents a true temporal expansion or merely a larger static basis cannot be answered. Additionally, Eq. (5) sums over S*T terms, so the comparison against a baseline with S=6 hidden units changes network capacity as well as architecture.","section":"Section II, Eq. (5)"},{"comment":"The paper claims 1000 independent Monte Carlo rounds, but reports only the mean training and test MSE values with no variance, confidence intervals, or statistical significance tests. Without this information, the difference between -19.67 dB and -4.431 dB cannot be distinguished from a hyperparameter artifact. The statement that the proposed method 'outperforms' the baselines is therefore not statistically substantiated.","section":"Section III, test-phase MSE reporting"},{"comment":"The text says the proposed ST-RBF learning rate is 'half of the RBF and FRBF learning rates i.e., 1e-2', but the baseline learning rate is 2e-5, so 'half' would be 1e-5, not 1e-2. This internal inconsistency suggests the 500x increase may be unintentional, and it further undermines confidence in the experimental setup. The authors should correct the statement or the value and explain which setting was actually used.","section":"Section III, learning-rate sentence"}],"minor_comments":[{"comment":"The heading contains the typo 'statio-temporal RBFNN' and should read 'spatio-temporal RBFNN'.","section":"Section II.A, heading"},{"comment":"The phrase 'calclulated' should be 'calculated'.","section":"Section III, testing paragraph"},{"comment":"The phrase 'Monte Carlos simulations protocol' is nonstandard; it should be 'Monte Carlo simulation protocol'.","section":"Section III, simulation setup"},{"comment":"The claim that the proposed output 'shows the exact match' is qualitative; a quantitative error metric or a zoomed plot would be more informative.","section":"Section III, Fig. 5"}],"recommendation":"reject","confidential_remarks":"The paper is very short and appears to be an extended abstract; its central empirical claim rests on a single confounded simulation. The definitional gaps in Eq. (5) and the learning-rate inconsistency are serious enough that I cannot recommend major revision: the contribution, as currently written, is essentially a re-indexed delta rule, and the claimed evidence for its benefit is not reliable. The heavy self-citation pattern (references [5]-[13] and [14]) is not itself a reason for rejection, but it does not compensate for the missing comparison with a matched-capacity RBF baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper is a re-indexing of the RBF output from S hidden units to S×T hidden units, with T never defined and the basis function φ(i,t) never specified. The derivation of the weight update reduces to the standard RBF delta rule. The empirical section claims a ~15 dB test-MSE improvement over conventional and fractional RBF baselines, but that comparison is badly confounded.\n\nWhat is actually new: essentially nothing architecturally. Equation (5) is Eq. (1) with the index range changed, and Eq. (11) is textbook gradient descent for an RBF. The authors do not explain how the temporal index enters the kernel, so the \"spatio-temporal\" label is doing no mathematical work. The paper is at least clearly written, and the derivation, such as it is, is algebraically correct. They also posted MATLAB code, which is more than many submissions do.\n\nThe soft spots are the load-bearing ones. The proposed ST-RBF runs with learning rate η=1e-2 while the baselines use η=2e-5 — a 500× difference that alone could explain the entire gap. The text even says the proposed η is \"half\" of the baseline, which is internally inconsistent. The proposed network has S×T hidden units versus S=6 for the baselines, adding capacity. T is never defined. No error bars or variance are reported despite 1000 Monte Carlo runs, so the −19.67 dB figure is just one number. The stress-test note is right: matching the learning rate and using an ordinary RBF with S×T hidden units would likely reproduce the reported gains, making the claimed architectural benefit an artifact of hyperparameter choice.\n\nI don't see circularity in the derivation, and the test signal is separate from training, so the problem is confounding, not fitting-to-the-test-set. The citation pattern is heavy self-citation, but that is common and not disqualifying by itself.\n\nMy take: this is not ready for publication. The empirical claim is unsupported as stated, and the novel content is a renamed increase in hidden units. A revision with a defined architecture, matched hyperparameters, equal capacity, and error bars could be reconsidered, but that is a re-do of the experiments, not a patch. I would not send it to peer review in its current form. For you: if you care about RBF variants for nonlinear system identification, the paper is a cautionary example of how not to run comparisons. No need to cite it.\n\nBest,\n[You]","headline":"A trivial re-indexing of RBF hidden units, with the claimed 10+ dB gains almost certainly an artifact of a 500x learning-rate mismatch and extra hidden units.","tokens_in":6388,"tokens_out":1222,"would_cite":false,"duration_ms":14340,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that indexing radial-basis hidden units by both space and time turns a standard RBF network into a markedly faster, more accurate identifier of nonlinear dynamic systems.","keywords":["spatio-temporal RBF","nonlinear system identification","radial basis function networks","gradient descent learning","fractional gradient descent","time-space orthogonality","kernel methods"],"falsifier":"Run the same nonlinear system identification task with the standard RBF given the same learning rate $10^{-2}$, the same $S\\times T$ hidden units, and the same basis function used for $\\varphi_{i,t}$, holding every other setting fixed. If the standard RBF then reaches approximately the same training and test MSE as the spatio-temporal RBF, the reported advantage is not due to the spatio-temporal architecture.","tokens_in":5379,"feed_emoji":"📉","tokens_out":6310,"duration_ms":57834,"temperature":0.7,"pith_summary":"This paper proposes a spatio-temporal radial basis function network for nonlinear system identification. The central idea is to give each hidden unit both a spatial index and a temporal index, so the output becomes a double sum over S spatial and T temporal terms, and gradient descent updates each weight using the corresponding basis response and the instantaneous error. The paper reports that on a benchmark nonlinear system with 1000-sample training and 200-sample test signals, the proposed network converges to a training MSE of -15.1286 dB and a test MSE of -19.67 dB, compared with roughly -1.7 dB and -4.4 to -5.0 dB for standard and fractional RBF networks. A sympathetic reader would care because the change is architectural and simple: if the gain holds with controlled hyperparameters, a small re-indexing of RBF units could improve temporal system identification without the data demands of deep networks.","feed_headline":"Splitting RBF units over time and space cuts error 15 dB","feed_subtitle":"Re-indexing RBF hidden units by time and space beats standard and fractional RBF in identification.","key_machinery":"The central object is the spatio-temporal RBF hidden layer, whose output is $y(k)=\\sum_{i=1}^S\\sum_{t=1}^T w_{i,t}(k)\\varphi_{i,t}(x,c_{i,t})+b(k)$. Each hidden unit carries both a spatial index $i$ and a temporal index $t$, giving $S\\times T$ effective units, and the weight update derived from the squared-error cost is $w_{i,t}(k+1)=w_{i,t}(k)+\\eta\\,\\varphi_{i,t}(x,c_{i,t})e(k)$. The mechanism carries the argument because every reported learning curve is a direct test of whether this re-indexed expansion plus the associated weight update identifies the nonlinear system.","core_discovery":"The paper's central claim is that separating time and space in the RBF hidden layer—modeling temporal dynamics with one index and spatial nonlinearity with another—yields a substantially better nonlinear system identifier than the standard architecture under gradient-descent training. The proposed model replaces $y=\\sum_{i=1}^S w_i\\varphi_i(\\|x-c_i\\|)+b$ by $y(k)=\\sum_{i=1}^S\\sum_{t=1}^T w_{i,t}(k)\\varphi_{i,t}(x,c_{i,t})+b(k)$, and the weight update becomes $w_{i,t}(k+1)=w_{i,t}(k)+\\eta\\,\\varphi_{i,t}(x,c_{i,t})e(k)$. The authors interpret this as exploiting time-space orthogonality, and they report that the spatio-temporal network reaches a training MSE of $-15.1286$ dB and a test MSE of $-19.67$ dB, against $-1.6813$ dB training and $-4.431$ dB test for the conventional RBF, and $-1.7444$ dB training and $-4.955$ dB test for the fractional RBF. The stated conclusion is that the spatio-temporal architecture, trained by the same gradient-descent rule adapted to the double sum, is what delivers the improvement.","pith_inferences":["The reported comparison changes several variables at once—learning rate ($10^{-2}$ versus $2\\times10^{-5}$), number of hidden units ($S\\times T$ versus $S$), and the unspecified $T$ and $\\varphi_{i,t}$—so the size of the architectural contribution is an open question; a controlled ablation with matched learning rate and matched parameter count is the natural next experiment.","If the benefit is genuinely architectural, the temporal index likely acts as an embedded delay line, making the spatio-temporal RBF a kernelized relative of tapped-delay-line and temporal-convolution models; this connection is not explored in the paper.","A direct testable extension is to report $T$ and the explicit form of $\\varphi_{i,t}(x,c_{i,t})$, since the paper leaves both undefined; specifying them would let others reproduce the network exactly."],"forward_implications":["On temporal system-identification tasks, substituting the double-sum spatio-temporal hidden layer for the standard RBF layer should yield lower steady-state error and faster convergence under the same squared-error cost.","The derived update rule is online and local, so it can be dropped into existing RBF training loops with negligible extra computation.","Because the derivation uses only the chain rule, the same spatio-temporal indexing applies to any differentiable kernel, such as Gaussian, multiquadric, or inverse multiquadric, and to gradient variants including fractional-order updates.","The approach preserves RBF's light data requirement and simple three-layer structure, so it offers a low-cost alternative to deep architectures for nonlinear identification."],"supporting_citations":[{"why":"Supplies the fractional gradient-descent RBF baseline and the standard RBF architecture that the proposed network is compared against.","marker":"[14]"},{"why":"Provides Cover's theorem, which the paper uses to justify mapping the input into a higher-dimensional kernel space.","marker":"[15]"},{"why":"Defines the RBF architecture and the Gaussian, multiquadric, and inverse multiquadric kernel functions used in the equations.","marker":"[16]"},{"why":"Cited as the source of the least-squares gradient-descent training method on which the weight update derivation is based.","marker":"[17]"}],"fun_headline_variants":["Time-space RBF separates dynamics, improves identification","RBF with time and space axes beats standard and fractional","Orthogonal time and space in RBF speeds convergence","Spatio-temporal RBF reaches -19.67 dB test MSE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that the reported gain comes from the spatio-temporal indexing, but the simulations also use a much larger learning rate ($10^{-2}$ versus $2\\times10^{-5}$), more hidden units ($S\\times T$ versus $S$), and leave $T$ and the exact $\\varphi_{i,t}(x,c_{i,t})$ unspecified; if those changes, not the indexing, produce the gain, the central claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Time-space RBF separates dynamics, improves identification","RBF with time and space axes beats standard and fractional","Orthogonal time and space in RBF speeds convergence","Spatio-temporal RBF reaches -19.67 dB test MSE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000505,"raw_usage":{"total_tokens":2437,"prompt_tokens":894,"completion_tokens":1543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1476}},"tokens_in":510,"tokens_out":1543,"duration_ms":11121,"temperature":1.0,"reasoning_tokens":1476,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:17:09.759947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same nonlinear system identification task with the standard RBF given the same learning rate $10^{-2}$, the same $S\\times T$ hidden units, and the same basis function used for $\\varphi_{i,t}$, holding every other setting fixed. If the standard RBF then reaches approximately the same training and test MSE as the spatio-temporal RBF, the reported advantage is not due to the spatio-temporal architecture.","supporting_citations":[{"cited_title":"Haykin, Neural Networks: A Comprehensive F oundation , 2nd ed","cited_arxiv_id":null,"evidence_quote":"Defines the RBF architecture and the Gaussian, multiquadric, and inverse multiquadric kernel functions used in the equations."}],"review_version":1}