{"id":"22c30c85-0117-460a-b24a-273f52d25132","arxiv_id":"2505.08514","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A spiking neural network with offline-learned, frozen convolution kernels classifies five street-object classes at 91.6% accuracy, compared with 92.3% for a similarly shaped CNN.","lead":"The authors add a fixed convolutional layer to a spiking neural network, learning the image filters ahead of time from unlabeled images so the network avoids weight sharing. On a five-class street-object task the spiking network reaches 91.6 percent accuracy, within a point of a similar conventional network, but no energy savings are measured.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convolution kernels are learned from the full G set before the 5-fold split, so test-fold images influence the fixed feature extractor; the 91.58% accuracy estimate is potentially optimistic and the CNN comparison is unfair.","rationale":"The reader's verdict is CONDITIONAL with moderate confidence, largely because of unproven transferability, missing code, and a suspicion of test-fold leakage. I read the full text and find the leakage issue to be the most load-bearing: the Section II kernels are learned on the full G set defined in Section IV before any train/test split, so the held-out folds are not truly held out from the feature extraction stage. The central claim rests on the accuracy comparison in Table 1, and that comparison is only meaningful if the fixed convolutions do not use test images. This is testable and fixable, so the paper should be accepted only on condition that the authors rerun with nested CV and report both numbers. The transferability assumption, while openly acknowledged, is less critical for this particular benchmark because the kernels are both learned and tested on the same Neovision2 domain. Energy claims are explicitly deferred, so they do not affect the accuracy-based central claim. My proposed check directly settles whether the accuracy result survives a clean evaluation protocol.","tokens_in":8800,"tokens_out":3787,"duration_ms":38745,"concrete_test":"Rerun the full pipeline with nested cross-validation: for each of the 5 folds, run the Section II kernel-construction algorithm using only the 80% training images of that fold, then evaluate the CSNN on the held-out 20% with those kernels. Compare the mean held-out accuracy to the reported 91.58% (and SD) and to CNN2's 92.34%. If the mean falls by more than ~1 percentage point or the difference from CNN2 becomes significant, the reported result is not a valid generalization estimate and the central claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV constructs G from all Neovision2 clips, shuffles, and runs the Section II convolution-learning algorithm on this entire set to obtain the fixed kernels. The accuracy evaluation then uses 5-fold cross-validation on the same images. Because the kernels are fit prior to the split, every test fold has already contributed to the feature extractor. The CNN baselines, in contrast, are trained only on the training folds (one epoch), so the comparison in Table 1 is not apples-to-apples. This is not a mere theoretical quibble: the kernel-learning algorithm is a competitive unsupervised procedure that adapts to the statistics of G, and the reported CSNN accuracy (91.58%) may be inflated by this transductive leakage. If the kernels were re-derived on each training fold alone, the measured accuracy could drop and the 'comparable to CNN2' claim could fail. The paper does not report any experiment isolating this effect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a convolutional spiking neural network (CSNN) for image classification in which the convolutional kernels are fixed and learned offline from a set of unlabeled images from the target domain, thereby avoiding weight sharing during SNN operation. The kernels are learned by an iterative unsupervised algorithm (Section II) and ported into a single convolutional layer followed by mean pooling; classification is performed by the CoLaNET spiking classifier. The method is evaluated on a 5-class object classification task derived from the Neovision2 benchmark, reporting mean accuracy 91.58% (SD 1.1) over 5-fold cross-validation, compared to 92.34% (SD 0.42) for a similarly shaped CNN (CNN2) and 94.35% (SD 0.32) for a larger CNN (CNN1). The authors conclude that the CSNN achieves accuracy comparable to traditional CNNs while preserving plasticity locality.","tokens_in":9051,"tokens_out":5117,"duration_ms":42862,"significance":"If the result holds, the paper offers a practical way to combine convolutional feature extraction with spiking networks on neuromorphic hardware, addressing the locality-principle objection to weight sharing. The kernel-learning algorithm is presented as a standalone procedure, and the evaluation includes two conventional CNN baselines. However, the strength of the evidence is limited by the evaluation protocol and by the untested domain-generality assumption. The paper would be a useful contribution to the neuromorphic SNN literature if these issues are resolved; the current version's central claim rests on a single dataset with a potentially optimistic accuracy estimate.","major_comments":[{"comment":"The set G used to learn the fixed convolution kernels is constructed from all 155,881 Neovision2 images before the 5-fold split is performed. Consequently, every test fold participates in building the feature extractor, so the reported 91.58% accuracy is measured under transductive leakage. The CNN baselines, by contrast, are trained only on training folds (one epoch), so the comparison in Table 1 is not apples-to-apples. The authors should either re-learn the kernels on each training fold only, or explicitly quantify the effect of this leakage (e.g., by comparing with kernels learned on a disjoint image set). This is load-bearing because the conclusion that CSNN is 'comparable' to CNN2 depends on the accuracy estimate.","section":"Section IV, paragraph 'In such a way, we created the set G...'"},{"comment":"All algorithmic hyperparameters and the pooling-layer calibration coefficient are selected using the same Neovision2 data on which accuracy is evaluated. For example, B is set to the mean brightness of G, l is derived from NE and the assumed brightness scale, and wmin/wmax are based on the observed brightness distribution; the calibration coefficient is chosen to make the mean pooling firing frequency approximately 50 Hz, again on this dataset. No sensitivity analysis or independent validation is provided, and the CoLaNET hyperparameters (microcolumn count 22, learning rate 0.0035, weight bounds) are reported without a selection procedure. The reported accuracy may therefore be optimistically biased. The authors should describe how these values were chosen and provide evidence that the result is robust to reasonable variations.","section":"Section IV, hyperparameter choices (K, NC, s, B, l, wmin, wmax) and the proportionality coefficient"},{"comment":"The central assumption that kernels learned from a representative image set G transfer to other classification tasks in the same domain is stated as 'can hardly be proven formally but it follows from common sense' (Section I). The paper provides no experiment that isolates transferability: kernels are learned from and evaluated on the same Neovision2 image collection. Because the entire fixed-feature pipeline collapses if kernels are task-specific rather than domain-generic, the authors should test this assumption, for example by learning kernels on one clip set and classifying objects from another clip set, or by varying the classification task while keeping the kernel-learning set fixed.","section":"Section I, 'Domain-level convolutions' assumption"},{"comment":"The claim that CSNN accuracy is 'comparable' to CNN2 is based on a single 5-fold cross-validation run. The mean difference is 0.76 percentage points, but the standard deviations (1.1 for CSNN, 0.42 for CNN2) overlap, and no statistical test or repeated-run variability is reported. With only five folds, the difference may not be significant. The authors should provide a paired comparison across folds or multiple independent runs, and state whether the difference is statistically reliable.","section":"Table 1 and Section V"}],"minor_comments":[{"comment":"The captions for Figures 2 and 3 appear to be swapped: Figure 2 (color-coded matrices) is labeled 'Color coded values of the convolution matrices' and Figure 3 (the network diagram) is also labeled 'Color coded values of the convolution matrices'. The captions should be corrected.","section":"Figure captions"},{"comment":"The pseudo-code contains an unclear line 'cbrt ← 0 for all <b, r, t> for which r = p & t = q'; this appears to be a leftover variable that is never used. The authors should clarify or remove it, and also define the shape of the convolution output tensor c explicitly.","section":"Section II, pseudo-code"},{"comment":"The paper does not report the class distribution of the 155,881 images or the number of examples per target class in the 5-class task. Class imbalance could affect both the CNN and CSNN results, and reporting it would help interpret the accuracy numbers.","section":"Section IV, dataset composition"},{"comment":"Reference [10] is cited as the basis of the kernel-learning algorithm, but the relation to that algorithm is only qualitative; a precise description of the differences (e.g., the threshold-based update rule and the WTA competition) would help reproducibility.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the idea is interesting. The main risk is the evaluation protocol; I would recommend asking the authors to address the leakage issue and the transferability test before publication. I have no concerns about authorship integrity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper describes a way to build a convolutional SNN without weight sharing — kernels are learned offline on a large image set from the domain, then ported as fixed synapses into an SNN whose classifier is CoLaNET. The idea is genuinely useful for neuromorphic hardware where weight sharing is expensive. The kernel-learning algorithm itself is a reasonable standalone reworking of the STDP-based approach from [10], and the pipeline with CoLaNET is new.\n\nWhere it holds up: the algorithm is clearly specified, the kernel visualizations show structured features, and the experimental setup is straightforward. The authors are honest that the energy advantage is not measured and that the domain-transfer assumption is informal. For an engineering preprint, that's fair.\n\nThe soft spot is the evaluation. The kernels are learned on the entire 155,881-image set G, which is built from all Neovision2 clips, and then 5-fold cross-validation is run on the same images. So each test fold has already contributed to the fixed feature extractor. The CNN baselines are trained only on the training folds. That makes the 91.58% vs 92.34% comparison not apples-to-apples, and the 'comparable with traditional CNNs' claim is likely optimistic. This isn't a footnote issue; it changes the headline result. The fix is easy: re-learn kernels per training fold and report the mean and spread, or at least show that kernels learned on one clip set transfer to a disjoint clip set.\n\nOther concerns are minor. Hyperparameters and the calibration coefficient are chosen heuristically on the same data; some ablation would help. No code is released, though a supplementary link exists. The energy claim is explicitly deferred, which is fine, but it means the main motivation remains unverified.\n\nWho's it for? People building SNN pipelines for neuromorphic processors, and researchers working on locality-friendly learning rules. It deserves a serious referee, but I would not accept the accuracy numbers as they stand. I'd ask for the split-fair evaluation and ideally an ablation on kernel transferability before publication.","headline":"Fixed convolutional kernels learned offline are a sensible idea, but the accuracy claim is inflated by learning kernels on the full dataset before the split.","tokens_in":9533,"tokens_out":2019,"would_cite":false,"duration_ms":18790,"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":"Fixed convolutional kernels give spiking net near-CNN accuracy","keywords":["convolutional spiking neural network","CoLaNET","synaptic plasticity","unsupervised feature learning","image classification","Neovision2","neuromorphic hardware","weight sharing"],"falsifier":"Learn the 28 kernels on one subset of the Neovision2 clips, classify objects from a disjoint subset, and compare with kernels learned on the target subset; a large drop below the reported 91.58% would refute the transfer assumption, as would a marked accuracy gain if the same network is retrained with trainable convolution weights.","tokens_in":8606,"feed_emoji":"🧠","tokens_out":6460,"duration_ms":62104,"temperature":0.7,"pith_summary":"This paper tries to establish that a spiking neural network can classify images nearly as well as a conventional convolutional network while leaving the convolutional weights completely fixed. Its strategy is to learn a small set of domain-level convolution kernels offline from unlabeled images, freeze them, and install them as non-plastic synaptic weights in a spiking convolutional front end. On the five-class Neovision2 benchmark the resulting network reaches 91.58% mean accuracy, against 92.34% for a similarly shaped trained CNN and 94.35% for a deeper CNN. If this holds, convolutional feature extraction becomes compatible with neuromorphic hardware constraints, because only the downstream classifier has to learn.","feed_headline":"Fixed convolutional kernels give spiking net near-CNN accuracy","feed_subtitle":"Frozen conv kernels learned from unlabeled images let a spiking classifier almost match a trained CNN on Neovision2.","key_machinery":"The load-bearing mechanism is an offline, unsupervised kernel-learning algorithm applied to a collection G of unlabeled images from the target domain. It maintains a synaptic-resource tensor W for each of NC candidate K by K kernels; on each image it identifies the most strongly stimulated kernel positions, strengthens the resources of synapses connected to the bright pixels that caused firing, and subtracts an equal total from the other synapses so each neuron's resources are conserved. A winner-takes-all constraint prevents two kernels from learning the same image patch, forcing the matrices to become distinct detectors of frequent local structures such as edges, spots, and boundaries. After learning, the resource values are converted to weights by a bounded monotone mapping, scaled by a calibration constant, and fixed as the weights of rate-coded integrate-and-fire neurons in a convolutional layer. The classifier is a columnar layered spiking network (CoLaNET) that receives the flattened, mean-pooled feature maps and does all of the plastic learning.","core_discovery":"The paper's central claim is that in-network learning of convolution kernels is unnecessary for a competitive spiking classifier. An unsupervised procedure, run on a representative set of unlabeled domain images, produces kernels that are ported as frozen weights into an SNN, and the full pipeline—frozen convolutional layer, average pooling, and a plastic columnar layered classifier—achieves 91.58% mean accuracy (SD 1.1) on Neovision2. This is within 0.76 percentage points of CNN2, the same-shape CNN trained end-to-end, and 2.77 points below the larger CNN1. The authors attribute the residual gap mainly to spike-rate encoding, which represents pixel brightness by only 11 spike counts rather than 256 gray levels, and point to the expected energy savings of neuromorphic execution as the compensating advantage.","pith_inferences":["If the domain-transfer assumption holds, the same 28 kernels should work for other tasks built from the same camera stream; a direct test would be to learn kernels on one subset of Neovision2 clips and classify objects from another subset.","The kernel-learning algorithm is a batch analogue of STDP-based unsupervised feature learning, so applying it recursively to pooled feature maps could plausibly build deeper convolutional SNNs without violating locality.","The reported comparison uses one epoch and a small five-class task; whether the accuracy gap widens on larger, harder benchmarks such as full ImageNet subsets is unresolved by this paper.","Increasing the spike-count encoding from 11 levels to finer rate or temporal codes should narrow the gap to CNN1; this is testable by lengthening the image presentation window."],"forward_implications":["A spiking convolutional network can be built without weight sharing or in-network kernel training, removing a major obstacle to mapping convolutional feature extraction onto neuromorphic processors.","The same fixed kernel set can in principle be reused across classification tasks within one domain, with only the classifier retrained.","On Neovision2 the fixed-kernel SNN is competitive with a same-shape CNN, and it uses far fewer neurons (9,412 vs. 42,885 for the larger CNN), so the trade-off is a few points of accuracy against large hardware savings.","Because only the classifier layer is plastic, the network is suited to one-pass online learning without storing the full training set."],"supporting_citations":[{"why":"Supplies the unsupervised winner-take-all spiking feature-learning approach that this paper adapts into a standalone offline kernel algorithm.","marker":"[10]"},{"why":"Supplies CoLaNET, the columnar layered spiking classifier that performs all plastic learning on the flattened feature maps.","marker":"[20]"},{"why":"Supplies the Neovision2/Tower dataset used to build the unlabeled set G and to run the five-class classification benchmark.","marker":"[21]"},{"why":"Is cited as sharing the assumption that domain-level convolutions can transfer across tasks within a domain.","marker":"[16]"},{"why":"Provides the synaptic-resource interpretation of the auxiliary tensor W and the constant-resource competition rule used in kernel learning.","marker":"[17–19]"}],"fun_headline_variants":["Unsupervised conv kernels: spiking net near CNN accuracy","Frozen kernels from unlabeled images: spiking net close to CNN","Spiking CNN with fixed kernels: almost CNN accuracy","Unsupervised kernels let spiking net near CNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Convolution kernels learned from a representative unlabeled image set in one domain transfer to arbitrary classification tasks in that domain; the paper itself says this cannot be formally proven and is accepted from common sense.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised conv kernels: spiking net near CNN accuracy","Frozen kernels from unlabeled images: spiking net close to CNN","Spiking CNN with fixed kernels: almost CNN accuracy","Unsupervised kernels let spiking net near CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000984,"raw_usage":{"total_tokens":4099,"prompt_tokens":796,"completion_tokens":3303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":412,"completion_tokens_details":{"reasoning_tokens":3234}},"tokens_in":412,"tokens_out":3303,"duration_ms":23548,"temperature":1.0,"reasoning_tokens":3234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:53:12.923069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Learn the 28 kernels on one subset of the Neovision2 clips, classify objects from a disjoint subset, and compare with kernels learned on the target subset; a large drop below the reported 91.58% would refute the transfer assumption, as would a marked accuracy gain if the same network is retrained with trainable convolution weights.","supporting_citations":[{"cited_title":"STDP-based spiking deep convolutional neural networks for object recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the unsupervised winner-take-all spiking feature-learning approach that this paper adapts into a standalone offline kernel algorithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Neovision2/Tower dataset used to build the unlabeled set G and to run the five-class classification benchmark."},{"cited_title":"Exploring parameter and hyper-parameter spaces of neuroscience models on high performance computers with learning to learn ,","cited_arxiv_id":null,"evidence_quote":"Is cited as sharing the assumption that domain-level convolutions can transfer across tasks within a domain."}],"review_version":1}