{"id":"5c4428a1-5f99-4c65-9935-e2b99f26159d","arxiv_id":"2412.05667","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Training each neuron of a multilayer perceptron with its own particle-swarm search, instead of using backpropagation, reaches accuracy comparable to gradient-based training on two real datasets and several synthetic ones.","lead":"This paper describes a way to train a neural network without the standard backpropagation algorithm: each neuron is optimized independently by a small particle-swarm search, and the best results are combined. A reader might care because it is another step toward training networks with only forward computations, which could avoid backpropagation's memory and gradient problems, though the demonstrations are small-scale.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported accuracies may be inflated because the validation-loss gate in Algorithm 1 selects final weights using data that may also be the scored held-out fold; the 'similar to MLP' claim is not yet established.","rationale":"Why this concern is load-bearing: the paper's headline is that a backpropagation-free method matches backpropagation-trained MLPs. That is an empirical assertion, and the only real-data evidence is Tables 1 and 2 from k-fold cross-validation. The algorithm's only mechanism for preventing the 'runaway' behavior admitted in Section 4 is the validation-loss gate (Algorithm 1, Step 14). Using that gate's objective as the evaluation metric is a direct form of selection on the test set. The paper neither documents a separate test partition nor provides code inspection; the repository is referenced but no commit hash or executed artifact is given. This makes the empirical claim currently unverifiable and potentially biased. I am not arguing the method cannot work; rather, the current write-up does not rule out the most common experimental-contamination failure mode. The reader's weakest_assumption also flagged the validation gate; I single it out as the decisive issue because the other concerns (lack of convergence proof, hyperparameter reporting) would not by themselves falsify the empirical claim, whereas test-set selection would. Verdict: keep CONDITIONAL, explicitly conditioned on demonstrating a valid train/validation/test separation; if the check shows leakage, the verdict should move to REJECT.","tokens_in":15198,"tokens_out":7598,"duration_ms":74927,"concrete_test":"Open the GitHub repository and identify the validation-label tensor passed to Algorithm 1 for the Rice and Dry Bean experiments; check whether it is the same data as the held-out fold used to compute the accuracy tables. Then rerun the two real-data experiments with a strict three-way split: a training partition, a disjoint validation partition for the Step 14 gate, and a test partition never used for selection (or nested cross-validation). If the code already has such a partition, the concern is resolved; if not, or if the proposed method's accuracy drops by more than the reported standard errors relative to MLP, the comparable-performance claim should be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (abstract; Section 5; Tables 1 and 2) is empirical: the proposed per-neuron PSO training reaches accuracies comparable to backpropagation-trained MLPs. For that claim to be meaningful, the reported accuracies must be unbiased estimates of generalization. Algorithm 1 explicitly requires 'independent validation data and labels' and, at Step 14, retains a network update only when the new validation loss is lower than the previous one. The final reported network is therefore the result of a selection process on validation loss. Sections 6.5 and 6.6 describe only k-fold cross-validation and never state that a separate, never-used-for-selection test split was reserved inside each training fold. If the validation data used in Step 14 is the same held-out fold whose accuracy is reported, the final network has effectively been selected on the test set, so the accuracies in Tables 1 and 2 are optimistically biased. This bias could explain part or all of the claimed similarity to MLP, and it is independent of whether the coordinate-wise PSO updates have a convergence proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a derivative-free training method for multi-layer perceptrons in which each neuron is equipped with its own set of PSO particles and is updated coordinate-wise while all other weights are held fixed. A validation-loss gate in Algorithm 1 accepts or rejects whole-network updates, and the method is evaluated on four synthetic problems and two real tabular datasets (Rice and Dry Bean). The authors claim that this per-neuron PSO approach overcomes both local-minima problems of gradient descent and the search-space limitation of whole-network PSO, achieving accuracies and loss curves comparable to backpropagation-trained MLPs.","tokens_in":15435,"tokens_out":5096,"duration_ms":44153,"significance":"If the empirical claims hold, the paper offers a simple, forward-only, derivative-free training scheme with per-neuron parallelism that is easy to implement and that the authors have released as open-source code. Those are genuine strengths. However, the evidence is limited to two small tabular datasets and synthetic examples, and the central comparability claim is weakened by the evaluation-protocol ambiguity discussed below, by missing baseline error bars, and by an unproven premise about decomposing the joint loss into per-neuron subproblems. The contribution is incremental relative to the large literature on evolutionary and derivative-free neural-network training, but the per-neuron decomposition is a distinct and potentially useful variation.","major_comments":[{"comment":"The final network is selected based on validation loss (Steps 6 and 14), but the paper never states that a separate test split, not used for this selection, is reserved inside each cross-validation fold. Sections 6.5 and 6.6 only say '4-fold cross-validation' and '13-fold cross-validation.' If the validation set used in Step 14 is the same held-out fold whose accuracy is reported in Tables 1 and 2, the reported accuracies are optimistically biased because the network has effectively been selected on the test set. Please clarify the exact data split (e.g., train/validation/test within each fold) and, if the validation gate used the test fold, re-run the experiments with nested splits before the comparability claim can be assessed.","section":"Algorithm 1, Steps 6 and 14; Sections 6.5 and 6.6"},{"comment":"The method's core premise is that fixing all other weights turns each neuron into an independent subproblem, but the joint loss is not separable into per-neuron losses. Updating each neuron by its own best particle does not necessarily decrease the joint loss; the paper itself notes in Section 4 that 'the nodes together perform runaway operations' and relies on the validation-loss gate in Algorithm 1 for stability. No convergence or quality guarantee is provided for this coordinate-wise PSO procedure. Please either provide a formal statement of what the per-neuron updates optimize (e.g., a descent or monotone-improvement condition on the joint loss) or substantially temper the Abstract and Introduction claims that the method 'overcome[s] the problem of gradient descent and the limitation of the PSO algorithm.'","section":"Sections 4.1 and 5; Eq. (5)"},{"comment":"The comparability claim rests on the last three rows of Tables 1 and 2, but standard deviations are reported only for the proposed PSO rows, not for the MLP or other baselines. Overlapping or comparable error bars are necessary to support 'performance is similar to MLP.' Please report means and standard deviations (or confidence intervals) for all methods and specify the MLP architecture, optimizer, learning rate, epochs, and any regularization used for the baselines.","section":"Tables 1 and 2"},{"comment":"The free PSO parameters—inertia weight w, acceleration constants c1 and c2, craziness constant c3, number of particles per neuron, batch size, weight initialization range, and stopping criterion—are not reported anywhere in Section 6 or in Algorithm 1. Without these values, the experiments are not reproducible from the text alone; the GitHub link is helpful but does not replace a hyperparameter table. Please add a table listing the values used for each dataset and synthetic experiment.","section":"Section 6 (all subsections)"}],"minor_comments":[{"comment":"The flowchart text is garbled and largely unreadable in the PDF; please replace it with a clear vector figure with legible font.","section":"Figure 3"},{"comment":"The MLP and the proposed method use different output-layer configurations (one output node with MSE loss vs. two output nodes with BCE loss), so the loss curves in Figures 4–7 are not directly comparable. Please explain how the comparison is intended despite the differing losses.","section":"Section 6.2"},{"comment":"The phrase 'follows the gradient descent' is misleading because the proposed method is derivative-free; it appears to mean only that the loss curve decreases similarly. Please rephrase to avoid implying gradient-based behavior.","section":"Sections 6.1–6.4 and 7"},{"comment":"Step 7 says loss values are set to zero after each epoch, but Eq. (5) accumulates batch losses into f_j_ik. Please clarify whether the accumulated loss is reset per epoch and how the personal-best comparison in Steps 9–11 interacts with this reset.","section":"Algorithm 1, Step 7 and Eq. (5)"},{"comment":"The statement that 'the original PSO paper had removed the craziness term' is a historical claim without a citation; please either cite the source or rephrase to describe only the version used here.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The validation-bias issue is the key gating risk. Even if the authors clarify that a separate validation split was used, the absence of baseline error bars and the lack of hyperparameter reporting make the empirical claim weaker than stated. The paper's scope is modest; whether the venue values the derivative-free, parallelizable per-neuron training idea may depend on how convincingly the evaluation-protocol ambiguity is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look, but the headline claim is not yet established. The novel bit is real: instead of running PSO over the whole weight vector, you give each neuron its own sub-swarm, fix everything else, update that neuron, and gate the whole-network update on validation loss. That is block coordinate descent with PSO as the inner solver, and the paper is candid about the runaway instability and the redundant-computation cost. That honesty counts. The synthetic curves do what they claim: PSO without the craziness term stalls, and the proposed method reaches loss comparable to an SGD-trained MLP. On Rice and Dry Bean, the accuracy numbers sit within about a point of the MLP baseline.\n\nThe soft spots are serious. First, the validation gate in Algorithm 1 can undermine the accuracy tables. Step 14 keeps a network update only if the new validation loss is lower, and the method needs 'independent validation data.' The experimental section only mentions k-fold cross-validation; it never says a separate test split was reserved inside each fold for final evaluation. If the holdout fold used for the gating is the same one whose accuracy is reported in Tables 1 and 2, those numbers are optimistically biased, and the 'similar to MLP' claim is unproven. That has to be fixed before the empirical result can be trusted. Second, there is no convergence or guarantee argument, and the per-neuron decomposition is motivated only by intuition. The abstract's language about 'overcoming' gradient descent and PSO limitations is too strong. Third, error bars exist only for the two PSO variants, not for the MLP or classical baselines, and hyperparameters are under-reported. The code is linked, but no commit hash or version is given, so I cannot pin down what exactly produced the tables.\n\nProportionately, this is a legitimate, simple idea that a competent referee can evaluate quickly. It is not a major breakthrough; the contribution is a modest extension of coordinate descent and cooperative PSO. But it is implementable, and the stability issue is acknowledged in the text. I would send it to peer review, with the referee instructed to require a clean train/validation/test split and to rerun the tables. The paper is not ready as is; a revised version with a proper split and full hyperparameters would be credible. The audience is people working on derivative-free or forward-only training methods, not those looking for large-scale results.","headline":"The per-neuron PSO idea is clear and implementable, but the validation-loss gate in Algorithm 1 may be selecting on the same held-out fold whose accuracy is reported, which means the central 'similar to MLP' claim is not yet established.","tokens_in":15939,"tokens_out":3073,"would_cite":false,"duration_ms":31156,"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":"The paper reports that a multi-layer perceptron can be trained without backpropagation by giving each neuron its own particle swarm, reaching accuracies on Rice and Dry Bean datasets comparable to a backpropagation-trained MLP.","keywords":["particle swarm optimization","backpropagation-free training","per-neuron training","forward-only learning","multi-layer perceptron","derivative-free optimization","validation gating","craziness term"],"falsifier":"Train the method on a synthetic classification task whose solution requires two hidden neurons to change their weights simultaneously, so that no single-neuron update reduces the loss; if per-neuron PSO with validation gating cannot match a backpropagation-trained baseline on that task, the central claim fails.","tokens_in":14993,"feed_emoji":"🧠","tokens_out":6801,"duration_ms":52928,"temperature":0.7,"pith_summary":"This paper claims that a multi-layer perceptron can be trained without backpropagation by treating each neuron as an independent search problem. For every neuron, a small swarm of particles explores candidate weight vectors while all other weights are held fixed, and the best particle's weights are selected; a validation-loss gate decides whether the combined new weights are kept. On synthetic classification tasks and on the Rice and Dry Bean datasets, the author reports loss curves and accuracies comparable to a backpropagation-trained MLP. If the claim holds, it offers a derivative-free, forward-only training alternative for small neural networks, with per-neuron updates that can be parallelized.","feed_headline":"Train MLPs with a per-neuron particle swarm, no backprop","feed_subtitle":"Per-neuron PSO reaches backprop-level accuracy on Rice and Dry Bean datasets — no gradients needed.","key_machinery":"The central object is the per-neuron particle swarm. Each neuron node hosts k PSO particles, each particle being a candidate weight vector for that node; the velocity update includes the standard PSO terms plus a craziness term $c_3 r_3$ to keep particles exploring. While one node is updated, all other network weights are fixed, so the node's loss subproblem depends only on its own particle positions. The per-node best particles are combined into the network weights, and a validation-loss gate (Equation 5) decides whether the new combined weights are kept, which restricts runaway behavior. This machinery replaces the backward pass with k forward loss evaluations per node, repeated batchwise across epochs.","core_discovery":"The central claim is that individual neuron nodes can be trained independently with particle swarm optimization and that the collective behavior of these separately trained neurons solves classification problems as effectively as backpropagation. The author splits the network into sub-problems: for each node, k particles carry candidate weight vectors; the loss is computed for each particle while all other node weights are frozen, and the minimum-loss particle is chosen. A validation-loss comparison after each batch decides whether the new combined weights replace the previous global best, which the author says contains the runaway problem of unconstrained per-node updates. The reported result is that the proposed method follows the same loss-reduction path as gradient descent and reaches accuracies within the spread of the MLP baseline on the Rice and Dry Bean datasets, while avoiding backpropagation entirely.","pith_inferences":["A comparison the paper does not report is against plain coordinate-wise random search without the PSO velocity memory; without it, the contribution of the swarm dynamics versus the per-neuron decomposition alone is untested.","The validation-loss gate repeatedly selects network states on the same validation data, which can bias the reported metrics; a clean test would hold out a third split for final evaluation.","The method's cost is k forward passes per neuron per batch, which the paper itself flags as redundant computation; scaling to larger networks would likely require sharing or approximating per-neuron loss evaluations.","If the approach scales, the per-neuron scalar reward resembles local learning rules in neuroscience, where each neuron aligns to a global performance score instead of receiving a backpropagated gradient."],"forward_implications":["Training requires no backward pass, so errors do not need to be transported backward through the network; weight updates come from forward loss evaluations alone.","Each neuron's weight update is independent, so node updates across the network can be computed in parallel without waiting for other neurons.","The method is insensitive to the choice of activation function in the reported tests: replacing ReLU with Tanh left the loss-reduction behavior unchanged.","Because updates are separated per node, the approach opens the door to arbitrary, non-standard connectivity between neurons without needing a global backpropagation rule.","On the Rice and Dry Bean datasets, the reported accuracy, specificity, and F1-scores land within one standard deviation of the MLP baseline in Tables 1 and 2."],"supporting_citations":[{"why":"Supplies the particle swarm optimization algorithm whose velocity and position updates each neuron's particles follow.","marker":"[18]"},{"why":"Introduces the craziness term included in the velocity update to keep particles exploring the search space.","marker":"[24]"},{"why":"Defines backpropagation, the training algorithm the proposed method claims to replace.","marker":"[33]"},{"why":"Provides the Rice dataset used for the first real-data comparison against MLP and other classifiers.","marker":"[7]"},{"why":"Provides the Dry Bean dataset used for the second real-data comparison.","marker":"[21]"},{"why":"Earlier PSO application cited when motivating the craziness term and the limited search space problem that per-neuron training addresses.","marker":"[25]"}],"fun_headline_variants":["No backprop: per-neuron particle swarm trains MLPs","Particle swarm per neuron: backprop-free MLP training","Neural nets without backprop, via per-neuron particles","Each neuron evolved by particles: no backprop needed","Per-neuron PSO replaces backprop for MLP training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that updating one neuron at a time while freezing every other neuron, and accepting the new weights only when validation loss improves, is enough to navigate the network's joint loss landscape to a good solution.","fun_headline_variants_meta":{"raw":{"variants":["No backprop: per-neuron particle swarm trains MLPs","Particle swarm per neuron: backprop-free MLP training","Neural nets without backprop, via per-neuron particles","Each neuron evolved by particles: no backprop needed","Per-neuron PSO replaces backprop for MLP training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000484,"raw_usage":{"total_tokens":2364,"prompt_tokens":894,"completion_tokens":1470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1386}},"tokens_in":510,"tokens_out":1470,"duration_ms":11397,"temperature":1.0,"reasoning_tokens":1386,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:29:08.594067+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the method on a synthetic classification task whose solution requires two hidden neurons to change their weights simultaneously, so that no single-neuron update reduces the loss; if per-neuron PSO with validation gating cannot match a backpropagation-trained baseline on that task, the central claim fails.","supporting_citations":[{"cited_title":"Particle swarm optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the particle swarm optimization algorithm whose velocity and position updates each neuron's particles follow."},{"cited_title":"Binary classifica tion posed as a quadratically con- strained quadratic programming and solved using particle s warm optimization","cited_arxiv_id":null,"evidence_quote":"Introduces the craziness term included in the velocity update to keep particles exploring the search space."},{"cited_title":"Learning representations by back-propagating errors","cited_arxiv_id":null,"evidence_quote":"Defines backpropagation, the training algorithm the proposed method claims to replace."},{"cited_title":"Classification of Rice V ari eties Using Artificial Intelligence Methods","cited_arxiv_id":null,"evidence_quote":"Provides the Rice dataset used for the first real-data comparison against MLP and other classifiers."},{"cited_title":"Quadratically constrained quadratic programming for classification using particle swarms and applications","cited_arxiv_id":"1407.6315","evidence_quote":"Earlier PSO application cited when motivating the craziness term and the limited search space problem that per-neuron training addresses."}],"review_version":1}