{"id":"fc9d2ce7-a109-4971-9f88-66820a24c7ee","arxiv_id":"2608.07066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A calibration-only framework jointly quantizes SNN weights and recurrent membrane states via channel-wise power-of-two scale coupling and mixed-precision allocation, preserving accuracy near the floating-point model.","lead":"PTQ4SNN quantizes both the weights and the internal membrane potentials of spiking neural networks to low bit-widths without retraining, using a small calibration set. The method keeps accuracy drops under one point on several vision benchmarks with about 4-bit membrane precision.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on a small input-only calibration set remaining representative across the full T-step membrane trajectory; without multi-split or calibration-sensitivity evidence, this is the least secured link.","rationale":"The reader's weakest_assumption identifies the same condition that I would stress: the method freezes all quantization parameters from a small input-only calibration set, and the paper gives no direct evidence that these parameters are stable across calibration draws or across the temporal trajectory at test time. I agree with the reader's conditional verdict. The paper has real strengths: it reports controlled ablations for the scale bridge and MPBA, honestly states that storage numbers are theoretical rather than measured hardware gains, and uses the same checkpoints and calibration protocol for baselines. However, the central empirical claim is \"consistently preserves competitive accuracy,\" and that consistency is currently supported by single-draw results without error bars or calibration-set robustness tests. The proposed check, re-running the pipeline over multiple calibration subsets and event-stream segments, would directly settle whether the frozen scales and bit assignments generalize. I do not see an internal inconsistency in the main construction; the concern is an unverified robustness condition, which is exactly why the paper should remain conditional rather than being accepted as a turnkey recipe.","tokens_in":12395,"tokens_out":11157,"duration_ms":100848,"concrete_test":"Run the full PTQ4SNN pipeline on SDT-2-256/CIFAR10-DVS (T=10) and on SDT-8-768/ImageNet with 5 independent random calibration subsets of the designated size, plus a calibration split taken from a different segment of the DVS event stream. Record W4/M4 accuracy, the element-count-weighted membrane precision, and the per-channel bit-assignment vector for each split. If the accuracy range across splits exceeds 1.0 point, or if the bit-assignment vectors disagree by more than a small tolerance, the calibration-representativeness assumption is not supported; if the range is small and assignments are stable, the concern is retired.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (13)-(15) estimate per-channel firing rates and a sensitivity score from calibration forward passes; these determine both the power-of-two bridge exponents k_c (Eq. 9) and the 2/4/8-bit allocation. Because membrane errors are threshold-sensitive, a scale that clips a rare membrane excursion, or a 2-bit assignment to a channel whose 4-bit sensitivity was low but whose 2-bit error is large, can flip a spike and alter the recurrent trajectory for subsequent timesteps. The sensitivity signal in Eq. 14 is computed at a 4-bit reference, so 2-bit channels are protected using an extrapolation from 4-bit behavior. The ablations use 1024 calibration samples and additionally tune the activity-sensitivity weight and the sparse-protection percentile on a held-out calibration subset, so the fitted allocation is itself a function of a small sample. Main results are single draws from unspecified calibration splits, with no error bars or calibration-set sensitivity analysis. If the calibration set does not capture the membrane distributions encountered during inference, especially for long event streams such as CIFAR10-DVS (T=10), the reported 0.38-1.00 point drops may not be reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents PTQ4SNN, a post-training quantization framework that jointly quantizes weights and recurrent membrane states in spiking neural networks using only a small calibration set. It introduces two components: a channel-wise Unified Scale Bridge that constrains each membrane scale as s_mem,c = s_w,c * 2^{k_c}, enabling shift-compatible scale conversion, and a Mixed-Precision Bit Allocation that assigns 2/4/8-bit membrane precision per channel according to firing activity and a gradient-based sensitivity score under an element-count-weighted average bit budget. The method is evaluated on ImageNet-1K classification (SDT-8-768, Meta-SpikeFormer, SEW-ResNet18), CIFAR10-DVS event classification, and Pascal VOC2012 semantic segmentation, reporting drops of 0.38-1.00 percentage points from floating-point checkpoints at W4/M4, with comparisons against weight-PTQ baselines (BRECQ, GPTQ) and a reimplemented FlowQ baseline.","tokens_in":12675,"tokens_out":7587,"duration_ms":66771,"significance":"If the results are robust, this is a useful contribution: it extends PTQ to recurrent membrane states, which are a substantial and previously mostly unquantized component of SNN inference cost. The reusable projection-LIF pair abstraction is clean, and the Unified Scale Bridge is a practical way to keep scale conversion shift-compatible. The paper is commendable for running controlled comparisons (same checkpoints, calibration samples, and first/last-layer settings across methods) and for explicitly disclaiming measured hardware gains and including a candid resource-accounting discussion in the appendix. The MPBA idea is interesting, and its independent benefit over uniform M4 is demonstrated on two datasets (Table 6).","major_comments":[{"comment":"The main results are single runs from one calibration set, with no error bars or calibration-split sensitivity analysis. Because the membrane scales (Eq. 9) and the 2/4/8-bit allocation (Eqs. 13-16) are fitted to calibration statistics, and membrane errors accumulate over the T-step trajectory (especially for CIFAR10-DVS at T=10), the central robustness claim needs evidence that the reported drops are stable across calibration draws and calibration sizes. Please report mean +/- standard deviation over at least three calibration splits/seeds and a calibration-size sweep.","section":"Experiments, Tables 1-3"},{"comment":"The sensitivity score g_c is computed at a 4-bit reference and used to allocate 2-bit channels. Since 2-bit quantization errors are strongly nonlinear and threshold-sensitive, the paper should validate that the 4-bit sensitivity ranking transfers to 2-bit channels. A direct check is to compare, on the calibration set, the g_c ranking with the actual per-channel loss error when quantizing to 2 bits, or to ablate the sensitivity term specifically for the 2-bit assignment. Without such evidence, the MPBA gain in Table 6 may be attributable to the activity term alone.","section":"Mixed-Precision Membrane Bit Allocation, Eq. (14)"},{"comment":"The appendix states that 'the first projection and classifier remain at higher precision in the SDT classification experiments,' but Tables 1 and 2 and the abstract claim W4 quantization without this caveat. This is a material omission: for SDT-8-768 and Meta-SpikeFormer the effective weight precision is higher than 4 bits for some layers. Please state the actual bit widths in the main text and tables, and report the corresponding average weight precision and the resulting memory savings, or adjust the W4 label to avoid overstating the configuration.","section":"Experimental Details vs. Abstract/Tables"},{"comment":"All FlowQ rows in Tables 1-3 are from the authors' reimplementation, but the paper does not validate this reimplementation against FlowQ's original results or release the code. Given that PTQ4SNN outperforms the reimplemented FlowQ by large margins in some settings (e.g., 75.16 vs. 63.74 on SDT-8-768), the comparison's fairness depends on the reimplementation's fidelity. Please provide a validation against published FlowQ numbers where checkpoints are available, a more detailed specification of the reimplementation, or a public code release.","section":"Baselines and Evaluation Protocol"}],"minor_comments":[{"comment":"The phrase 'the completeBRECQreusebaseline' is missing spaces and should read 'the complete BRECQ reuse baseline.'","section":"Main text near Table 1"},{"comment":"The variables b_nonstem and b_all are used in the caption but their computation, especially the exclusion of the protected 16-bit stem, should be defined explicitly in the table caption.","section":"Table 4"},{"comment":"The term z^2 is ambiguous between an element-wise square and a squared norm; please state the intended operation.","section":"Equation (14)"},{"comment":"Please clarify whether the reported experiments quantize the leak factor tau and, if so, how; if tau remains floating-point, the 'shift-compatible' claim should be qualified.","section":"Unified Scale Bridge, Eq. (11)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid PTQ contribution with controlled evaluation and a candid appendix, which are strengths. However, the single calibration run and the undisclosed high-precision first/last layers are likely to be seen as overselling the 'W4' claim in a competitive venue. The FlowQ reimplementation should be released or validated. The requested experiments (calibration sensitivity, sensitivity-metric validation, and protocol disclosure) are feasible and would substantially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: PTQ4SNN is a solid engineering contribution that extends post-training quantization to recurrent membrane states in SNNs, and it largely works. If I were editing this venue, I'd send it to review.\n\nWhat's genuinely new: the channel-wise power-of-two scale bridge (s_mem,c = s_w,c * 2^k_c) and the activity/sensitivity-driven 2/4/8-bit allocation for membrane channels. FlowQ quantizes membranes but with uniform layer-wise precision; NeuronQuant and SNNQ don't touch membrane states. So this fills a real gap.\n\nThe experimental core is honest and controlled. They reimplement BRECQ, GPTQ, and FlowQ on the same checkpoints, compare fair W4/M4 baselines, and run ablations isolating the bridge and the allocation. Results are consistent across static classification, event-based CIFAR10-DVS, and segmentation. Drops under 1 point on the transformer backbones are believable. The supplementary is transparent about what they did and didn't measure: they don't claim hardware latency gains, just tensor storage and a cost model.\n\nSoft spots, in order of seriousness. First, no code, and the main results have no error bars or calibration-split variation. The allocation hyperparameters are tuned on a held-out calibration subset, and the sensitivity score is computed at 4-bit reference then used to justify 2-bit assignments. Second, the 'M4' label is slightly generous: first/last layers and stem stay at higher precision, and the true average includes protected 16-bit channels in the segmentation/cost tables. They disclose this, so it's a labeling issue more than a result issue. Third, the calibration set is 1024 samples; the stress-test worry that these statistics may not cover long event streams is legitimate but not fatal—this is standard PTQ practice and the paper's ablations give no sign of collapse.\n\nThe central claim holds up: a calibration-only method that quantizes membrane states to about 4 bits with minimal accuracy loss. I'd like to see code and error bars before treating it as a turnkey recipe.\n\nVerdict: send it to review. I'd cite it in my SNN quantization work. Worth a reading group if anyone in the group cares about deployment.","headline":"Solid, incremental PTQ for SNN membranes with controlled experiments; deserves a serious referee despite calibration-stability caveats.","tokens_in":13181,"tokens_out":2225,"would_cite":true,"duration_ms":18789,"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 calibration-only method quantizes SNN membrane states to about 4 bits with under 1-point accuracy drops across architectures.","keywords":["post-training quantization","spiking neural networks","membrane potential quantization","mixed-precision bit allocation","unified scale bridge","spike-driven transformer","event-based classification","semantic segmentation"],"falsifier":"A concrete check would be to measure per-channel membrane firing rates on the held-out test set and compare them with calibration-set rates; if a substantial fraction of channels have test firing rates outside the calibrated range, or if a long event stream much longer than the calibration stream degrades accuracy monotonically, the representativeness assumption is false. A direct experiment: calibrate on only two of the ten CIFAR10-DVS temporal bins and evaluate on all ten; accuracy should stay near the all-bin calibration result if the claim is robust.","tokens_in":12192,"feed_emoji":"⚡","tokens_out":6705,"duration_ms":58020,"temperature":0.7,"pith_summary":"This paper tries to establish that the recurrent membrane potential, not just the weights, can be a first-class target for post-training quantization in spiking neural networks. It proposes PTQ4SNN, which uses only a small calibration set and keeps backbone weights frozen, jointly quantizing weights to 4 bits and membrane states to an average of 4 bits per channel. The key claim is that this can be done with negligible accuracy loss across convolutional SNNs, spike-driven Transformers, event-based recognition, and semantic segmentation. For instance, ImageNet accuracy drops are 0.74 and 0.38 points on SDT-8-768 and Meta-SpikeFormer, and 1.00 point on CIFAR10-DVS. If true, this removes the last floating-point state that remains in low-bit SNN inference and lowers recurrent storage and data movement without retraining.","feed_headline":"SNN membrane states cut to 4 bits with under 1-point drops","feed_subtitle":"Calibration-only method ties each membrane scale to its weight scale by a power of two and assigns 2/4/8 bits per channel.","key_machinery":"The central object is the projection-LIF pair, where an LIF neuron is a leaky integrate-and-fire spiking neuron. Two mechanisms carry the argument: the Unified Scale Bridge, defined by the identity $s_{\\text{mem},c} = s_{w,c} \\cdot 2^{k_c}$, couples each membrane channel's quantization scale to the preceding weight channel's scale by an integer power of two, so scale conversion becomes a shift while the quantization range adapts to the membrane distribution; and Mixed-Precision Bit Allocation, which computes per-channel firing rates $r_c$ and a gradient-based sensitivity $g_c$, combines them into a score $a_c = \\beta \\hat{r}_c + \\gamma \\hat{g}_c$, and assigns 2/4/8-bit widths by score quantiles under an element-count-weighted average budget $\\sum_c b_c N_c / \\sum_c N_c \\approx b_m$. The bridge exponent $k_c$ is calibrated only after the per-channel bit width is assigned.","core_discovery":"On its own terms, the paper claims that membrane states are quantizable at the same precision as weights, provided the quantization is done channel-wise and is aware of firing activity and sensitivity. The central discovery is that a channel-wise power-of-two bridge between weight scale and membrane scale, combined with mixed 2/4/8-bit allocation under an average-bit budget, prevents the temporal error accumulation that makes naive membrane quantization collapse. Direct weight-scale reuse drops ImageNet accuracy by 71 points on SDT-8-768, while PTQ4SNN drops only 0.74 points. The method is formulated over reusable projection-LIF pairs and reports strong W4/M4 results across static classification, event-based classification, and Pascal VOC2012 segmentation, all without backbone retraining.","pith_inferences":["A natural extension would be to test the shift-compatible bridge in an integer-only accelerator, where the recurrent membrane update could run with shifts instead of floating-point multiplies; the paper's cost model counts this benefit but does not measure silicon-level latency or energy.","The sensitivity score built from a gradient-spike inner product is a local proxy, and one could test whether tracking firing-rate drift over longer sequences, such as T=100 event streams, improves allocation without increasing calibration cost.","The same activity-and-sensitivity bit allocation could transfer to other recurrent stateful layers such as RNNs or state-space models that store per-channel hidden states, though the paper evaluates only LIF membrane channels.","A potential failure point the paper does not resolve is that channels silent during calibration but active at inference would receive 2 bits and may clip; a calibration set covering rare-event channels would be needed before deployment on long or distribution-shifted streams."],"forward_implications":["A pretrained SNN can be deployed with both weights and membrane states quantized to near 4 bits using only calibration inputs, with no retraining and no original training data.","The same projection-LIF recipe applies across convolutional SNNs and spike-driven Transformers, so membrane quantization no longer needs an architecture-specific training pipeline.","Channel-wise mixed precision beats uniform 4-bit membrane precision at the same average bit budget: +0.440 points on CIFAR-10 and +0.522 points on ImageNet-1K on SEW-ResNet18.","A hardware-oriented estimate shows packed membrane resident state shrinking from 30.29 MiB at M32 to 5.28 MiB under an average M4 budget with per-channel tags and shift metadata, roughly an 8x reduction in state storage; actual latency and energy gains depend on a target integer implementation."],"supporting_citations":[{"why":"Defines FlowQ, the layer-wise power-of-two membrane scale-coupling baseline that PTQ4SNN extends to channel-wise coupling and consistently outperforms.","marker":"Asim, Aung, and Lee 2026"},{"why":"Supplies BRECQ, a generic weight PTQ baseline whose W4/M4 direct scale reuse collapses, providing the controlled comparison that motivates membrane-aware scales.","marker":"Li et al. 2021"},{"why":"Supplies GPTQ, the other generic weight PTQ baseline whose W4/M4 reuse also fails without membrane-aware scaling.","marker":"Frantar et al. 2023"},{"why":"Shows that prior joint weight-membrane quantization (MINT) required quantization-aware training, positioning PTQ4SNN as a calibration-only alternative.","marker":"Yin et al. 2024"},{"why":"Q-SNNs also quantizes membrane states during training, illustrating the training dependence that PTQ4SNN removes.","marker":"Wei et al. 2024"},{"why":"NeuronQuant is the closest existing SNN PTQ work but leaves membrane states in floating point, defining the gap PTQ4SNN fills.","marker":"Li et al. 2025"},{"why":"Quantizes spike-driven Transformers at low weight precision during training, serving as the Transformer-side contrast that PTQ4SNN matches without retraining.","marker":"Qiu et al. 2025"},{"why":"Defines AdaRound, a weight-rounding PTQ baseline used in the experimental protocol and controlled comparisons.","marker":"Nagel et al. 2020"}],"fun_headline_variants":["Power-of-two scale bridge lets SNN membranes hit 4 bits, <1pt drop","Membrane-aware quantization: 4-bit weights and states, <1pt ImageNet drop","Scale-bridge quantization cuts SNN membranes to 4 bits without retraining","Mixed 2/4/8-bit membrane allocation keeps SNN accuracy under 4-bit budget"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The per-channel firing rates and sensitivity scores gathered on a small calibration set remain representative over the whole inference trajectory, so the chosen 2/4/8-bit assignments and power-of-two scales do not clip or under-quantize membrane channels when longer or different inputs arrive.","fun_headline_variants_meta":{"raw":{"variants":["Power-of-two scale bridge lets SNN membranes hit 4 bits, <1pt drop","Membrane-aware quantization: 4-bit weights and states, <1pt ImageNet drop","Scale-bridge quantization cuts SNN membranes to 4 bits without retraining","Mixed 2/4/8-bit membrane allocation keeps SNN accuracy under 4-bit budget"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001458,"raw_usage":{"total_tokens":5854,"prompt_tokens":917,"completion_tokens":4937,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":4844}},"tokens_in":533,"tokens_out":4937,"duration_ms":30779,"temperature":1.0,"reasoning_tokens":4844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:27:28.465440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check would be to measure per-channel membrane firing rates on the held-out test set and compare them with calibration-set rates; if a substantial fraction of channels have test firing rates outside the calibrated range, or if a long event stream much longer than the calibration stream degrades accuracy monotonically, the representativeness assumption is false. A direct experiment: calibrate on only two of the ten CIFAR10-DVS temporal bins and evaluate on all ten; accuracy should stay near the all-bin calibration result if the claim is robust.","supporting_citations":[{"cited_title":"2026 31st Asia and South Pacific Design Automation Conference (ASP-DAC) , pages =","cited_arxiv_id":null,"evidence_quote":"Defines FlowQ, the layer-wise power-of-two membrane scale-coupling baseline that PTQ4SNN extends to channel-wise coupling and consistently outperforms."},{"cited_title":"Proceedings of the 32nd ACM International Conference on Multimedia , pages =","cited_arxiv_id":null,"evidence_quote":"Q-SNNs also quantizes membrane states during training, illustrating the training dependence that PTQ4SNN removes."}],"review_version":1}