{"id":"34e37c11-f174-4156-94f5-be0984dac18b","arxiv_id":"2506.05679","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A binary-encoded integer spiking neuron with range alignment matches or beats ANN accuracy on ImageNet, COCO, and CIFAR100 at lower energy cost.","lead":"This paper introduces IBRA-LIF, a spiking neuron that fires integer values during training and converts them to binary spikes during inference, reporting state-of-the-art accuracy on ImageNet, COCO, CIFAR100, and CIFAR10-DVS. The authors claim their spiking networks match or exceed artificial neural networks with the same architecture while retaining 6.3x lower energy consumption.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (6) omits the least-significant bit (sum over b=1..B with weights 2^b), so as written binary inference reconstructs 2*floor(O/2) instead of O, leaving the main mechanism not reproducible as printed.","rationale":"The reader's identified concern about the global scaling factor N=100 transferring across tasks is real but secondary: the paper already reports strong ImageNet and COCO numbers with N=100, so the issue mainly affects generality and missing per-layer range statistics, not whether the central construction is internally correct. The more immediate problem is Eq. (6). The paper's core contribution is exact binary expansion of integer-valued activations, and every downstream claim about exponential capacity and energy efficiency uses that expansion. As printed, the summation starts at b=1 and uses weights 2^b, which discards the least-significant bit and doubles every retained bit position. Consequently, odd integer multiples of 1/N cannot be represented at inference even though Eq. (10) can produce them during training. This would make the reported accuracies impossible to reproduce from the equations alone. The most plausible explanation is a simple off-by-one typo, and the intended construction is standard and likely sound, so this does not warrant rejection. But it does warrant a correction, an explicit numerical example, and ideally code before the central claim can be fully verified. That conclusion matches the reader's CONDITIONAL verdict, so no change to the verdict is needed.","tokens_in":12693,"tokens_out":13158,"duration_ms":146138,"concrete_test":"Implement Eq. (6) literally for O_t_l = 3 and B = 3: compute sum_{b=1}^3 2^b ((3 >> b) & 1). The result is 2, not 3. Then run the supplied or re-implemented model on CIFAR10 with the corrected bit positions b=0..B-1 versus the printed b=1..B, keeping D=5.11 and N=100; if the as-written version does not reproduce the reported 97.10% in Table 1, the experiments used an undocumented formula.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the exact binary reconstruction in Section 3.2. As written, Eq. (6) defines O_t_l = sum_{b=1}^B 2^b ((O_t_l >> b) & 1). This drops bit position 0 and weights the remaining positions one power of two too high. For any integer x, the right-hand side evaluates to 2*floor(x/2): for example, x=3 reconstructs as 2 and x=5 as 4. Equation (10) produces activations with step 1/N, so with N=100 all odd integer multiples of 0.01 are unreachable at inference. This is not a cosmetic indexing issue: exact integer reconstruction is precisely what justifies the exponential capacity gain and the AC-based energy comparison, and as printed it creates a training/inference mismatch for odd integer outputs. The reported accuracies imply the implementation used bit positions 0..B-1 or weights 2^{b-1}, but the paper does not say this and no code is provided. Until Eq. (6), and correspondingly Eq. (12), are corrected or shown to be a typo, the central mechanism cannot be reproduced from the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IBRA-LIF, a spiking neuron that trains with integer-valued activations produced by rounding and clipping a scaled membrane potential, and at inference converts the rounded integer into a binary spike expansion, using the linearity of convolution to perform accumulate-only operations with re-parameterized weights. A range alignment factor N is introduced to scale the membrane potential into the neuron's output range. The authors report state-of-the-art results on CIFAR10/100, ImageNet, CIFAR10-DVS, and COCO, including 74.19% top-1 accuracy on ImageNet with ResNet34, 49.2% mAP@50:95 on COCO, and 6.3x energy efficiency relative to the ANN baseline, together with near-lossless ANN-to-SNN conversion at one timestep.","tokens_in":12907,"tokens_out":10344,"duration_ms":99807,"significance":"The core idea is attractive: binary decomposition of integer activations can multiply the effective temporal information capacity while preserving spike-driven, accumulate-only inference, provided the binary reconstruction is exact. The linearity argument in Eq. (8) and the weight re-parameterization in Eqs. (13)-(15) are mathematically sound, and the reported empirical gains over existing SNN methods are large and consistent across classification and detection. If the implementation matches the corrected version of the method, this is a practically valuable contribution to high-performance low-energy SNNs. However, the manuscript as submitted contains an off-by-one error in the central binary reconstruction equation, leaves the neuron dynamics under-specified, and provides no code or statistical validation, so the paper is not currently reproducible from the text.","major_comments":[{"comment":"The binary expansion has an off-by-one indexing error. With O^{t,b}_l = (O^t_l >> b) & 1 and weights 2^b for b = 1..B, the reconstructed value is sum_{b=1}^B 2^b bit_b = 2*floor(O/2), not O. For example, O=3 reconstructs to 2 and O=5 to 4. Because Eq. (10) produces activations at multiples of 1/N, odd multiples of 1/N are unreachable at inference. This invalidates the exact reconstruction that justifies the exponential capacity gain and the AC-based inference in Eqs. (8)-(9). Please correct the summation to b=0..B-1 with weights 2^b, or b=1..B with weights 2^{b-1}, adjust the definition of B accordingly, and confirm that the reported experiments used the corrected convention.","section":"Sec. 3.2, Eq. (6) and Eq. (12)"},{"comment":"The hyperparameters N=100 and D=5.11 are selected on CIFAR10 and then transferred without modification to CIFAR100, ImageNet, CIFAR10-DVS, and COCO. The paper provides no per-layer membrane-potential range statistics and no theoretical bound relating the global factor N to the distribution of V_pre. If the activation range differs by architecture or dataset, the high binary bits will remain unused and the reported capacity gains will disappear. Please provide per-layer range statistics or a sensitivity analysis (for example N=10 and N=1000) on at least ImageNet and COCO.","section":"Sec. 4.2, Table 1; Secs. 4.3-4.5"},{"comment":"Several headline claims rest on differences that cannot be assessed without variance information, such as +0.35% over the ANN on ImageNet in Table 5 and +0.11% over the ANN on CIFAR100 ResNet19 in Table 2. All results are single runs with no error bars or multiple seeds, and no code is released. The need for verification is amplified by the inconsistency between Eq. (6) and the reported results. Please report mean and standard deviation over at least three seeds for the smaller datasets and clarify the code-release plan.","section":"Secs. 4.3-4.5, Tables 2, 5, and 6"},{"comment":"The manuscript never specifies the membrane-potential dynamics for IB-LIF or IBRA-LIF after the output becomes an integer, or an integer divided by N. Equation (3) is the reset rule for a binary LIF; for integer outputs it is unclear whether the reset subtracts O, O*N, or something else. Similarly, Eq. (16) defines a surrogate gradient on V but does not state how the round operation and the 1/N scaling in Eq. (10) are treated in the backward pass. These omissions leave the method under-specified and leave the training/inference consistency of the quantization step unanalyzed. Please add the full charging, firing, and resetting equations and the exact backward rule.","section":"Secs. 3.1-3.3, Eqs. (1)-(4), (10), and (16)"}],"minor_comments":[{"comment":"The table is difficult to read: the values in the D×N column appear to be the exponent B rather than the product, and entries such as \"102\" and \"1002\" are unreadable. Please regenerate the table with clear column headers and superscripts.","section":"Sec. 4.2, Table 1"},{"comment":"The abstract reports 49.1% mAP@50:95 on COCO, while Table 6 reports 49.2% for the largest IBRA-LIF model; please reconcile the two numbers.","section":"Abstract and Sec. 4.5, Table 6"},{"comment":"The tables report T×D, but the actual number of inference timesteps is T×B, where B is the number of bits. Reporting T×B as well would make the comparisons with prior work more transparent.","section":"Secs. 4.3-4.5, Tables 2-6"},{"comment":"The notation D_N is introduced but later written as \"DN\" without consistency; please define D_N once and use it uniformly.","section":"Sec. 3.3, Eq. (10)"},{"comment":"There are several typos and unfinished items: \"Emergy\" in Table 4, \"conprised\" in Sec. 3.3, and reference [41] contains the placeholder \"Accessed: insert date here.\"","section":"Throughout"},{"comment":"The accuracy column contains leading minus signs, for example \"−196.89%\", which should be removed.","section":"Sec. 4.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The off-by-one error in Eq. (6) is at the center of the proposed mechanism, and the manuscript as submitted is inconsistent with the reported results. Please ask the authors to provide either code/pseudocode or a corrected derivation, and to verify whether the energy and accuracy numbers change under the corrected indexing. The 6.3x energy-efficiency claim should also be clarified as being relative to the ANN baseline, since Table 4 shows IBRA-LIF at 0.44 mJ versus 0.15 mJ for I-LIF. Finally, Table 1 appears to have been produced from LaTeX with lost superscripts; the authors should be asked to resubmit a clean version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper has a real idea and unusually strong numbers, but you cannot reproduce the mechanism from the manuscript as printed. Eq. (6) defines the binary reconstruction with bits b=1..B and weights 2^b. That drops bit 0 and double-weights every included bit; for any integer x it reconstructs 2*floor(x/2). So odd activation values are unreachable at inference. That is a load-bearing detail—the exponential capacity claim and the energy comparison both rely on exact integer reconstruction. The reported accuracies imply the implementation used bits 0..B-1, but nothing in the paper says so. This is probably a typo, but it needs to be fixed before the method is citable as reproducible.\n\nWhat is genuinely new: replacing I-LIF's D unit spikes with log2(D) weighted binary spikes, and the range alignment scaling that lets the neuron actually use the high bits. The linearity argument in Eqs. (7)-(9) is correct, and the re-parameterization turning the RA MACs into ACs is a neat practical trick. The experimental results are consistent across ImageNet, CIFAR100, CIFAR10-DVS, and COCO, and the A2S near-lossless conversion at 1 timestep is a notable datapoint.\n\nSoft spots, in order. The equation issue above is the big one. Second, the range alignment N=100 is chosen by ablation on CIFAR10 and then silently transferred to all other datasets and architectures; no per-layer range statistics or generalization argument is given. Third, no code, no error bars, no multiple seeds—given the +3.45% claim over prior SOTA, that matters. Fourth, the tables list T×D=1×5.11, but the actual inference timestep count is T×B with B=9 for D×N=511. So the energy comparisons are hard to audit. These are addressable, not fatal; the central design is sound.\n\nWho this is for: anyone working on SNN representation capacity or energy-efficient inference. It deserves a serious referee; a good reviewer would ask for corrected equations, code, and per-task N justification, but the core contribution is worth engaging with.","headline":"The binary-encoding neuron idea is new and the results are strong, but a bit-indexing typo in Eq. (6) and flaky cost accounting mean the paper is not reproducible as printed.","tokens_in":13494,"tokens_out":2080,"would_cite":true,"duration_ms":20370,"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":"Spiking neurons that fire integer codes, converted to binary spikes at inference, match or beat the same-architecture ANN on ImageNet (74.19% vs 73.84%) at a computed 6.3× lower energy cost.","keywords":["spiking neural networks","integer firing","binary encoding","range alignment","information expression capacity","direct training","ANN-to-SNN conversion","energy-efficient inference"],"falsifier":"Run the trained ImageNet network and record, layer by layer, how often each of the 9 binary digits of $O^t_l \\times N$ is active. If the top bit-planes stay silent in most layers, the exponential capacity expansion is not actually being used, and a different explanation (such as plain integer quantization or range normalization) would need to account for the accuracy gain.","tokens_in":12470,"feed_emoji":"⚡","tokens_out":14207,"duration_ms":136051,"temperature":0.7,"pith_summary":"The paper sets out to close the accuracy gap between spiking neural networks and ordinary artificial neural networks without giving up spike-driven, accumulate-only inference. Its proposal, the IBRA-LIF neuron, fires positive integer values during training and then, at inference, decomposes each integer into binary bit-planes that behave like parallel 0/1 spike streams. Because convolution is linear, this decomposition is exact, and a single timestep carries exponentially more information than a single bit. A range-alignment scaling factor prevents the neurons from only ever firing small values, which would leave most of that capacity unused. The reported outcome is that directly trained SNNs of this kind beat same-architecture ANNs on ImageNet (74.19% versus 73.84%) and on CIFAR100, at a computed 6.3× lower energy cost.","feed_headline":"Integer-firing neurons beat their ANN twins at 6.3× lower energy","feed_subtitle":"On ImageNet a directly trained spiking ResNet34 hits 74.19% top-1, 0.35% above the same-architecture ANN.","key_machinery":"The load-bearing object is an exact identity for linear layers: $W \\sum_{b} 2^{b} O^{b} = \\sum_{b} 2^{b} W O^{b}$. It converts one integer-valued forward pass into $B$ parallel 0/1 spike streams whose weighted sum is bit-identical to the original, turning MAC-heavy convolutions into $B$ accumulate-only operations and giving exponential capacity growth for linear energy growth. Two supporting mechanisms carry the training side: the range-alignment factor $N$ that scales the membrane potential before rounding and clipping (with the division folded into the weights as $W^{N} = W/N$ at inference), and the straight-through surrogate gradient $\\varphi(V)=1$ for $0 \\le V \\le D$ that lets backpropagation pass through the non-differentiable integer firing.","core_discovery":"IBRA-LIF, on the paper's own terms, is a spiking neuron that replaces the binary firing step with an integer emission. During training the neuron outputs $O^{t} = \\mathrm{Clip}(\\mathrm{round}(V_{\\mathrm{pre}}^{t} \\times N),\\, 0,\\, DN)/N$, which keeps the information loss of direct training low; during inference the integer is rewritten in binary as $O^{t}_{l} = \\sum_{b=1}^{B} 2^{b}\\, O^{t,b}_{l}$ with $O^{t,b}_{l} = (O^{t}_{l} \\gg b)\\ \\&\\ 1$, so the $B$ bit-planes are ordinary 0/1 spike streams. Because convolution is linear, the next layer computes $\\sum_{b=1}^{B} 2^{b}\\, W^{N}_{l+1} O^{t,b}_{l}$ with a re-parameterized weight $W^{N} = W/N$: pure accumulate operations on spikes. The authors claim that this yields an exponentially larger theoretical expression capacity per unit energy than unary integer firing, and they report directly trained IBRA-LIF reaching 74.19% top-1 on ImageNet with ResNet34 — 0.35% above the same architecture's ANN — and 49.2% mAP@50:95 on COCO, at a computed 6.3× energy saving over the ANN, plus near-lossless ANN-to-SNN conversion at 73.04%.","pith_inferences":["A per-layer or learned scaling factor $N$ is a natural next step the paper does not run; layer-wise activation-range statistics could set $N_l$ where the single global $N=100$ leaves bit-planes silent, likely recovering more capacity in deep layers.","Because the binary decomposition relies only on linearity, the integer-plus-bit-plane trick transfers to other spike-driven architectures — Transformers, segmentation heads, or event-stream models — where it acts as an exact refinement of the temporal code rather than an approximation.","Ablating the training/inference asymmetry would localize the source of the gain: training forwards raw integers while inference uses binary reconstruction, and swapping one for the other would quantify how much of the improvement is true capacity expansion versus quantization side effects.","The paper itself notes that current hardware confines IBRA-LIF to synchronous neuromorphic chips; mapping the bit-plane expansion onto asynchronous hardware, where spike timing replaces a global clock, is an open engineering step that would decide how much of the theoretical energy saving survives in practice."],"forward_implications":["Directly trained spiking networks can match or beat the same-architecture ANN: +0.35% on ImageNet (74.19% vs 73.84%) and +0.11% on CIFAR100 (81.67% vs 81.56%).","ANN-to-SNN conversion becomes near-lossless and fast: 73.04% top-1 on ImageNet versus 71.12% for the best prior conversion baseline, which needed 64 timesteps.","Object detection improves to 66.2% mAP@50 and 49.2% mAP@50:95 on COCO, surpassing same-backbone prior SNN detectors and a similarly sized YOLOv5.","Estimated energy stays low: the N=100 model is computed at 6.3× the ANN's efficiency on the CIFAR10 energy analysis, and the N=10 model uses 63% of LIF's energy while improving accuracy.","Effective latency multiplies: a network run for T timesteps executes as T×B bit-plane steps, each an accumulate-only 0/1 convolution, so expression capacity grows exponentially with B at linear energy cost."],"supporting_citations":[{"why":"This is the I-LIF neuron the paper extends, and its SpikeYOLO detector is the main COCO baseline that IBRA-LIF outperforms.","marker":"[30]"},{"why":"Real Spike is the prior real-valued-spike method whose marginal gains motivate the move to integer firing.","marker":"[10]"},{"why":"Ternary Spike is the prior ternary-emission baseline that IBRA-LIF surpasses on ImageNet and CIFAR100.","marker":"[14]"},{"why":"This work supplies the straight-through surrogate gradient that makes direct training through integer firing possible.","marker":"[35]"},{"why":"SNNC-AP is the ANN-to-SNN conversion baseline that IBRA-LIF's 73.04% conversion result beats with far fewer timesteps.","marker":"[27]"},{"why":"YOLOv5 is the similarly sized ANN detector that IBRA-LIF surpasses on COCO mAP and in energy efficiency.","marker":"[41]"}],"fun_headline_variants":["Integer-fire SNN hits 74.19% on ImageNet, beats ANN twin","Spiking neurons match ANNs at 6.3x lower energy","Binary-range alignment makes SNNs outperform ANNs","SNN neuron with integer spikes: 74.19% ImageNet, 6.3x energy cut","New spiking neuron: ANN-beating accuracy at a fraction of power"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that one global scaling factor, N=100, tuned once on a small dataset, correctly matches the neuron's voltage range to its integer output range in every layer and every task; if the ranges drift in deeper or wider networks, the high part of the integer code never fires and the reported gains disappear.","fun_headline_variants_meta":{"raw":{"variants":["Integer-fire SNN hits 74.19% on ImageNet, beats ANN twin","Spiking neurons match ANNs at 6.3x lower energy","Binary-range alignment makes SNNs outperform ANNs","SNN neuron with integer spikes: 74.19% ImageNet, 6.3x energy cut","New spiking neuron: ANN-beating accuracy at a fraction of power"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1687,"prompt_tokens":1117,"completion_tokens":570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":733,"completion_tokens_details":{"reasoning_tokens":468}},"tokens_in":733,"tokens_out":570,"duration_ms":5637,"temperature":1.0,"reasoning_tokens":468,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:13:01.763510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained ImageNet network and record, layer by layer, how often each of the 9 binary digits of $O^t_l \\times N$ is active. If the top bit-planes stay silent in most layers, the exponential capacity expansion is not actually being used, and a different explanation (such as plain integer quantization or range normalization) would need to account for the accuracy gain.","supporting_citations":[{"cited_title":"Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection","cited_arxiv_id":null,"evidence_quote":"This is the I-LIF neuron the paper extends, and its SpikeYOLO detector is the main COCO baseline that IBRA-LIF outperforms."},{"cited_title":"Real spike: Learning real-valued spikes for spiking neural net- works","cited_arxiv_id":null,"evidence_quote":"Real Spike is the prior real-valued-spike method whose marginal gains motivate the move to integer firing."},{"cited_title":"Ternary spike: Learning ternary spikes for spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Ternary Spike is the prior ternary-emission baseline that IBRA-LIF surpasses on ImageNet and CIFAR100."},{"cited_title":"Enabling deep spiking neural net- works with hybrid conversion and spike timing dependent backpropagation","cited_arxiv_id":null,"evidence_quote":"This work supplies the straight-through surrogate gradient that makes direct training through integer firing possible."},{"cited_title":"A free lunch from ann: Towards efficient, accurate spiking neural networks calibration","cited_arxiv_id":null,"evidence_quote":"SNNC-AP is the ANN-to-SNN conversion baseline that IBRA-LIF's 73.04% conversion result beats with far fewer timesteps."},{"cited_title":"YOLOv5: A state-of-the-art real-time object de- tection system.https://docs.ultralytics.com,","cited_arxiv_id":null,"evidence_quote":"YOLOv5 is the similarly sized ANN detector that IBRA-LIF surpasses on COCO mAP and in energy efficiency."}],"review_version":1}