{"id":"fd3235c7-9c09-416d-b1f6-e5db6762aae2","arxiv_id":"2412.16219","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper presents a training-free ANN-to-SNN conversion framework using layer-wise adaptive firing patterns, sensitivity-based spike compression, and entropy-based early exit to cut energy and latency while maintaining accuracy.","lead":"An ANN-to-SNN conversion method that needs no retraining, using adaptive burst-firing neurons plus spike compression and input-dependent timesteps, reports ImageNet-level accuracy at few timesteps with large theoretical energy savings. It may make spiking networks practical for energy-constrained deployment while preserving ANN accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 5's combined row contradicts the text: on CIFAR-10 the full method loses 0.87% accuracy while the text claims a 0.13% gain, undermining the headline 'together improves accuracy and energy' claim.","rationale":"The reader correctly identifies the layer-independence assumption in Eq. 11 as an unvalidated methodological pillar, and that assumption is the most plausible mechanism for the combined method degrading CIFAR-10 accuracy in Table 5. However, the more directly load-bearing issue for the paper's headline is the internal contradiction between Table 5 and the text: the 'all three techniques' row shows a 0.87% accuracy drop on CIFAR-10 even as the text claims a 0.13% gain and the abstract cites this table for a 70.1% energy saving. A false or inconsistent number in the table that supports the headline quantitative claims is a concrete, checkable failure, whereas the independence assumption alone would not necessarily falsify the empirical results. I therefore keep the reader's CONDITIONAL verdict: the paper's central claim should be accepted only after the released code reproduces the ablation and the contradiction is resolved. This is not an accusation of misconduct; the paper provides code, so the check is straightforward and fair.","tokens_in":12407,"tokens_out":12649,"duration_ms":120706,"concrete_test":"Run the released repository on CIFAR-10 with ResNet-20 and reproduce the four ablation toggles of Table 5, using the same T=8 setting; record accuracy and total spike-based energy for the baseline, AdaFire-only, and AdaFire+SSC+IAT rows. If the all-three row reproduces 95.47%, the text's '+0.13% accuracy enhancement' is false and the central claim should be revised to an energy-efficiency claim with an accuracy trade-off. If it reproduces ~96.47%, the table cell is erroneous and every energy/accuracy row in Table 5 should be regenerated before the abstract's numbers are used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract and the ablation section is that AdaFire+SSC+IAT together deliver state-of-the-art accuracy with large energy savings. In Table 5, however, the all-three row reports CIFAR-10 accuracy 95.47% against a 96.34% baseline and 96.69% for AdaFire-only, a 0.87% loss; the surrounding text states the synergistic application yields a 0.13% accuracy enhancement and a 70.12% energy reduction. The two statements cannot both be true for the same row. If the table is correct, the framework does not jointly improve accuracy on CIFAR-10; the headline must be weakened to energy savings at an accuracy cost. If the text is correct, the table cell is wrong and the other quantitative rows need re-derivation. This is not a cosmetic typo: the abstract's 'up to 70.1%' energy figure and the 'SOTA performance' claim come from this table. The likely source is the unvalidated independence assumption of Eq. 11 used by the Pareto search; when configurations are chosen per layer without checking cross-layer effects, jointly optimal behavior is not guaranteed, and the observed degradation is exactly the kind of failure mode one would expect. The paper provides no ablation or experiment testing this assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free ANN-to-SNN conversion framework combining three techniques: AdaFire (layer-wise adaptive burst firing, formalized in Eq. 8), SSC (layer-wise threshold compression, Eqs. 12-14), and IAT (input-adaptive early-exit timesteps, Eq. 17). Configuration search is performed with sensitivity and energy estimators under a per-layer independence assumption (Eq. 11), using a Pareto-frontier approach. The authors report large accuracy gains over the Calibration baseline on ImageNet, neuromorphic datasets, object detection, and segmentation, together with substantial theoretical energy savings; code is provided.","tokens_in":12690,"tokens_out":5667,"duration_ms":53656,"significance":"If the results hold, this is a practically valuable training-free conversion pipeline that improves low-timestep ANN-to-SNN accuracy while cutting energy. The burst-firing formalism in Eq. 8 and the threshold-compression formulation are coherent, and the breadth of tasks (2D, 3D, event-driven, detection, segmentation) is a strength. The paper also ships code and reports very low setup time. However, an internal contradiction in the main ablation table and an unvalidated independence assumption prevent accepting the headline 'simultaneous accuracy gain and energy saving' claim as stated.","major_comments":[{"comment":"The text and Table 5 contradict each other on the central result. The paragraph states that the combined application of AdaFire, SSC, and IAT yields a 70.12% energy reduction and a 0.13% accuracy enhancement on CIFAR-10, while the all-three row reports 95.47% accuracy against a baseline of 96.34% and an AdaFire-only accuracy of 96.69%. That is a 0.87 percentage-point loss relative to the baseline and a 1.22 percentage-point loss relative to AdaFire alone. The same inconsistency propagates to the abstract's 'up to 70.1%' energy savings and the claim of state-of-the-art performance. Please correct the table or the text, and re-derive the affected conclusions; as written, the two statements cannot both be true.","section":"Ablation Study, Table 5"},{"comment":"Eq. 11 reduces the exponential search over layer configurations to a sum of per-layer sensitivity terms by assuming that each layer's sensitivity is independent of other layers' configurations. The paper gives no proof, no validation, and no ablation of this assumption. If cross-layer interactions are non-negligible, the Pareto-selected configuration can be globally suboptimal, which would explain the degraded all-three CIFAR-10 accuracy in Table 5. Please add explicit evidence, for example a small-network comparison between layer-wise independent search and joint search, or a measurement of interaction magnitudes for a subset of layers.","section":"Adaptive-Firing Neuron Model, Pareto Frontier Driven Search Algorithm, Eq. 11"},{"comment":"The paper states that sensitivity, defined by the KL divergence in Eq. 9, is 'demonstrated inversely related to SNN performance (shown in the Appendix)', but the appendix is not included in the manuscript. Because Eq. 11's optimization objective and Eq. 15's constraint both rely on this sensitivity proxy, the missing support is load-bearing. Please include the appendix material or provide an in-text experiment demonstrating the inverse relationship.","section":"Preliminary, Performance Metric, Eq. 9"}],"minor_comments":[{"comment":"The row headers contain only checkmarks, which makes the method combination in each row ambiguous; please add explicit labels such as 'AdaFire only', 'AdaFire + SSC', 'AdaFire + IAT', and 'AdaFire + SSC + IAT'.","section":"Table 5"},{"comment":"Two energy columns contain doubled closing parentheses: '(-27.75%))' and '(-60.25%))'.","section":"Table 5"},{"comment":"The figure caption and inline text include garbled glyphs such as 'A/glyph1197et' and 'A/glyph1197/glyph1197'; the source image should be replaced with a clean version.","section":"Figure 1"},{"comment":"The energy expression 'E = total spikes / 1e-3 * mu (in Watts)' is dimensionally unclear; please specify the unit conversion and state explicitly why spike amplitude in Eqs. 13-14 does not enter the energy count.","section":"Eq. 10"},{"comment":"The text describes an accuracy increase of '11.53%' for ImageNet; since this is a difference in percentage points, please state it as '11.53 percentage points' to avoid ambiguity.","section":"Ablation Study"}],"recommendation":"major_revision","confidential_remarks":"The authors should be asked to resolve the Table 5 inconsistency and provide the missing appendix and independence-assumption validation before the paper is considered further. The manuscript also shows signs of hasty formatting (garbled glyphs, duplicate parentheses), which should be corrected in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper has a good core idea and one serious internal inconsistency. The AdaFire neuron model (Eq. 8) that lets each neuron fire up to phi times per timestep, with per-layer phi found via a Pareto search, is a natural and well-motivated extension of the calibration line. The reported gains over the Calibration baseline at low timesteps are large — e.g., 73.53% vs 25.33% on ImageNet VGG-16 at T=8 — and the mechanism is coherent. That part looks solid.\n\nThe problem is Table 5 and the text around it. The abstract and ablation text claim the full framework (AdaFire+SSC+IAT) delivers a 0.13% accuracy enhancement on CIFAR-10 alongside a 70.12% energy cut. Table 5 shows the full framework at 95.47% accuracy versus a 96.34% baseline — a 0.87% loss — and 1.22% below AdaFire alone (96.69%). The text and table cannot both be right. For CIFAR-100 and ImageNet the combined row does improve accuracy, so the framework isn't uniformly bad, but the CIFAR-10 row directly contradicts the headline 'improves accuracy and energy.' This needs to be fixed before publication, either by correcting the table or weakening the claim.\n\nA second soft spot is the independence assumption under Eq. 11. The search decomposes into per-layer optimizations with no validation that layer interactions are negligible. The authors don't test this, and the CIFAR-10 degradation could be a symptom of exactly that failure mode. They should add an ablation that compares the per-layer search against a few joint configurations.\n\nMinor: several hyperparameters (alpha_base, beta, delta, Starget) are not specified in the main text, and the energy numbers are theoretical. That's fine if labeled, but say so clearly.\n\nOverall, the AdaFire core is a genuine contribution and the accuracy results justify a serious referee. The paper is not ready as-is. I'd send it to review with the expectation of a major revision. My citation of the work in the next year: yes, for the AdaFire neuron model.","headline":"Solid core idea with a serious internal inconsistency in the main ablation table that undermines the combined-claim.","tokens_in":13228,"tokens_out":3513,"would_cite":true,"duration_ms":28226,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Training-free ANN-to-SNN conversion reaches state-of-the-art accuracy at 8 timesteps by letting neurons burst-fire up to a per-layer maximum, cutting theoretical energy by up to 70.1% on CIFAR-10.","keywords":["spiking neural networks","ANN-to-SNN conversion","burst firing","sensitivity-based search","energy-efficient inference","adaptive timesteps"],"falsifier":"Run the layer-wise search on a fixed architecture and dataset twice, once with the layers optimized in forward order and once with the order reversed; if the selected burst-firing limits and threshold ratios, or the resulting accuracy-energy frontier, change materially between the two runs, the independence assumption that makes the search tractable is false.","tokens_in":12199,"feed_emoji":"⚡","tokens_out":7626,"duration_ms":62934,"temperature":0.7,"pith_summary":"This paper argues that the main obstacle to fast, accurate ANN-to-SNN conversion is not clipping or quantization but an 'unevenness error' that appears when input spikes arrive unevenly and a neuron can fire at most once per timestep. It claims this error can be largely removed without any training by giving each layer its own burst-firing limit, the maximum number of spikes a neuron may emit in one timestep, chosen by a Pareto search over estimated accuracy and energy. Two further mechanisms, threshold-ratio spike compression and entropy-based early exit, then cut spike count and latency. If correct, converted SNNs can match or beat retraining-based methods at 8 timesteps while needing only minutes of setup instead of hundreds of GPU-hours, with theoretical energy savings of up to 70.1%, 60.3%, and 43.1% on CIFAR-10, CIFAR-100, and ImageNet.","feed_headline":"Burst firing makes ANN-to-SNN conversion accurate at 8 timesteps","feed_subtitle":"Training-free calibration with per-layer burst spikes cuts theoretical energy by up to 70 percent on CIFAR-10.","key_machinery":"The load-bearing object is the adaptive-firing neuron model, AdaFire, formalized in Eq. 8 as $s^{(\\ell+1)} = \\mathrm{ClipFloor}(W^{(\\ell)} s^{(\\ell)}, T, V_{\\mathrm{th}}^{(\\ell)}, \\varphi^{(\\ell)})$, where $\\varphi^{(\\ell)}$ lets a neuron fire up to that many times per timestep. This replaces the ordinary single-spike clip and is what attacks the unevenness error. Around it sit two efficiency mechanisms: Sensitivity Spike Compression rescales each layer's threshold by a ratio $\\rho^{(\\ell)}$ so regular spike trains are compressed into fewer, larger spikes, and Input-aware Adaptive Timesteps sets a time-dependent confidence boundary $\\alpha_t$ that lets easy inputs exit early. A Pareto-frontier search under the assumption that layers are independent chooses the per-layer $\\varphi$ and $\\rho$ values, turning an exponential search into a sum of per-layer sensitivity terms.","core_discovery":"The paper's central claim is that a training-free conversion framework can simultaneously improve both accuracy and efficiency of converted SNNs by adapting three per-layer or per-input quantities. The adaptive-firing neuron model replaces the standard ClipFloor operation, Eq. 8, with one that permits up to $\\varphi^{(\\ell)}$ spikes per timestep, expanding each neuron's effective output range and thereby shrinking the unevenness error, defined in Eq. 7 as the difference between averaged SNN output and ANN activation. Layer-specific $\\varphi$ and threshold ratio $\\rho$ values are selected by a sensitivity-driven Pareto search that treats the total sensitivity as a sum of per-layer terms. Input-aware adaptive timesteps use an entropy-based confidence measure with a time-dependent boundary to exit early on easy inputs. The paper reports that this combination achieves state-of-the-art accuracy at low timesteps across static, event-driven, 3D, detection, and segmentation benchmarks and saves up to 70.1%, 60.3%, and 43.1% theoretical energy on CIFAR-10, CIFAR-100, and ImageNet, respectively.","pith_inferences":["Editorial inference: the energy numbers are theoretical spike-count savings, so real neuromorphic hardware may show smaller gains because memory access and routing often dominate energy.","Editorial inference: the independence assumption behind the Pareto search likely degrades on deeper or more coupled architectures; testing it by re-optimizing one layer while others are fixed would reveal how much the reported trade-offs depend on it.","Editorial inference: the layer-wise burst-firing idea could likely be folded into trained SNNs or quantization-aware training, not just conversion, as a general way to trade spike count for representational range.","Editorial inference: the entropy-based adaptive timestep rule is architecture-agnostic and could be tested directly on spiking transformers or other attention-based SNNs."],"forward_implications":["Converted SNNs can reach state-of-the-art accuracy at T=8 without any retraining, cutting setup cost from hundreds of GPU-hours to under an hour.","Unevenness error, the dominant conversion error at low timesteps, can be reduced by burst firing rather than by longer simulation.","Layer-specific threshold compression can cut theoretical spike-based energy by more than half on CIFAR-10 with under one percentage point accuracy loss.","Entropy-based early exiting can cut latency roughly 2.4-fold and energy 2.7-fold while slightly improving accuracy.","The same framework transfers to object detection and segmentation, where it reaches comparable mAP with far fewer timesteps than prior spiking detectors."],"supporting_citations":[{"why":"The training-free calibration baseline this framework extends; its layer-wise threshold fitting is the starting point AdaFire replaces.","marker":"(Li et al. 2021a)"},{"why":"Defines the quantized clip-ReLU conversion and the error taxonomy (clipping, quantization, unevenness) the paper's motivation and ImageNet comparisons build on.","marker":"(Bu et al. 2021a)"},{"why":"Supplies the definition of unevenness error as the difference between averaged SNN output and ANN activation, the specific error AdaFire is designed to reduce.","marker":"(Hao et al. 2023a)"},{"why":"Provides the sensitivity metric and the Pareto-frontier search inspiration used to choose per-layer configurations.","marker":"(Cai et al. 2020)"},{"why":"Documents diverse intrinsic burst-firing patterns in cortical neurons, the biological motivation for layer-specific firing limits.","marker":"(Connors and Gutnick 1990)"},{"why":"Argues bursts are a unit of neural information that make transmission reliable, the rationale for using burst firing to reduce conversion error.","marker":"(Lisman 1997)"},{"why":"A temporal early-exit SNN method that the input-aware adaptive timesteps technique builds on for input-dependent inference time.","marker":"(Li et al. 2024)"},{"why":"Optimal conversion method used as a state-of-the-art re-training baseline in Table 2 comparisons.","marker":"(Deng and Gu 2021)"}],"fun_headline_variants":["Burst spike calibration makes SNNs accurate at 8 timesteps","Training-free conversion with burst firing cuts energy by 70%","Adaptive firing neurons shrink conversion error for efficient SNNs","Unified SNN conversion: state-of-the-art accuracy, 70% less energy","Burst spikes: the key to fast, accurate, energy-efficient SNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole search relies on assuming each layer's sensitivity to its own burst-firing and threshold settings can be evaluated without considering the settings of other layers; if layer interactions matter, the chosen configurations may not be globally optimal and the reported accuracy-energy savings could fail on deeper or more coupled networks.","fun_headline_variants_meta":{"raw":{"variants":["Burst spike calibration makes SNNs accurate at 8 timesteps","Training-free conversion with burst firing cuts energy by 70%","Adaptive firing neurons shrink conversion error for efficient SNNs","Unified SNN conversion: state-of-the-art accuracy, 70% less energy","Burst spikes: the key to fast, accurate, energy-efficient SNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000285,"raw_usage":{"total_tokens":1729,"prompt_tokens":1043,"completion_tokens":686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":591}},"tokens_in":659,"tokens_out":686,"duration_ms":6393,"temperature":1.0,"reasoning_tokens":591,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:54:40.339823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the layer-wise search on a fixed architecture and dataset twice, once with the layers optimized in forward order and once with the order reversed; if the selected burst-firing limits and threshold ratios, or the resulting accuracy-energy frontier, change materially between the two runs, the independence assumption that makes the search tractable is false.","supporting_citations":[],"review_version":1}