{"id":"c103fd0e-39ac-4006-be05-bc8309f17a87","arxiv_id":"2608.01090","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frozen-backbone spiking pseudo-ensemble trained to maximize head disagreement on blur-corrupted inputs improves OOD detection on remote sensing shifts at lower cost than a deep ensemble.","lead":"The paper attaches several small heads to one frozen spiking neural network, and trains them to agree on clean images but disagree on blurry ones. This gives cheaper ensemble-style uncertainty, improving out-of-distribution detection on remote sensing data with fewer parameters and computations than a full ensemble.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADPE's headline UCM results may be inflated by selecting hyperparameters on UCM itself; the main comparison needs a held-out validation split.","rationale":"The reader's verdict is CONDITIONAL with moderate confidence, and my read does not move it. I agree that the paper's cleanest comparison is CEPE vs ADPE, which isolates the disagreement objective and shows consistent AUROC/FPR improvements across two backbones and four OOD datasets. That evidence supports the claim that blur-induced disagreement transfers to real shifts. The weaker spot is experimental protocol: the supplementary ablation selects Kh and lambda_dis on UCM, and UCM is a headline test set. This is a concrete, manuscript-evidenced concern rather than a speculative assumption. The 'broad plateau' defense is itself computed on UCM, so it does not remove the selection bias. However, the effect is likely moderate: the chosen setting is near the center of a flat region, AID and Global-LULC results were not used for selection and still show gains, and the spiking/RS setting makes exact overfitting less obvious. A held-out validation check would settle it. I therefore recommend no change to the CONDITIONAL verdict: the paper should be accepted only if the authors either show the ablation was on a validation split or confirm the UCM result under a pre-registered, UCM-free hyperparameter choice. The absence of released code remains a reproducibility limitation but is not the load-bearing issue.","tokens_in":19014,"tokens_out":8040,"duration_ms":72114,"concrete_test":"Hold out UCM completely. Before any final evaluation, repeat the Table 8/9 ablation on a validation split that excludes UCM and AID, for example a held-out EuroSAT validation set corrupted with the same box-blur kernels, and record which (Kh, lambda_dis) it selects. Then recompute Table 2's ADPE(3,5) UCM and AID rows using the locked configuration. If ADPE(3,5) still matches or exceeds DE(5,1) on UCM and AID after this pre-registered choice, the concern is resolved; if the UCM advantage shrinks or reverses, the headline overstates the method's OOD gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is not the blur-to-shift transfer assumption, which the head-to-head CEPE vs ADPE comparisons across UCM, AID, Global-Near, and Global-Far empirically support. It is the use of UCM for model selection. In the Hyperparameter Ablation (supplementary, Tables 8 and 9), Kh in {2,...,7} and lambda_dis in {0.1,...,0.5} are swept using UCM FPR@95 on ResNet19-SNN backbones with seeds {0,1,2}. The main experiments, including Table 2's ADPE(3,5) UCM rows for both backbones, then use the selected Kh=5 and lambda_dis=0.3. UCM is one of the two OOD sets in the headline claim ('matches or improves on a five-model deep ensemble on UCM and AID'). Because the same UCM metric that is later reported was used to pick the configuration, the UCM leg of the central claim is optimistically biased and is not a clean out-of-sample evaluation. The claimed 'broad region of competitive performance' (within ~1.2 FPR@95 points) mitigates the selection effect, and AID/Global-LULC were not used for tuning, so the cross-dataset improvements remain informative. But the specific UCM-vs-deep-ensemble comparison cannot be taken at face value without an independent validation split.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Agree–Disagree Spiking Pseudo-Ensembles (ADPE) for out-of-distribution detection in remote sensing. A frozen SNN backbone is equipped with multiple lightweight spiking heads; the heads are trained to agree on clean in-distribution images and to disagree on box-blurred versions of the same images, via a Jensen–Shannon divergence term. The authors evaluate Spikformer and ResNet19-SNN backbones on EuroSAT as in-distribution data and UCM, AID, and two Sentinel-2 Global LULC subsets as OOD sets, comparing against deep ensembles, MC dropout, last-layer Laplace, Mahalanobis distance, and cross-entropy-trained pseudo-ensembles (CEPE). The headline result is that ADPE(3,5) (three backbones, five heads each) matches or improves a five-model deep ensemble on UCM and AID while using about 38% fewer parameters and 40% fewer backbone evaluations. The paper also reports consistent AUROC improvements over CEPE across both backbones and all four OOD datasets.","tokens_in":19322,"tokens_out":5939,"duration_ms":53295,"significance":"If the results hold, the method provides a practically useful efficiency–uncertainty trade-off for SNN deployment and introduces a clean, OOD-data-free diversity-promoting objective. The paper is unusually thorough on the experimental side: two architectures, four OOD datasets, multiple uncertainty metrics, and detailed ablations. The authors are also explicit about the key limitation—that blurred inputs are not assumed to match the true test-time OOD distribution—and they provide honest discussion of the mixed FPR@95 results. However, the central empirical claims rest on a few experimental-design choices that currently weaken the internal validity of the headline comparisons; these are fixable and do not invalidate the overall idea.","major_comments":[{"comment":"The hyperparameters Kh=5 and λ_dis=0.3 are selected by sweeping Kh and λ_dis on UCM FPR@95 (Tables 8 and 9, using ResNet19-SNN backbones with seeds 0,1,2). The main experiments, including the UCM rows of Table 2, then use this configuration for both backbones. Because UCM is one of the two datasets in the headline claim 'matches or improves a five-model deep ensemble on UCM and AID,' the UCM leg of that claim is optimistically biased and is not a clean out-of-sample evaluation. The reported flatness of the ablation surface mitigates the effect, but it does not remove the selection bias. I request that the authors either (a) perform model selection on a held-out validation split (e.g., a subset of EuroSAT with synthetic corruptions, or a separate OOD dataset such as one of the Global-LULC subsets) and then report the chosen configuration on UCM, or (b) explicitly quantify the selection bias by reporting how the selected configuration performs when the selection is repeated on different OOD targets.","section":"Supplementary, Tables 8 and 9; Table 2"},{"comment":"The main text states that the only difference between CEPE and ADPE is the disagreement term on blurred inputs, but the supplementary reports that CEPE heads receive minibatches from their own independently shuffled data loaders, whereas ADPE uses a shared minibatch. This is a real confound: the two methods differ not only in the objective but also in the per-head data stochasticity. The shared-minibatch requirement is inherent to the JSD objective, but the authors should either implement a CEPE variant with a shared minibatch as an additional control, or explicitly discuss how the different data loading schemes could affect head diversity. Without this, the attribution of the observed improvements to the disagreement objective is not fully supported.","section":"Supplementary, 'Backbone and Pseudo-Ensemble Training'; Section 'Effect of Agree–Disagree Training'"},{"comment":"The deep-ensemble baseline DE(5,1) is reported with a standard deviation of zero because only one subset of five backbones exists. Meanwhile, ADPE(3,5) reports means and standard deviations over ten backbone subsets. The claimed improvements over the deep ensemble on ResNet19-SNN are small (e.g., UCM AUROC 92.10 vs. 91.84; AID AUROC 93.12 vs. 92.79), and without a measure of the deep ensemble's variance across independent training seeds or a statistical test, it is unclear whether these differences are meaningful. The authors should provide either confidence intervals for the deep-ensemble baseline (e.g., by retraining multiple sets of five backbones with different seed sets) or a statistical comparison that accounts for the fact that only one ensemble was evaluated.","section":"Table 2; Section 'Overall OOD Detection Performance'"}],"minor_comments":[{"comment":"The caption says 'MSP-based OOD detection,' but the Mahalanobis row reports results from its native feature-distance score rather than MSP. Please clarify this in the caption so the row is not misinterpreted.","section":"Table 2 caption"},{"comment":"The main text uses Blur-S, Blur-M, and Blur-L in Table 1 without specifying the kernel sizes; the supplementary explains these are 5, 9, and 11. Please define the kernel sizes in the main text when the abbreviations are first introduced.","section":"Table 1 and Eq. (10)"},{"comment":"The value λ_dis=0.3 is presented as a fixed design choice, but the supplementary shows it is the result of an ablation. Please state explicitly in the main text that this value was selected via the ablation and, if possible, cite the ablation tables there.","section":"Section 'Agree–Disagree Training', Eq. (12)"},{"comment":"The parameter counts P_b≈12.5M and P_h≈0.1M are presented without a derivation; a brief note on how these numbers were obtained (e.g., from the architecture configurations in the supplementary) would aid reproducibility.","section":"Section 'Inference and Computational Efficiency'"},{"comment":"The supplementary describes CEPE as using separate optimizers and separate shuffling, but does not specify the random seed that controls the sharding; please clarify the seed handling for these per-head loaders.","section":"Supplementary, 'Backbone and Pseudo-Ensemble Training'"},{"comment":"The paper does not provide a code-availability statement. Given the Reproducibility section's detail, a public code link or a statement of intention to release code would strengthen the manuscript.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The model-selection-on-UCM issue is the most serious concern, but it is fixable with a held-out validation split or a sensitivity analysis. The CEPE data-loading confound is also fixable with an additional baseline. The underlying method is interesting and the cross-dataset results (AID, Global-LULC) suggest that the transfer from blur-induced disagreement is real beyond UCM; I believe the paper can be made acceptable after these revisions. The authors should also be encouraged to report the deep-ensemble baseline with some measure of variability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you spend time on this. First, the paper is a clean efficiency result: frozen SNN backbone with several lightweight heads trained under an agree–disagree objective (cross-entropy on clean images, JSD maximization on box-blurred ones). That combination is new in the spiking setting, and the CEPE vs ADPE comparison cleanly isolates the disagreement term. Second, the headline UCM numbers are partly tuned. The supplementary ablation (Tables 8/9) sweeps Kh and λ_dis using UCM FPR@95 on ResNet19-SNN seeds {0,1,2}, and the main experiments then use those values for both backbones. UCM is one of the two datasets in the central claim, so that leg is not a clean out-of-sample evaluation.\n\nWhat is genuinely good: the evaluation spans two SNN architectures, four OOD sets (UCM, AID, Global-Near, Global-Far), and several baselines. ADPE(3,5) matches or beats a five-model deep ensemble on AID and both Global-LULC shifts, which were not used for tuning. The parameter and computation accounting is transparent: about 38% fewer parameters and 40% fewer backbone evaluations. The blur-selection diagnostic is honest; they use deep-ensemble uncertainty on ID corruptions, not OOD data, and they explicitly disclaim that blurred inputs model test-time shifts.\n\nThe soft spots are real but moderate. The UCM selection issue is the main one; the ablation shows a broad plateau (selected config within about 1.2 FPR@95 points of the best), so the bias is limited, but a validation split or an explicit acknowledgment of tuning would make the UCM claim trustworthy. Minor: the DE(5,1) comparison has zero variance because only one subset exists; no code is released, so reproducing the backbone training is nontrivial; and the MI-based results are partly aligned with the training objective, though the MSP and entropy improvements over CEPE are stronger evidence and hold consistently.\n\nFor whom: people working on efficient uncertainty for SNNs or resource-limited remote sensing. It is a solid subfield contribution, not a new framework.\n\nRecommendation: send it to peer review. Ask the authors to show the hyperparameter ablation on a held-out validation split or demonstrate robustness to separate validation. The core method and the cross-dataset evidence are worth publishing; the UCM leg just needs to be cleaned up.","headline":"Solid SNN pseudo-ensemble efficiency result, but the headline UCM comparison is partly tuned and needs a held-out validation split.","tokens_in":19850,"tokens_out":5132,"would_cite":true,"duration_ms":41585,"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":"A frozen spiking backbone fitted with five lightweight heads, trained to agree on clean images and disagree on box-blurred ones, matches a five-model deep ensemble for remote-sensing out-of-distribution detection while using roughly 38%…","keywords":["spiking neural networks","out-of-distribution detection","pseudo-ensemble","deep ensemble","diversity collapse","agree-disagree objective","box blur","remote sensing"],"falsifier":"Re-run ADPE training with the disagreement term evaluated on additive Gaussian noise instead of box blur, holding all other hyperparameters fixed; if the reported UCM and AID AUROC advantages over CEPE disappear or reverse, then the structure-preserving property of box blur carries the claimed diversity transfer.","tokens_in":18800,"feed_emoji":"🛰️","tokens_out":17463,"duration_ms":133051,"temperature":0.7,"pith_summary":"Spiking neural networks promise low-cost inference for satellites and edge devices, but they lack reliable out-of-distribution (OOD) detection. The paper proposes a cheap substitute for deep ensembles: freeze one pretrained SNN backbone, attach several lightweight classification heads, and train those heads so they agree on clean images but deliberately disagree on box-blurred versions of the same images. This agree-disagree training prevents diversity collapse, the tendency of heads trained independently with cross-entropy to converge to correlated predictions. With three backbones and five heads each on EuroSAT as in-distribution data, the resulting ADPE(3,5) matches or improves the OOD detection of a five-model deep ensemble on UCM and AID while using about 38% fewer parameters and 40% fewer backbone evaluations. If correct, this would give resource-constrained remote-sensing systems ensemble-quality uncertainty at a fraction of the deployment cost.","feed_headline":"Frozen SNN plus five lightweight heads matches a 5-model ensemble","feed_subtitle":"Agree-disagree training on blurred images achieves it with 38% fewer parameters for OOD detection.","key_machinery":"The load-bearing object is the agree-disagree spiking pseudo-ensemble: a frozen SNN backbone (Spikformer or ResNet19-SNN) with $K_h$ lightweight spiking heads broadcast to the same temporal features, trained under $L_{\\mathrm{AD}}=L_{\\mathrm{CE}}-\\lambda_{\\mathrm{dis}}D_{\\mathrm{JS}}$ with $\\lambda_{\\mathrm{dis}}=0.3$. The disagreement signal is the generalized Jensen-Shannon divergence $$D_{\\mathrm{JS}} = \\frac{1}{K_h|\\mathcal{I}_b|} \\sum_{i\\in\\mathcal{I}_b} \\sum_{k=1}^{K_h} \\mathrm{KL}(p_k(\\tilde{x}_i) \\| \\bar{p}(\\tilde{x}_i)),$$ computed only on box-blurred inputs $\\tilde{x}_i$, with kernel size sampled uniformly from $\\{5,7,9,11\\}$ and $30\\%$ of each minibatch blurred. Clean images contribute only cross-entropy, so every head keeps solving the original classification task; blurred images carry no labels and only push the heads toward different predictive distributions. A diagnostic with a five-model Spikformer ensemble selects box blur because medium and large blur kernels produce the strongest rise in ensemble entropy and mutual information while preserving coarse scene structure. Because the backbone is frozen and evaluated once per input, the added cost of more heads is small: five heads add roughly $4\\%$ parameter overhead per backbone.","core_discovery":"The paper's central claim is that shared-backbone pseudo-ensemble failure is not inevitable. A frozen spiking backbone's time-resolved features can be broadcast to several independently initialized spiking heads, but when those heads are trained with cross-entropy alone they suffer diversity collapse: every head sees the same representation, labels, and objective, so their predictions become correlated. The paper replaces that training with an agree-disagree objective, $$L_{\\mathrm{AD}} = L_{\\mathrm{CE}} - \\lambda_{\\mathrm{dis}} D_{\\mathrm{JS}}, \\quad \\lambda_{\\mathrm{dis}} = 0.3,$$ where $L_{\\mathrm{CE}}$ applies cross-entropy to clean in-distribution images and $D_{\\mathrm{JS}}$ is a generalized Jensen-Shannon divergence among head predictions on box-blurred copies of the same images. Because blurred inputs are generated from the training set, the diversity signal requires no external outlier data or teacher ensemble. The paper reports that ADPE(3,5) matches or exceeds a five-model deep ensemble on UCM and AID with roughly 38% fewer parameters and 40% fewer backbone evaluations, and that it consistently improves MSP (maximum-softmax-probability)- and mutual-information-based OOD detection over cross-entropy pseudo-ensembles for both Spikformer and ResNet19-SNN.","pith_inferences":["The paper does not test shifts that alter only high-level semantics while preserving texture statistics, such as rearranged land-cover configurations; on such shifts the blur-induced disagreement may be uninformative.","Because the loss needs only predictions and blurred inputs, the same agree-disagree recipe could be attached to non-spiking backbones or to larger head banks, but that extension is untested.","The reported FPR@95 is less monotonic than AUROC, so converting the AUROC gains into a reliable fixed operating threshold may require calibration or adaptive disagreement inputs.","The blur choice was made once from a diagnostic study; automatically selecting disagreement transformations per dataset could yield stronger or more robust diversity."],"forward_implications":["A single frozen backbone with five heads already improves MSP-based OOD AUROC over one deterministic model (for Spikformer, from 89.53% to 94.82% on UCM), so within-backbone diversity is itself a usable source of uncertainty.","ADPE(3,5) needs three backbone evaluations and about 3.12 full-model parameter equivalents instead of five, so ensemble-like OOD detection is available on memory- and energy-limited hardware.","The disagreement objective improves over cross-entropy-trained pseudo-ensembles for both a Transformer-based and a convolutional SNN backbone, indicating the mechanism is not tied to one architecture.","Detection remains harder for semantically close shifts (Global-Near) than for far shifts (Global-Far), so the method improves but does not remove the difficulty of near-OOD acceptance."],"supporting_citations":[{"why":"Defines the deep ensemble whose five-member configuration ADPE(3,5) is compared against and whose linear cost motivates the pseudo-ensemble.","marker":"Lakshminarayanan, Pritzel, and Blundell 2017"},{"why":"Documents the diversity-collapse failure mode in shared-input multihead architectures and motivates the explicit diversity term.","marker":"Havasi et al. 2020"},{"why":"Supplies the agree-to-disagree training principle that the paper adapts from auxiliary OOD data to box-blurred in-distribution images.","marker":"Pagliardini et al. 2023"},{"why":"Provides the Spikformer backbone architecture used for the Transformer-based experiments.","marker":"Zhou et al. 2022"},{"why":"Provides the ResNet19-SNN backbone architecture used for the convolutional experiments.","marker":"Deng et al. 2022"},{"why":"Supplies the EuroSAT in-distribution training and test images.","marker":"Helber et al. 2019"},{"why":"Supplies the UCM OOD benchmark used in the main comparison.","marker":"Yang and Newsam 2010"},{"why":"Supplies the AID OOD benchmark used in the main comparison.","marker":"Xia et al. 2017"},{"why":"Supplies the Sentinel-2 Global LULC data from which the near and far OOD subsets are built.","marker":"Benhammou et al. 2022"}],"fun_headline_variants":["Agree-disagree training stops diversity collapse in spiking heads","Frozen SNN heads trained with agree-disagree match deep ensembles","Pseudo-ensemble diversity collapse solved by agree-disagree loss","Lightweight spiking heads match 5-model ensemble with 38% fewer params","Diversity promotion on blurred inputs recovers ensemble-style uncertainty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, acknowledged in the paper when selecting disagreement inputs, is that disagreement induced on box-blurred in-distribution images transfers to real distribution shifts; the authors state that blurred samples 'are not assumed to follow the true test-time OOD distribution.'","fun_headline_variants_meta":{"raw":{"variants":["Agree-disagree training stops diversity collapse in spiking heads","Frozen SNN heads trained with agree-disagree match deep ensembles","Pseudo-ensemble diversity collapse solved by agree-disagree loss","Lightweight spiking heads match 5-model ensemble with 38% fewer params","Diversity promotion on blurred inputs recovers ensemble-style uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000971,"raw_usage":{"total_tokens":4171,"prompt_tokens":1032,"completion_tokens":3139,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":3063}},"tokens_in":648,"tokens_out":3139,"duration_ms":20446,"temperature":1.0,"reasoning_tokens":3063,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:12:19.222496+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run ADPE training with the disagreement term evaluated on additive Gaussian noise instead of box blur, holding all other hyperparameters fixed; if the reported UCM and AID AUROC advantages over CEPE disappear or reverse, then the structure-preserving property of box blur carries the claimed diversity transfer.","supporting_citations":[{"cited_title":"Proceedings of the 18th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems , pages =","cited_arxiv_id":null,"evidence_quote":"Supplies the UCM OOD benchmark used in the main comparison."},{"cited_title":"Scientific Data , volume =","cited_arxiv_id":null,"evidence_quote":"Supplies the Sentinel-2 Global LULC data from which the near and far OOD subsets are built."}],"review_version":2}