{"id":"7fad7262-2023-4026-b746-3d11f8a41de3","arxiv_id":"2412.12696","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ALADE-SNN combines dynamic network expansion with adaptive logit alignment and OtoN weight suppression to improve class incremental learning in spiking neural networks.","lead":"This paper presents ALADE-SNN, a spiking neural network method for class incremental learning that expands the network with each new task and adds an adaptive correction to rebalance old and new class logits. It reports 75.42% average incremental accuracy on CIFAR100-B0 over 10 steps, surpassing one SNN baseline and roughly matching DNN methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'surpasses SOTA SNN' claim rests on one cross-paper baseline, DSD-SNN, whose rehearsal protocol is not shown to match ALADE's; the +14.95 gap is uncontrolled.","rationale":"The reader identified the cross-paper baseline problem for the DNN comparison. I agree with that concern and sharpen it: the more absolute claim in the abstract, 'surpasses state-of-the-art SNN-based continual learning algorithms,' is supported by only one SNN baseline, DSD-SNN, and the paper does not demonstrate that DSD-SNN uses the same memory/rehearsal protocol. This is a load-bearing weakness because the headline number gap (+14.95) could be an artifact of comparing rehearsal-based ALADE against a non-rehearsal or differently configured baseline. The reader's CONDITIONAL verdict already requires strengthened baselines, so my read does not change the verdict; it reinforces the condition. I did not find an internal inconsistency in the adaptive logit alignment or OtoN suppression mechanisms sufficient to reject the method outright, and the ablation study provides partial support for the individual components.","tokens_in":15169,"tokens_out":10278,"duration_ms":94393,"concrete_test":"Obtain DSD-SNN's original code and configuration and run it under ALADE's exact CIFAR100-B0-10steps protocol (memory=2000, herding exemplars, Spiking ResNet-19, T=4); if DSD-SNN's published setting is rehearsal-free or uses a different memory, run ALADE under that setting instead. Report both Avg and Last. If the gap narrows to within seed noise or DSD-SNN wins, the SOTA claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's claim to 'surpass state-of-the-art SNN-based continual learning algorithms' is supported in Table 3 by exactly one SNN baseline, DSD-SNN (60.47% Avg on CIFAR100-B0-10steps vs ALADE's 75.42%, and 57.39% vs 72.73% on 20 steps). The Implementation Details state that 'The results are directly taken from their respective papers,' but no evidence is given that DSD-SNN was evaluated under the same protocol: ALADE uses a fixed memory of 2000 exemplars with herding (B0 protocol) and a Spiking ResNet-19 with T=4, while DSD-SNN's memory budget, backbone, timestep, and rehearsal strategy are not reported in the comparison table. If DSD-SNN is rehearsal-free or uses a smaller memory, the large gap is expected and not evidence of superiority. Moreover, other SNN CIL methods mentioned in the paper (SOR-SNN, SSA-SNN, ESL-SNN) are not included, so the 'surpasses state-of-the-art' claim is not established. The same cross-paper issue affects the ANN comparison, but the paper hedges that as 'comparable'; the SNN-SOTA claim is absolute and therefore more load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ALADE-SNN, a class-incremental learning (CIL) method for spiking neural networks built on the DER dynamic-expansion framework. It introduces three components: adaptive logit alignment (a correction term gamma applied to new-class logits during classifier fine-tuning), OtoN suppression (zeroing weights from frozen old features to new classes while training the new feature extractor), and knowledge distillation. The main experiments are on CIFAR100-B0 and CIFAR100-B50 benchmarks, reporting an average incremental accuracy of 75.42±0.74% on CIFAR100-B0-10steps. The paper claims that ALADE-SNN matches DNN-based methods and surpasses state-of-the-art SNN continual learning algorithms.","tokens_in":15448,"tokens_out":9167,"duration_ms":78819,"significance":"If the results are properly supported, ALADE-SNN would be a meaningful advance in SNN continual learning: it proposes a simple, plug-in logit-alignment mechanism with an 'adaptive' update rule, validates on standard CIFAR100 protocols with three seeds and standard deviations, and is implemented in an open-source framework. The observation that imbalanced replay data biases logit distributions is a useful empirical finding. However, the strength of the central claims is currently limited by uncontrolled cross-paper comparisons and an internally inconsistent ablation result, so the significance of the work is not fully established in this version.","major_comments":[{"comment":"The abstract's claim that ALADE-SNN 'surpasses state-of-the-art SNN-based continual learning algorithms' is not established by Table 3, which compares against a single SNN baseline, DSD-SNN, whose results are 'directly taken from their respective papers' (Implementation Details). No evidence is provided that DSD-SNN was evaluated under the same protocol: ALADE-SNN uses a fixed memory of 2000 exemplars with herding and a Spiking ResNet-19 with T=4, while DSD-SNN's memory budget, backbone, timestep, and rehearsal strategy are not reported in the table. The +14.95-point gap on CIFAR100-B0-10steps may therefore reflect protocol differences rather than algorithmic superiority. Moreover, other SNN CIL methods mentioned in the paper (SOR-SNN, SSA-SNN, ESL-SNN) are absent from the comparison, so the 'surpasses SOTA' claim is not supported by the evidence presented.","section":"Results / Implementation Details (Table 3)"},{"comment":"The claim that ALADE-SNN 'matches the performance of DNN-based methods' rests on an uncontrolled cross-paper comparison. Table 3 and Table 4 compare ALADE-SNN (Spiking ResNet-19, T=4, herding memory) with ANN results from other papers (ResNet18/ViT, different training protocols), and the Implementation Details state that these numbers are taken verbatim. Differences in backbone, timestep, memory selection, training epochs, and hyperparameters can easily account for several accuracy points. To support the 'comparable to DNN' claim, the authors should run matched-setting ANN baselines (e.g., DER with ResNet18 under the same memory and data-order protocol) in their own framework, or at least explicitly restrict the claim to a qualitative comparison.","section":"Results, Tables 3-4 and Implementation Details"},{"comment":"The ablation results do not support the statement that knowledge distillation 'brings a slight gain.' In Table 5, removing knowledge distillation from ALADE-SNN increases average incremental accuracy from 75.42%±0.74 to 75.51%±1.31 and final accuracy from 63.13%±0.40 to 63.60%±0.75; adding knowledge distillation to the DER-SNN baseline gives 74.47%±1.85 versus 74.22%±1.06, a difference well within one standard deviation. The text's claim of a gain is therefore inconsistent with the reported numbers, and the authors should either provide a statistical significance test or revise the claim. This is load-bearing because the paper lists knowledge distillation as one of its three components.","section":"Results, Table 5 (Ablation experiment)"},{"comment":"The adaptive update of the correction term gamma in Algorithm 1 appears to have a sign/stability issue that should be clarified. In Table 1, the authors observe that new-task ground-truth logits are higher than old-task logits in DER (e.g., 13.32 vs 10.45 at step 3). With Eq. (1) adding gamma only to new-class logits, the measured difference Delta_e = s_new/n_new - s_old/n_old is inflated by gamma. The update gamma_e <- gamma_{e-1} + Delta_e then increases gamma whenever new logits are higher, which can cause gamma to grow without bound unless the fine-tuning dramatically reduces the true new-class logits. The paper does not report gamma's initialization, a clipping mechanism, or observed gamma trajectories. The authors should demonstrate that the update converges (or is bounded) and provide empirical gamma values, or the central logit-alignment mechanism's reproducibility is in doubt.","section":"Dynamically Expandable SNNs with Adaptive Logits Alignment, Algorithm 1 and Eq. (1)"}],"minor_comments":[{"comment":"There is a typo in 'biogical comuting processes'; it should read 'biological computing processes'.","section":"Introduction"},{"comment":"The column header '3st avg/std acc' should be '3rd avg/std acc'.","section":"Table 1"},{"comment":"The initial value of the correction term gamma is not specified; please state the initialization used in all experiments (e.g., gamma_0 = 0).","section":"Eq. (1) and Algorithm 1"},{"comment":"The variables 'Delta_st' and 'delta_Delta' are used before being explicitly defined; please clarify their roles or rename them to avoid confusion.","section":"Algorithm 1"},{"comment":"The sentence containing '1.7625µJ vs 3.367pJ' is confusingly written and the unit comparison is unclear; please rewrite or remove it.","section":"Results, energy estimation"},{"comment":"The caption should explicitly state that '-' denotes removing a component from ALADE-SNN and '+' denotes adding it to DER-SNN*.","section":"Table 5 caption"},{"comment":"The acronym 'OtoN' is not defined at first use; please define it (e.g., 'old-to-new suppression').","section":"OtoN Suppression"}],"recommendation":"major_revision","confidential_remarks":"The core concern is that the two headline claims—'surpasses SNN SOTA' and 'matches DNN methods'—depend on cross-paper numbers that are not protocol-matched. The ablation inconsistency for knowledge distillation is also a direct contradiction with the text. If the authors can provide matched-setting baselines, clarify or bound the gamma update, and fix the ablation discussion, the paper could become publishable. I would not reject outright because the method itself is plausible and the experiments include seeds and standard deviations, but the current evidence is not sufficient for the claimed strengths."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of ALADE-SNN. The core idea is a legitimate small step: take the DER dynamic expansion framework, adapt it to SNNs, and tackle the logit imbalance with an adaptive correction term that is updated online during the classifier fine-tuning, plus an OtoN suppression mask that blocks frozen old features from being mapped to new classes during feature training and releases them during fine-tuning. The analysis in Section 3, where they compare logit statistics of an oracle vs DER variants, is a good way to motivate the design. Credit where due: the method is clearly described, the ablation shows the alignment term gives roughly a point, and the three-seed standard deviations are included.\n\nThe soft spots are about evaluation, not the mechanism. The abstract's claim to surpass state-of-the-art SNN continual learning methods is built on exactly one baseline, DSD-SNN, and there's no evidence the protocols match. DSD-SNN may not use a 2000-exemplar rehearsal buffer; if it doesn't, the +14.95 gap is an artifact of the replay budget, not a demonstration of superiority. The paper cites SOR-SNN, SSA-SNN, and ESL-SNN but doesn't list them in the tables, so the 'surpasses' claim is not established. The ANN comparison is also cross-paper with a different backbone (Spiking ResNet-19, T=4 vs ResNet18/ViT), so 'matches DNN performance' is an overstatement. Within the ablation, the knowledge distillation component shows no clear gain (removing it gives 75.51 vs 75.42, both inside the std), so that part needs justification or removal. The alpha and beta in the adaptive gamma are hand-set, no sensitivity study.\n\nNone of this kills the central idea. The method is plausible and the alignment mechanism is a reasonable alternative to fixed scaling. But the claims need to be trimmed and the SNN baselines need to be run under matched conditions before I'd trust the headline numbers. I'd send this to a serious referee with a request to strengthen the comparison and soften the SOTA claim. I wouldn't cite it in my current work, but it's a fair contribution to the SNN CIL niche.","headline":"A plausible DER-based SNN CIL method with an adaptive logit alignment, but the 'surpasses SOTA SNN' claim hinges on one unmatched baseline; worth refereeing after the claims are reined in.","tokens_in":15945,"tokens_out":4226,"would_cite":false,"duration_ms":37396,"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 spiking neural network that aligns logits across tasks reaches 75.42% average incremental accuracy on CIFAR100 class-incremental learning.","keywords":["spiking neural networks","class-incremental learning","catastrophic forgetting","dynamic architecture","logit alignment","continual learning","neuromorphic computing","CIFAR100"],"falsifier":"Re-run each ANN baseline that the paper cites, such as DER, Dytox, and TCIL, on the same CIFAR100-B0 splits, memory budgets, and training schedules used for ALADE-SNN without changing the baseline methods; if their average incremental accuracy stays above 75.42% under matched conditions, the claimed parity with DNN-based methods would be contradicted, and if it falls below, the claim would be supported.","tokens_in":14986,"feed_emoji":"🧠","tokens_out":8233,"duration_ms":67216,"temperature":0.7,"pith_summary":"This paper tries to show that spiking neural networks can do class-incremental learning as well as conventional deep networks when the architecture grows with each task and the classifier's logits are actively rebalanced. The proposed ALADE-SNN appends a new spiking feature extractor at every incremental step, freezes older extractors, and adds a correction term to the logits during a classifier-only fine-tuning stage to close a bias it identifies between old and new classes. It also suppresses the classifier weights that map frozen old features to new classes during representation learning, releasing them during fine-tuning. The reported result is 75.42±0.74% average incremental accuracy on CIFAR100-B0 over 10 steps, which the authors present as surpassing prior SNN continual-learning methods and reaching parity with ANN-based ones. The value of the claim, if it holds, is that energy-efficient event-driven networks would be a viable option for learning from a stream of classes rather than only from fixed batches.","feed_headline":"Spiking network reaches 75.42% on incremental CIFAR100","feed_subtitle":"Adaptive logit alignment plus dynamic expansion beats prior spiking continual learners on CIFAR100-B0.","key_machinery":"The load-bearing objects are the adaptive logit-alignment correction and the Old-to-New (OtoN) weight suppression, operating inside the DER-style dynamic expansion. The correction is a scalar $\\gamma$ added only to new-class logits via $O = \\mathcal{H}_s(\\Phi_s(x)) + \\gamma \\cdot \\mathrm{mask}_s$ during the classifier fine-tuning stage; $\\gamma$ is updated from the measured difference $\\Delta_e$ between average new-class and old-class logits, with two hyper-parameters that decide when the correction has taken effect. OtoN suppression zeros the classifier weights connecting frozen old features to new classes while the new extractor is trained, then releases them during fine-tuning. Together they rebalance the gradient and the final predictions between old and new tasks without adding new network structure or a new loss term.","core_discovery":"The central claim is that catastrophic forgetting in class-incremental learning is substantially a logit-imbalance problem, and that a spiking network built on dynamic expansion plus two targeted corrections can solve it. On each new task, ALADE-SNN freezes the previous set of feature extractors, appends a new Spiking ResNet-19 extractor, and trains it with TET classification loss, an auxiliary task-prediction loss, and a knowledge-distillation loss. During this representation stage, the weights that map frozen old features to new classes are zeroed to stop old representations from misleading the new extractor; in the subsequent classifier fine-tuning stage those weights are released and an adaptive scalar $\\gamma$ is added to the new-class logits so that the average ground-truth logits of old and new tasks converge. The paper reports this yields 75.42±0.74% average incremental accuracy on CIFAR100-B0 with 10 incremental steps, 78.67% with 5 steps, and 72.73% with 20 steps, surpassing the compared SNN continual-learning baseline and sitting close to the ANN methods it cites.","pith_inferences":["The authors do not test this, but the adaptive logit-alignment correction should transfer to non-spiking dynamic-expansion methods, since it only modifies logits and adds no structure or loss.","One testable extension the paper leaves open is whether OtoN suppression generalizes to a broader rule, such as blocking all cross-task classifier mappings during representation learning and releasing them during fine-tuning.","Because the gap to ANN methods widens with more steps and the paper attributes it to the short time window, a concrete follow-up would be to vary the spike time window or backbone depth and measure whether the degradation flattens."],"forward_implications":["ALADE-SNN raises average incremental accuracy on CIFAR100-B0-10steps to 75.42%, about 15 points above the published DSD-SNN result it compares against.","On CIFAR100-B0-5steps the method scores 78.67%, about 1.8 points above the cited DER result, and on CIFAR100-B50-5steps it also surpasses the ANN baselines listed.","Ablations attribute roughly one percentage point of the gain to adaptive logit alignment, with smaller contributions from OtoN suppression and knowledge distillation.","The gap to attention-based ANN methods grows as the number of incremental steps increases, which the paper attributes to the limited time window of SNNs."],"supporting_citations":[{"why":"Supplies the DER dynamic-expansion backbone that ALADE-SNN builds on.","marker":"Yan, Xie, and He 2021"},{"why":"Supplies the TET training loss and the surrogate-gradient training recipe used for the SNN feature extractors.","marker":"Deng et al. 2021"},{"why":"Defines the Spiking ResNet-19 backbone used for every feature extractor.","marker":"Fang et al. 2021"},{"why":"Defines the CIFAR100-B0 benchmark, memory budget, and the average incremental accuracy metric.","marker":"Rebuffi et al. 2017"},{"why":"DSD-SNN is the SNN continual-learning baseline that ALADE-SNN is claimed to surpass.","marker":"Han et al. 2023b"},{"why":"Supplies the herding algorithm that selects the exemplar memory set at each step.","marker":"Welling 2009"},{"why":"Supplies the knowledge-distillation loss used in the representation-learning stage.","marker":"Hinton, Vinyals, and Dean 2015"}],"fun_headline_variants":["Logit imbalance is the real culprit in continual learning: ALADE-SNN","Spiking net solves incremental learning via adaptive logit alignment","75.42% on CIFAR100: ALADE-SNN beats spiking baselines","Dynamic spiking network adapts to new tasks without forgetting","Forgetfulness in spiking nets cured by logit alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that ALADE-SNN matches DNN-based methods assumes that published accuracy numbers from different backbones and training setups can be compared directly with the SNN results; if the ANN baselines were re-run under identical conditions, the parity result could change.","fun_headline_variants_meta":{"raw":{"variants":["Logit imbalance is the real culprit in continual learning: ALADE-SNN","Spiking net solves incremental learning via adaptive logit alignment","75.42% on CIFAR100: ALADE-SNN beats spiking baselines","Dynamic spiking network adapts to new tasks without forgetting","Forgetfulness in spiking nets cured by logit alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000716,"raw_usage":{"total_tokens":3244,"prompt_tokens":995,"completion_tokens":2249,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":2155}},"tokens_in":611,"tokens_out":2249,"duration_ms":15307,"temperature":1.0,"reasoning_tokens":2155,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:49:14.198445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run each ANN baseline that the paper cites, such as DER, Dytox, and TCIL, on the same CIFAR100-B0 splits, memory budgets, and training schedules used for ALADE-SNN without changing the baseline methods; if their average incremental accuracy stays above 75.42% under matched conditions, the claimed parity with DNN-based methods would be contradicted, and if it falls below, the claim would be supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DER dynamic-expansion backbone that ALADE-SNN builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the herding algorithm that selects the exemplar memory set at each step."}],"review_version":1}