{"id":"76bcbac3-c249-4de1-94a5-9d5bcbb34d82","arxiv_id":"2411.17431","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Injecting uniform noise before activation quantization in ANN training improves converted low-latency SNN accuracy without runtime noise correction.","lead":"Spiking neural networks can be made more accurate at low latency by adding random noise during the training of the quantized neural network that is later converted into the spiking network. This training-time trick avoids the extra computation of runtime spike-correction methods and works on very deep ResNet models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unbiasedness theorem in §4.4 applies only for T≫p, yet Noise Adaptor's gains are largest at small T; the claimed low-latency mechanism is unverified.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the paper's theoretical bridge relies on an unbiasedness theorem that does not justify the claimed mechanism for low-latency SNNs. My stress-test sharpens this by noting that the empirical gains are concentrated at low T, exactly where the theory (which requires T≫p) is least applicable. This is not a fatal flaw—the within-paper comparison in Table 1 is a genuine controlled experiment, and the method may still work for reasons not yet articulated. However, the paper's central explanatory claim ('better accounts for dynamic differences') is unsupported, and the lack of code or error bars further weakens reproducibility. The reader's CONDITIONAL verdict remains appropriate: the paper should be published only if the mechanism is tested (e.g., via the proposed ablation) or the claims are softened to a purely empirical contribution. I therefore leave the verdict unchanged.","tokens_in":13904,"tokens_out":13188,"duration_ms":121439,"concrete_test":"On CIFAR-10 ResNet-18, train a control variant where noise is added after rounding instead of before, e.g., \\hat v = s·round(clip(v/s,0,p)) + \\eta with \\eta ~ U(−0.5s,0.5s), so the transition probability is independent of the fractional part of v/s and E[\\hat v] ≠ clip(v,0,sp). Evaluate SNN accuracy at T=2 and T=8 and compare to the standard Noise Adaptor results in Table 1 (95.26% and 95.95%). If the control matches those accuracies, the specific pre-rounding noise placement and the §4.2 fractional-distance mechanism are not load-bearing; if it drops substantially, the claimed mechanism is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central causal claim is that injecting U(−0.5,0.5) noise before rounding (Eq. 5) models the occasional noise of low-latency SNNs and thereby improves conversion. The only theoretical support is §4.4/Eq. (10): the noisy quantizer's expected output equals clip(v,0,sp), which is the rectified-ReLU response curve of an SNN only in the limit T≫p. However, the method is explicitly aimed at low-latency SNNs, and its largest measured gains occur at small T (CIFAR-10 T=2: +1.61% over baseline; ImageNet T=8: +3.14%). At these time steps the SNN response curve is a coarse staircase (Fig. 2), not the smooth clip function, so the unbiasedness theorem does not cover the operating regime where the method claims to help. Even when the expectation equals clip, training loss is nonlinear in activations; unbiased activations do not imply an unbiased loss or an optimal SNN conversion. The §4.2 assertion that transition probability depends on fractional distance to the next quantization level is also never measured against actual SNN spike-count fluctuations. The reported gains could therefore stem from stochastic-rounding regularization rather than from the claimed noise-adaptation mechanism. This does not refute the empirical improvements, but it leaves the paper's explanatory claim unverified for precisely the low-latency setting it targets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Noise Adaptor, a training-time modification for activation-quantized ANNs used in ANN-to-SNN conversion. Uniform noise epsilon ~ U(-0.5,0.5) is added to the pre-activation before clipping and rounding (Eqs. 5-6), while keeping the learnable quantization scale. The authors report that this improves the accuracy of converted low-latency SNNs on CIFAR-10 and ImageNet, avoids runtime noise-correction techniques, and scales to ResNet-101/152. A supplementary proof establishes that the noisy quantizer's expected output equals clip(v,0,sp), the rectified-ReLU response in the limit of large simulation time.","tokens_in":14173,"tokens_out":8822,"duration_ms":80193,"significance":"If the empirical results are reproducible, the method offers a simple and effective drop-in modification to quant-ANN-to-SNN conversion pipelines, backed by a clean controlled evaluation in Table 1 (with/without noise injection under the same pipeline). The supplement's proof of Eq. (10) is a correct self-contained derivation of the stochastic-rounding unbiasedness property, and the experiments cover networks up to ResNet-152, addressing a scalability gap in the literature. The main weakness is that the theoretical analysis only supports the large-T limit, while the method's practical gains appear at small T, so the mechanistic explanation of the low-latency benefit remains unverified.","major_comments":[{"comment":"The theoretical support does not cover the operating regime where the method's gains are largest. Eq. (10) shows that the expected value of the noise-injected quantizer equals clip(v,0,sp), which is the response curve of the SNN only as T grows large. At small T, the SNN response is a staircase (Figure 2), so the unbiasedness theorem does not imply that the ANN quantizer matches the actual SNN response. Moreover, because the training loss is nonlinear in activations, unbiased activations do not imply an unbiased loss or optimal conversion. Since the largest accuracy improvements are reported at low latency (CIFAR-10 T=2: +1.61%; ImageNet T=8: +3.14% in Table 1), the paper's central claim that noise injection models occasional noise during low-latency conversion is not supported by the provided theory. The authors should either provide a theory or measurements covering finite T, or explicitly restrict the theoretical claim to large T and present the low-T gains as empirical.","section":"§4.4, Eq. (10), Figure 2"},{"comment":"The noise model's central assumption is unmeasured. The paper assumes that the probability of a spike-count transition in the SNN depends on the fractional distance to the next quantization level, and models this by adding uniform noise before rounding. No experiment in the paper verifies that the actual occasional-noise statistics in the converted SNN match this model. Without such validation, the reported gains could equally be explained by stochastic rounding acting as a regularizer. I recommend adding a diagnostic experiment that compares the distribution of predicted state transitions (from Eq. (5)) with measured spike-count fluctuations during SNN inference, or tempering the causal language throughout Section 4.","section":"§4.2, Eqs. (5)-(6)"},{"comment":"The cross-method comparison mixes architectures. The claim that Noise Adaptor 'consistently achieves the highest SNN accuracy across all time steps' is supported by comparing Noise Adaptor on ResNet-50/101/152 with QCFS and SlipReLU on ResNet-34. The only matched architecture for all three methods is ResNet-34, where Noise Adaptor starts from a lower ANN accuracy (71.26%) yet still attains good SNN accuracy; this should be highlighted and the cross-architecture comparison should be clearly qualified. In addition, Table 1 omits the without-NA baseline at T=256 for ImageNet, so the headline 77.07% figure is not directly compared to the same pipeline without noise injection at that time step.","section":"§5.3, Table 2"}],"minor_comments":[{"comment":"The word 'catagory' should be 'category'.","section":"Section 2"},{"comment":"The entry '36,31%' for Noise Adaptor ResNet-34 at T=16 should be '36.31%'.","section":"Table 2"},{"comment":"The sentence 'like doubling it to ϵ ∼ U(−0.5, 0.5)' is confusing; if the base noise is U(−0.5, 0.5), doubling the amplitude would yield U(−1, 1). Please correct the example.","section":"Section 4.2"},{"comment":"The expression for the scale gradient in line 14 is hard to read; the intended division by sqrt(len(x_clip) * p) (or similar) should be written with explicit parentheses.","section":"Algorithm 1"},{"comment":"The heading 'Updations on the Conversion Process' should be 'Updates to the Conversion Process'.","section":"Section 4.3"},{"comment":"The column headers 'CIFAR-10 ANN T T T' and 'ImageNet ANN T T T' are ambiguous; please list the T values explicitly, e.g., 'ANN, T=1, T=2, T=4, T=8, T=16'.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The empirical core is solid and the method is simple, so I believe the paper can be made publishable after the authors either add evidence for the noise-adaptation mechanism or soften the explanatory claims. The main risk is that the theoretical section in §4.4 may be seen as overclaiming for the low-latency setting. The paper would also benefit from a careful proofreading pass and an explicit caveat about the architecture-matched comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Give this one a serious referee. The controlled comparison in Table 1 is the real result: same pipeline, with and without noise injection, and the noise-injected version improves SNN accuracy at nearly every time step, including the low-latency regime (T=2 on CIFAR-10, T=8 on ImageNet). That is a clean, useful empirical finding, and it is new to me. The method is simple and training-only, which matters for hardware deployment, and the extension to ResNet-101/152 is a genuine if incremental scaling data point.\n\nWhat is not established is the causal story. The paper claims U(−0.5,0.5) noise before rounding models occasional spike-count noise, but this is asserted, not measured. The theoretical support (Eq. 10) is a correct unbiasedness proof for the stochastic quantizer, but it applies to the expected activation, not to SNN accuracy; at small T the SNN response is a coarse staircase, so the bridge to the low-latency setting is missing. The stress-test note is right that the largest gains are at small T where the theorem is least relevant. Training a stochastic quantizer may simply be providing helpful regularization, and the paper does not distinguish that from the claimed noise-adaptation mechanism.\n\nAlso: no code released, no error bars, and cross-method comparisons rely on published numbers from different training protocols. The pooling-agnostic claim is a reasonable observation but not a systematic study. There are typos (catagory, compatability) and the stochastic-rounding literature is not cited; the authors discuss the difference but should acknowledge prior art more explicitly.\n\nNone of this refutes the empirical improvement. The method is simple enough to be worth testing, and the controlled baseline gives me reasonable confidence. But the paper would be stronger if the authors measured the actual spike-count fluctuation statistics they claim to model, or at least softened the mechanism claims and framed the method as a stochastic-rounding technique that empirically helps conversion.\n\nVerdict: send to review. The empirical claim is sufficiently clean and the method is sufficiently simple that a good referee can push on the missing verification without much effort. I'd be surprised if the headline results vanish on replication, but I would not be surprised if the mechanism turns out to be unremarkable stochastic-rounding regularization.","headline":"Noise-injected quantization empirically improves low-latency SNN conversion, but the claimed mechanism is unsupported; the controlled comparison justifies peer review.","tokens_in":14692,"tokens_out":1448,"would_cite":true,"duration_ms":15325,"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":"Adding a uniform random jitter to the ANN activation quantizer before clip-and-round, called Noise Adaptor, makes converted low-latency spiking networks more accurate, reaching 96.72% on CIFAR-10 and 77.07% on ImageNet without runtime…","keywords":["spiking neural networks","ANN-to-SNN conversion","low-latency SNN","activation quantization","noise injection","stochastic rounding","quantization-aware training","ResNet"],"falsifier":"Run a converted SNN at a fixed time step, record each neuron's spike-count deviation from the value the ANN quantizer predicts for its pre-activation, and condition those deviations on the fractional distance to the next quantization level. The paper's model predicts that the probability of a one-count deviation equals that fractional distance; if the measured probabilities deviate systematically, or if deviations do not depend on fractional distance at all, the noise-injection premise is wrong.","tokens_in":1643,"feed_emoji":"⚡","tokens_out":1635,"duration_ms":82204,"temperature":0.7,"pith_summary":"The paper sets out to show that a one-line change to how low-bit artificial networks are trained can fix the main accuracy loss in ANN-to-SNN conversion: adding a uniform random value to the signal just before it is rounded during training. This makes the quantizer stochastic, and the paper argues this mirrors the occasional spike-noise that appears when the trained ANN is simulated as a spiking network. The claimed result is that converted spiking networks run accurately at short simulation times without runtime noise correction, and that the method scales to activation-quantized ResNet-101 and ResNet-152. On CIFAR-10 the converted ResNet-18 reaches 96.72% at T=16; on ImageNet the converted ResNet-50 reaches 77.07% at T=256.","feed_headline":"Uniform jitter in ANN training improves converted spiking networks","feed_subtitle":"Adding U(-0.5,0.5) noise before rounding beats runtime spike-noise correction and scales to deep ResNets.","key_machinery":"The central object is the noise-injected activation quantizer, called Noise Adaptor: instead of $\\hat{v}^l = s_l \\lfloor \\mathrm{clip}(v^l/s_l, 0, p)\\rceil$, the forward pass uses $\\hat{v}^l = s_l \\lfloor \\mathrm{clip}(v^l/s_l + \\epsilon, 0, p)\\rceil$ with $\\epsilon \\sim U(-0.5, 0.5)$, and gradients are obtained with a straight-through estimator plus a trainable scale factor $s_l$. The mechanism does two jobs at once: the noise creates a Bernoulli rounding rule whose up-probability equals the fractional distance to the next quantization level, which is the paper's model of how occasional noise perturbs a spiking neuron's spike count, and the expectation of the quantizer equals the clipped-ReLU response curve of a spiking neuron at long simulation times, so the ANN is implicitly trained toward the SNN's target response.","core_discovery":"The central claim is that training an activation-quantized ANN with $\\epsilon \\sim U(-0.5, 0.5)$ added to $v^l/s_l$ before the clip-and-round step, as in Eqs. (5)-(6), produces a network whose conversion to a spiking network is more accurate at low latency than conversion of the same architecture trained by ordinary deterministic quantization. The injected noise makes the rounding decision probabilistic: a pre-activation whose fractional distance to the next quantization level is $w$ rounds up with probability $w$, so the expected quantizer output is exactly the clipped linear function $\\mathrm{clip}(v^l,0,s_l p)$ (Eq. 10, proven in the supplement). The paper interprets this as training the ANN to be robust to spike-count changes of one, matching the behavior of occasional noise in spiking neurons, and reports consistent accuracy improvements over the no-noise baseline across CIFAR-10 and ImageNet at every tested time step.","pith_inferences":["If the benefit comes from making the quantizer unbiased, then the specific choice of U(-0.5,0.5) may be replaceable by any zero-mean noise of the same support; a direct comparison against standard stochastic rounding would separate the paper's spike-dynamics explanation from the simpler unbiased-expectation effect.","The paper's analysis suggests the advantage should be largest when the simulation time T is larger than the quantization level count p but not yet so large that the SNN response curve is smooth; plotting the accuracy gap against T/p would test the implicit-optimization story.","The compatibility experiments show that runtime corrections still help at very early time steps, so a hybrid recipe of noise-injected training plus cheap correction only for very small T could deliver both ultra-low latency and high accuracy, a combination the paper does not emphasize.","Because the method changes only the training-time quantizer and is pooling-agnostic, it should transfer to other conversion-based tasks such as object detection or transformer-based SNNs, though the paper does not demonstrate this."],"forward_implications":["A network trained with Noise Adaptor reaches higher SNN accuracy at every tested time step than the same architecture trained by deterministic activation quantization, so low-latency SNN inference no longer depends on runtime noise correction for the reported settings.","Because runtime correction such as negative spikes or offset spikes can be omitted, the converted SNN keeps the standard integrate-and-fire neuron and a single-pass control flow, simplifying deployment on neuromorphic hardware.","The method extends conversion to activation-quantized ResNet-101 and ResNet-152; these very deep models require more time steps to approach ANN accuracy, while ResNet-34 and ResNet-50 already give strong accuracy at T=32 and T=64 respectively.","Max-pooling pre-trained ANNs can be used for SNN conversion by swapping in average-pooling during noise-aware quantization training, so existing full-precision checkpoints remain usable.","ANN accuracy slightly drops when Noise Adaptor is applied, yet SNN accuracy rises, indicating that raw ANN accuracy is not the right proxy for conversion quality."],"supporting_citations":[{"why":"Defines the QCFS quant-ANN-to-SNN conversion framework and response-curve analysis that Noise Adaptor extends and compares against.","marker":"[3]"},{"why":"Introduces the occasional-noise problem in low-latency conversion and the runtime negative-spike correction that Noise Adaptor is designed to replace.","marker":"[22]"},{"why":"Provides the TTRBR deep ResNet conversion baseline that Noise Adaptor outperforms on ImageNet.","marker":"[29]"},{"why":"Describes residual-membrane-potential noise reduction, a runtime-free correction baseline relevant to low-latency conversion.","marker":"[15]"},{"why":"Presents offset-spike correction, used in the paper to test Noise Adaptor's compatibility with runtime noise correction.","marker":"[16]"},{"why":"Gives the SlipReLU baseline, another conversion method that avoids runtime noise correction during SNN inference.","marker":"[19]"}],"fun_headline_variants":["Training-time noise boosts converted spiking nets","Inject noise in ANN training for better SNNs","Probabilistic rounding improves ANN-to-SNN conversion","Jittered quantization yields low-latency SNNs","Noise-injected ANN conversion scales to deep SNNs"],"cache_read_input_tokens":16896,"weakest_assumption_plain":"The method's success rests on the assumption that the random spike-count flutter observed in a low-latency spiking network is faithfully represented by adding uniform noise to the value being rounded during ANN training; if the real flutter follows a different rule, the injected noise tunes the network for the wrong perturbation.","fun_headline_variants_meta":{"raw":{"variants":["Training-time noise boosts converted spiking nets","Inject noise in ANN training for better SNNs","Probabilistic rounding improves ANN-to-SNN conversion","Jittered quantization yields low-latency SNNs","Noise-injected ANN conversion scales to deep SNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2316,"prompt_tokens":924,"completion_tokens":1392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":1317}},"tokens_in":540,"tokens_out":1392,"duration_ms":10612,"temperature":1.0,"reasoning_tokens":1317,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:07:17.862803+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a converted SNN at a fixed time step, record each neuron's spike-count deviation from the value the ANN quantizer predicts for its pre-activation, and condition those deviations on the fractional distance to the next quantization level. The paper's model predicts that the probability of a one-count deviation equals that fractional distance; if the measured probabilities deviate systematically, or if deviations do not depend on fractional distance at all, the noise-injection premise is wrong.","supporting_citations":[{"cited_title":"Optimal ann-snn conversion for high- accuracy and ultra-low-latency spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the QCFS quant-ANN-to-SNN conversion framework and response-curve analysis that Noise Adaptor extends and compares against."},{"cited_title":"Quantization frame- work for fast spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Introduces the occasional-noise problem in low-latency conversion and the runtime negative-spike correction that Noise Adaptor is designed to replace."},{"cited_title":"Training much deeper spiking neural networks with a small number of time-steps","cited_arxiv_id":null,"evidence_quote":"Provides the TTRBR deep ResNet conversion baseline that Noise Adaptor outperforms on ImageNet."},{"cited_title":"Reducing ANN-SNN Conversion Error through Residual Membrane Potential","cited_arxiv_id":"2302.02091","evidence_quote":"Describes residual-membrane-potential noise reduction, a runtime-free correction baseline relevant to low-latency conversion."},{"cited_title":"A unified optimization framework of ann-snn conversion: Towards optimal mapping from activa- tion values to firing rates","cited_arxiv_id":null,"evidence_quote":"Gives the SlipReLU baseline, another conversion method that avoids runtime noise correction during SNN inference."}],"review_version":1}