{"id":"01419f4c-3b9e-4843-9d78-207f57576ce3","arxiv_id":"2608.08317","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A modular spiking network with isolated per-class experts and a hidden-activity loss is claimed to match dense accuracy with far fewer spikes and parameters, though key training details and internal numbers are inconsistent.","lead":"This paper proposes a spiking neural network built from independent class-specific expert pathways, trained with a loss that suppresses activity in wrong experts. The authors report accuracy close to dense networks with much lower firing rates and synaptic operations, plus resistance to forgetting when old experts are frozen. The reported results contain internal contradictions and an unspecified output-layer training mechanism.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Output-layer readout is never trained or specified, so the reported SNN accuracies are not reproducible from the described method.","rationale":"The reader's weakest_assumption is exactly that the push-pull loss never touches the output layer, and the reader's verdict is REJECT. My reading confirms this: Eq. (7) is a sum over hidden activations only, while Sec. 4.8 states classification is made by the output neuron with the maximum number of spikes. No output-layer loss, freezing rule, reset rule, or scaling is described anywhere in the manuscript. A spike-count readout is not a differentiable function of the hidden excitations unless the output weights are chosen so that the correct block's hidden activity maps deterministically to the correct output neuron; the paper provides no such construction. If one instead imagined an ad hoc readout such as 'pick the class whose expert has the highest total hidden spike count', the paper does not state that either, and that rule would trivially differ from the reported confusion matrices. The FMNIST weight visualizations (Figure 3, FC3 (L2 -> Out)) show nonzero structured output weights, so weights must exist in the experiment, but the training objective that produced them is undocumented. That omission is fatal to reproducibility of the central claim. I also note the paper's own internal inconsistencies (parameter savings described as both two-fold and order-of-magnitude; firing-rate savings described as 'several orders of magnitude' while Table 4 shows at most ~10x; CIFAR-10 accuracy given as ~66% in the introduction and 82-83% in the results), which reinforce the reader's rejection, but the output-readout gap is the most fundamental, load-bearing concern because without a specified readout the reported numbers cannot be traced to the described method.","tokens_in":28320,"tokens_out":2082,"duration_ms":16131,"concrete_test":"Inspect the released repository and identify the exact rule that sets the output-layer weights for the independent-experts and dense-hybrid models, or re-run the training loop with the output layer excluded from every gradient update, trying several standard initializations (e.g., zero, uniform, He) and readout scales. If no initialization and no auxiliary loss reproduces the reported accuracies (98.15% MNIST, 91.68% FMNIST, 82.7%/83.3% CIFAR-10), then the described method is under-specified and the claimed results are not reproducible from Eqs. (6)-(7) alone.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that D-SNN matches dense accuracy with an order-of-magnitude fewer parameters, lower firing rates, and forgetting protection. Every reported accuracy is computed from output-neuron spike counts (Sec. 4.8). Yet the push-pull loss L_PP in Eq. (7) is defined exclusively over hidden-layer activities z_alpha of Eq. (6): it measures S_true and S_false within the two hidden expert layers only. The output layer therefore receives no gradient under L_PP. The paper never specifies output-weight initialization, scaling, a secondary loss, or any readout-mapping rule. Since z_alpha is strictly positive and L_PP only balances total activity between correct and incorrect expert blocks, it does not constrain which output neuron fires most. Unless the omitted code contains an undocumented readout rule, the accuracies, SynOps, and forgetting results cannot be generated by the method as stated. This is not merely a presentation gap: it is the single most load-bearing assumption because all downstream claims are measured through that untrained output layer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Decomposable Spiking Neural Network (D-SNN) in which the hidden layers are partitioned into class-specific experts with block-diagonal connectivity. Training is performed in a continuous ANN using a 'Push-Pull' loss L_PP (Eq. 7) that compares the total hidden-layer activity in the correct expert block with the total activity in all wrong blocks; the learned weights are then ported directly into a Leaky Integrate-and-Fire SNN, and classification is read out over T=100 simulation steps as the output neuron with the most spikes. The authors report results on MNIST, Fashion-MNIST, and CIFAR-10/100, claiming that this modular design matches dense baselines in accuracy while using far fewer parameters, producing much lower firing rates and synaptic operations, and providing protection against catastrophic forgetting when experts are frozen during continual learning.","tokens_in":28416,"tokens_out":8340,"duration_ms":77101,"significance":"If the claims were fully supported, the paper would offer a simple, transparent, and parameter-efficient recipe for modular spiking classifiers, and the three-way comparison among dense CCE, dense hybrid, and independent experts is a useful decomposition that separates the effect of the loss function from the effect of structural isolation. The manuscript also has positive reproducibility signals: standard public datasets, a stated code repository, SNN-based verification of all reported accuracies, and a scaling study up to 100 classes. However, the central claims rest on a training procedure that is incompletely specified, and several headline statements are contradicted by the paper's own tables.","major_comments":[{"comment":"The training loss L_PP in Eq. (7) is a function only of the hidden-layer post-activations z_alpha defined in Eq. (6); the output layer never appears in L_PP, no secondary loss for output weights is defined, and no initialization or scaling rule for the output layer is stated. Since Section 4.8 determines classification from the output neuron with the maximum spike count, the described training procedure cannot be expected to produce the reported accuracies unless some omitted readout mechanism exists. This is load-bearing because every accuracy, efficiency, and forgetting result is measured through that output readout. The authors must specify exactly how the output weights are trained or initialized, or add an explicit output-loss term, before the experimental claims can be evaluated.","section":"§4.6/§4.8, Eq. (7)"},{"comment":"The reported MNIST baseline is inconsistent across the paper: Section 2.3 states dense cross-entropy achieves 98.19%, Table 1 reports 98.55% for Dense CCE, and Table 4 reports 98.18% for Dense CCE. A single consistent set of baseline numbers is required, since the central claim of accuracy parity depends on these comparisons.","section":"§2.3, Table 1, Table 4"},{"comment":"The abstract's claim of 'an order of magnitude fewer parameters' relative to fully dense networks is not supported by the paper's own tables. For MNIST, independent experts use about 272K parameters versus about 460K for Dense CCE (a factor of 1.7); for Fashion-MNIST the ratio is about 351K versus 539K (a factor of 1.5); and for CIFAR-10 the independent experts and dense hybrid have essentially identical parameter counts (≈350K versus ≈352K). The text in Section 2.3 correctly describes a 'two-fold reduction' in some places, so the abstract and the phrase 'an order of magnitude' should be corrected to match the actual data.","section":"Abstract, §2.3, Tables 1–3"},{"comment":"The continual-learning comparison is confounded by an asymmetric freezing protocol: in Phase 2, the independent experts and dense hybrid freeze all weights belonging to old-class experts, while the dense CCE baseline keeps all weights trainable. The paper's own control experiment, reported in Section 2.8, shows that when old experts are left plastic the independent experts' accuracy on legacy tasks drops to about 40%, indicating that structural isolation alone is not sufficient. The claim that physical separation provides 'inherent protection' against catastrophic forgetting should be either retracted or supported by a comparison in which the dense baseline receives an analogous class-dependent freezing or regularization scheme.","section":"§2.8, §4.11"},{"comment":"The large SynOps reductions attributed to the method are largely forced by the block-diagonal architecture rather than by learning: for a dense backend the number of backend connections scales as O(M^2), while a block-diagonal expert mask reduces this to O(M) by construction. Similarly, L_PP in Eq. (7) explicitly suppresses activity in non-target experts, so a substantial part of the firing-rate reduction is a direct effect of the objective. The paper should separate architectural arithmetic from learned sparsity when presenting the efficiency results, and should not present the SynOps scaling in Section 2.7 as a discovered phenomenon.","section":"§2.5, §2.7, Table 4"}],"minor_comments":[{"comment":"Section 4.5 states that the shared receptor for CIFAR-10/100 comprises 'two' trainable 3x3 convolutional layers, whereas Section 4.10 describes a three-layer convolutional topology with two 5x5 layers and one 3x3 layer; this discrepancy should be resolved.","section":"§4.5 vs §4.10"},{"comment":"The caption of Table 4 says MNIST and Fashion-MNIST both used a base filter size of F=16, but Section 4.9 specifies MNIST with F=8 and Fashion-MNIST with F=16; please align the caption with the architecture descriptions.","section":"Table 4 caption vs §4.9"},{"comment":"The CIFAR-10 accuracy values in Table 4 (82.74, 83.27, 82.82) are consistent with the text, but Table 3 omits the Dense CCE row entirely; adding it would make the three-way comparison easier to verify.","section":"Table 4, CIFAR-10 rows"},{"comment":"Several references are incomplete or inconsistently formatted, for example [29] appears as a bare arXiv identifier without authors or title; the reference list should be checked for completeness.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript links to a public GitHub repository. I did not inspect that repository; if the code contains an output-layer training or readout rule that is absent from the manuscript, the authors must still document it in the paper, and the reported numbers should be re-verified under the stated protocol. If no such rule exists, the experimental section is not reproducible and rejection would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a promising but currently unreproducible paper. The block-diagonal spiking expert idea combined with a hidden-activity log-ratio loss is a reasonable combination, and the three-way comparison (dense CCE, dense hybrid, independent experts) is a clean way to separate the loss effect from the architectural effect. But the loss in Eq. (7) is defined over hidden activations only, while all reported accuracies come from output spike counts. The output weights are never trained and their initialization is never specified, so the numbers cannot be generated by the method as written. That stress-test concern holds up on reading.\n\nWhat the paper does well: the push-pull loss is a well-defined, hyperparameter-free objective that directly encourages activity localization; the ANN-to-SNN conversion with direct input encoding is a sensible, cheap way to get deterministic verification; and the code is linked. The authors also honestly note the structural similarity to categorical cross-entropy, which is the right kind of self-awareness.\n\nThe soft spots are real. The output-layer gap is the load-bearing one; if the code contains an undocumented readout rule, it needs to be in the paper. Second, the headline numbers overstate the tables: parameter savings are about two-fold, not an order of magnitude (the FMNIST methods section even says '10-fold' while Table 2 shows ~351K vs ~539K); CIFAR-10 accuracy is 66% in the introduction and 83% in Section 2.4; firing-rate savings in Table 4 are single-digit factors, not 'several orders of magnitude'. Third, the continual-learning advantage is largely from freezing old experts, which the dense baseline cannot do; the authors admit this, but the abstract's 'inherent protection' overstates it. Fourth, the SynOps scaling argument is fine as a gain ratio, but because total width is fixed, per-expert capacity shrinks as M grows; the claim that accuracy stays similar across M deserves scrutiny.\n\nBottom line: the direction is worth a referee. The output-layer specification is fixable, and the code may already contain the missing detail. But as submitted, the paper is not reproducible from the text, and the abstract overclaims. I would send it to review with a request for major revision: specify the readout, correct the numbers, and temper the language. This is a solid reject-in-current-form, resubmit-after-clarification, not a desk reject.","headline":"The push-pull expert idea is plausible, but the central training scheme never touches the output layer, so the reported accuracies are not reproducible from the text.","tokens_in":662,"tokens_out":1431,"would_cite":false,"duration_ms":40106,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","92B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A spiking network whose class pathways are physically isolated experts can match dense accuracy at a fraction of the parameters and synaptic operations, and resists catastrophic forgetting by freezing old experts.","keywords":["spiking neural networks","modular architecture","push-pull loss","one-vs-all experts","catastrophic forgetting","energy-efficient neuromorphic computing","ANN-to-SNN conversion","interpretability"],"falsifier":"Run the training protocol exactly as stated, initialize the output-layer weights randomly, and verify that no secondary loss touches them; the SNN readout (output neuron with the most spikes) would then be driven by untrained weights and cannot reproduce the reported 98% MNIST accuracy. Equivalently, computing the gradient of $L_{PP}$ with respect to the output weight matrix shows it is identically zero under Eq. (7), so any reported accuracy requires an output-weight initialization or auxiliary objective that the paper does not disclose.","tokens_in":27990,"feed_emoji":"🧠","tokens_out":8440,"duration_ms":71829,"temperature":0.7,"pith_summary":"The paper sets out to prove that a neural network does not need globally entangled weights to classify images. It reports a Decomposable Spiking Neural Network (D-SNN) in which each output class owns a structurally isolated 'expert' pathway, and a biomimetic push-pull loss forces activity into the correct expert while suppressing all others. The network is trained as a continuous ANN and then ported unchanged to a spiking network, with all reported accuracies measured in the spiking domain. The central claim is that this modular design matches fully dense baselines on MNIST, Fashion-MNIST, and CIFAR-10/100 while using roughly half the hidden parameters of the paper's own dense baselines and cutting firing rates and synaptic operations by one to three orders of magnitude. If true, it gives a practical route to auditable, energy-efficient neuromorphic classifiers without expensive spatio-temporal backpropagation.","feed_headline":"Modular spiking net keeps accuracy with 10x fewer operations","feed_subtitle":"Block-diagonal experts trained by a push-pull loss fire less, forget less, and make decisions auditable.","key_machinery":"The load-bearing object is the block-diagonal expert decomposition used together with the Push-Pull loss of Eq. (7): $L_{PP} = \\sum_m [\\log(S^m_{\\mathrm{false}} + S^m_{\\mathrm{true}}) - \\log(S^m_{\\mathrm{true}})]$, where $S^m_{\\mathrm{true}} = \\sum_{\\alpha \\in B_{k_m}} z_\\alpha$ and $S^m_{\\mathrm{false}} = \\sum_{\\alpha \\notin B_{k_m}} z_\\alpha$ are sums of hidden post-activations inside and outside the correct expert block. This objective acts only on hidden-layer activity, not on output logits, so the training signal itself enforces sparse, class-specific routing. The second ingredient is structural: weight matrices between expert blocks are zeroed by construction, which eliminates cross-talk, lowers the fan-out from $O(M^2)$ to $O(M)$, and lets the network be verified as $K$ independent One-vs-All classifiers.","core_discovery":"The discovery the authors report is that routing a push-pull loss through a block-diagonal expert architecture produces a spiking network whose classification accuracy is statistically indistinguishable from a dense network, while the hidden activity is confined almost entirely to the expert of the correct class. The loss is $L_{PP} = \\sum_m [\\log(S^m_{\\mathrm{false}} + S^m_{\\mathrm{true}}) - \\log(S^m_{\\mathrm{true}})]$, built from hidden-layer activities: $S^m_{\\mathrm{true}}$ sums the surrogate post-activations in the correct expert and $S^m_{\\mathrm{false}}$ sums them everywhere else. Optimizing this quantity forces wrong experts to stay metabolically quiet and the correct expert to dominate, so the decision can be read off from which physical pathway fired. In SNN verification this yields 98.15% on MNIST and 91.68% on Fashion-MNIST (matching dense baselines), 82.7-83.3% on CIFAR-10 with only 5 neurons per expert, and a synaptic-operations reduction of roughly 37- to 160-fold on the two small datasets. The same frozen-expert mechanism keeps old-class accuracy at 80-90% in split-MNIST continual learning, where the dense cross-entropy baseline falls to 20-30%.","pith_inferences":["A testable extension the paper does not report: use the total spike count in each expert as a per-class confidence score, since the activity matrices on CIFAR-100 are near-diagonal; this could give cheap rejection of ambiguous or out-of-distribution inputs.","The frozen-expert design hints at a physical neuromorphic architecture where experts occupy separate cores or chips and new classes are added by appending a unit, without retraining old ones; the paper motivates but does not build this hardware.","Because the SynOps advantage over dense networks is $O(M)$ rather than $O(M^2)$, the gain should widen on many-class problems beyond the 100-class scale tested here, such as large handwriting sets; a scaling measurement on such a dataset would test that prediction."],"forward_implications":["If the central claim holds, dense global coupling is unnecessary for these classification tasks: a per-class OvA expert layout reaches the same SNN-verified accuracy with a fraction of the synaptic operations.","Because training happens entirely in the continuous ANN and is ported unchanged to the SNN, the approach avoids BPTT and its memory overhead, cutting training time by roughly 10-50x on consumer hardware.","The structural $O(M)$ fan-out instead of $O(M^2)$ means the efficiency gain over dense networks should grow with the number of classes, as illustrated by the CIFAR-100 scaling tests up to $M=100$.","Freezing the weights of previously learned experts gives a direct, regularization-free defense against catastrophic forgetting in class-incremental learning, holding old-class accuracy near 80-90% in split MNIST.","Because only the correct expert fires during inference, every classification has an explicit physical origin, making the network auditable in a way dense networks are not."],"supporting_citations":[{"why":"Supplies the One-vs-All decomposition that gives each class its own expert pathway.","marker":"[42]"},{"why":"Provides the biologically motivated inhibitory-plasticity mechanism the push-pull loss is modeled on.","marker":"[44]"},{"why":"Supplies the three-layer convolutional 'retina' architecture and the ANN-to-SNN conversion baseline used for CIFAR-10/100.","marker":"[25]"},{"why":"Defines the spatio-temporal backpropagation (STBP) supervised baseline that the method's efficiency claims are measured against.","marker":"[37]"},{"why":"Provides the MNIST benchmark dataset for the accuracy and continual-learning experiments.","marker":"[50]"},{"why":"Provides the Fashion-MNIST benchmark for the accuracy and synaptic-ops comparisons.","marker":"[51]"},{"why":"Provides the CIFAR-10 and CIFAR-100 datasets used for scaling and class-count experiments.","marker":"[58]"},{"why":"Defines the SynOps metric used to quantify synaptic-operation efficiency.","marker":"[62]"},{"why":"Is the sleep-replay continual-learning method the D-SNN's freezing approach is positioned against.","marker":"[82]"}],"fun_headline_variants":["Brains inspire spiking AI that forgets less and fires less","Modular spiking net: 10x fewer parameters, better recall","Isolated expert neurons: fewer ops, no catastrophic forgetting","Modular spiking net: efficient, auditable, anti-forgetting","Biologically modular spiking net: fewer synapses, clearer decisions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies depend on the output-layer weights being set in a way the paper never specifies, because the push-pull loss in Eq. (7) is a function only of hidden-layer activations and therefore supplies no gradient to those weights.","fun_headline_variants_meta":{"raw":{"variants":["Brains inspire spiking AI that forgets less and fires less","Modular spiking net: 10x fewer parameters, better recall","Isolated expert neurons: fewer ops, no catastrophic forgetting","Modular spiking net: efficient, auditable, anti-forgetting","Biologically modular spiking net: fewer synapses, clearer decisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000865,"raw_usage":{"total_tokens":3784,"prompt_tokens":1010,"completion_tokens":2774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":2684}},"tokens_in":626,"tokens_out":2774,"duration_ms":17872,"temperature":1.0,"reasoning_tokens":2684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:09:55.297284+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the training protocol exactly as stated, initialize the output-layer weights randomly, and verify that no secondary loss touches them; the SNN readout (output neuron with the most spikes) would then be driven by untrained weights and cannot reproduce the reported 98% MNIST accuracy. Equivalently, computing the gradient of $L_{PP}$ with respect to the output weight matrix shows it is identically zero under Eq. (7), so any reported accuracy requires an output-weight initialization or auxiliary objective that the paper does not disclose.","supporting_citations":[{"cited_title":"Plasticity in inhibitory networks improves pattern separation in early olfactory processing.Communications biology, 8(1):590, 2025","cited_arxiv_id":null,"evidence_quote":"Provides the biologically motivated inhibitory-plasticity mechanism the push-pull loss is modeled on."},{"cited_title":"Spatio-temporal back- propagation for training high-performance spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the spatio-temporal backpropagation (STBP) supervised baseline that the method's efficiency claims are measured against."},{"cited_title":"Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278– 2324, 1998","cited_arxiv_id":null,"evidence_quote":"Provides the MNIST benchmark dataset for the accuracy and continual-learning experiments."},{"cited_title":"Learning multiple layers of features from tiny images","cited_arxiv_id":null,"evidence_quote":"Provides the CIFAR-10 and CIFAR-100 datasets used for scaling and class-count experiments."},{"cited_title":"Activity pruning for efficient spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the SynOps metric used to quantify synaptic-operation efficiency."},{"cited_title":"Sleep- like unsupervised replay reduces catastrophic forgetting in artificial neural networks","cited_arxiv_id":null,"evidence_quote":"Is the sleep-replay continual-learning method the D-SNN's freezing approach is positioned against."}],"review_version":1}