{"id":"56216efb-93b6-4931-aaff-81deb62e282d","arxiv_id":"2412.12525","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CREST combines a conjoint learning rule, a multi-scale spatiotemporal event integrator, and a spatiotemporal IoU loss to improve accuracy and estimated energy efficiency of event-based object detection with spiking neural networks.","lead":"CREST is a new framework for training spiking neural networks to detect objects in event-camera data, using a faster training trick and a multi-scale event encoder. If its energy estimates hold on real hardware, it could make low-power, high-speed vision viable for drones and edge devices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 100x energy-efficiency claim is an unmeasured consequence of Eqs. 25–26 charging LIF neurons a full decay MAC while charging FSN neurons zero decay; the headline ratio needs a hardware-aware sensitivity check.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern that I would flag: the energy-efficiency claim depends on Eqs. 24–26, which are analytic estimates rather than measurements, and the 100x figure is dominated by charging LIF neurons for decay MACs while charging FSN neurons nothing for decay. This is not a manufactured objection; the paper itself states that FSN-based SNNs have negligible decay costs (Eq. 26) and credits the authors' own STELLAR hardware paper, but no measured power data for CREST is provided. The accuracy contributions, including the CLR ablations, MESTOR ablations, and ST-IoU comparisons, are internally consistent and support the detection-quality part of the claim. The main risk is confined to the energy multiple, and the reader's CONDITIONAL verdict already captures that risk. I found no independent reason to move the verdict to ACCEPT or REJECT: the accuracy results are credible, the energy model is a legitimate but unverified hardware abstraction, and the missing sensitivity analysis can be supplied without changing the method. Hence UNCHANGED is the right verdict, with the concrete test above offering a way to decide whether the energy model survives closer scrutiny.","tokens_in":17144,"tokens_out":7039,"duration_ms":68454,"concrete_test":"Recompute the Gen1 table (Table 2) energy entries with a symmetric cost model: add a static per-neuron-per-timestep energy e_static to Eq. 26 for FSN-based CREST, and set the LIF decay term in Eq. 25 to 0.9 pJ (an AC cost) or to zero (analog leak). Sweep e_static from 0 to 1 pJ. If the CREST-to-EMS-34 energy ratio drops below roughly 20x under any of these settings, the 'up to 100x' claim should be softened; if it remains above 50x, the concern is non-load-bearing. A stronger version of this check is to run CREST on the STELLAR test chip with power instrumentation and compare it against an LIF-based detector on the same process.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of up to 100x energy efficiency rests on Eqs. 24–26. Eq. 25 charges every LIF-based SNN a full MAC (4.6 pJ) for membrane-potential decay on every non-spiking neuron-time step, while Eq. 26 assumes FSN-based SNNs have no decay cost at all. On Gen1, this asymmetry accounts for most of the gap: CREST DenseNet121-16 is listed at 3.48 mJ while EMS-34 is 650.13 mJ, of which roughly 624 mJ comes from the (1−fr) decay term. The claim is therefore not a measurement of CREST on FSN hardware but an arithmetic consequence of two very different cost models. The STELLAR citation supports the FSN design concept, but this manuscript reports no measured power data for CREST. If a real FSN implementation incurs non-negligible static, memory, or control energy, or if a LIF implementation realizes decay with a cheap analog leak instead of a 4.6 pJ MAC, the advertised 100x multiple can shrink by an order of magnitude. This does not affect the accuracy results, which are credible, but it makes the energy-efficiency headline conditional on an untested hardware model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CREST proposes a conjointly-trained spike-driven framework for event-based object detection and recognition. The method combines a surrogate DL-Net with discrete-level activations to approximate FSN-based spatiotemporal backpropagation, a multi-scale spatiotemporal event representation (MESTOR), and a spatiotemporal IoU loss (ST-IoU) that mixes a spiking-density term with CIoU. The paper reports state-of-the-art accuracy on NCARs, Gen1, and PKU-Vidar-DVS compared with prior SNN detectors, and claims up to 100x energy efficiency over SOTA SNNs on the basis of an analytical energy model.","tokens_in":17410,"tokens_out":3711,"duration_ms":34638,"significance":"If the accuracy results hold, CREST makes a useful contribution to event-based vision: it demonstrates that a surrogate-training scheme built around few-spikes neurons can train deep SNN detectors to competitive accuracy on three established benchmarks, and the ablations isolate the effects of the learning rule, the MESTOR representation, and the ST-IoU loss. The release of code and the use of multiple backbones and datasets are strengths. The central caveat is that the headline energy-efficiency claim is not a measurement: it follows from an analytical model that assigns full MAC costs to LIF decay and zero decay cost to FSN neurons, so the 100x number is contingent on hardware assumptions that are not validated in this manuscript. The accuracy results themselves are credible and do not depend on that assumption.","major_comments":[{"comment":"The 'up to 100x energy efficiency' claim is computed, not measured. Eq. (25) charges LIF neurons a 4.6 pJ MAC for every non-spiking neuron-time step, while Eq. (26) assigns FSN neurons zero decay cost, citing the authors' STELLAR hardware paper (Mao et al. 2024). This asymmetry drives most of the reported gap: on Gen1, CREST (DenseNet121-16) is listed at 3.48 mJ versus EMS-34 at 650.13 mJ, and the (1−fr) decay term in Eq. (25) contributes roughly 624 mJ of that total. If a real FSN implementation incurs non-negligible static, memory, or control energy, or if a LIF implementation realizes decay with a cheaper operation than a full 4.6 pJ MAC, the advertised multiple shrinks substantially. Please either report measured power on FSN hardware, or provide a sensitivity analysis that treats the LIF decay cost and FSN overhead as parameters; without that, the energy claim should be reworded as conditional on the STELLAR hardware model.","section":"Experimental Settings, Eqs. (24)-(26), and Table 2"},{"comment":"The final values of the ST-IoU weights a and b are never reported. The text and figure show experiments with (a,b) = (0.5,0.5), (0.3,0.7), and (0.1,0.9), but it is not stated which pair is used in Tables 1-3. Since ST-IoU is one of the two novel components claimed to improve detection accuracy, the missing values prevent reproduction and a fair comparison. Please report the exact weights used for each reported result.","section":"Eq. (23) and Figure 5(b)"},{"comment":"The MESTOR spatiotemporal-continuity stage uses 'fixed-value filters of size 3×3' in Eq. (15), but the filter values are never specified. These values directly determine which pixels are retained through the threshold condition in Eq. (17), and they are essential for reproducing MESTOR. Please provide the exact filter kernel (or state that it is random and fixed) in the main text or supplementary material.","section":"Eq. (15) and Eq. (17)"},{"comment":"All experimental results are single-run, with no error bars or multiple-seed reporting. Several accuracy gaps over prior work are modest (e.g., 0.952 vs 0.937 on NCARs in Table 5, and 0.360 vs 0.321 mAP50:95 on Gen1 in Table 2). Without variance estimates, the claim of 'superior recognition & detection performance' is not statistically supported. Please report mean ± std over at least three seeds for the main comparisons and for the key ablations.","section":"Tables 1-6"}],"minor_comments":[{"comment":"Spiking-IoU is defined as an absolute difference of spike densities, not as an intersection-over-union; the name is misleading and should be clarified in the text.","section":"Eq. (22)"},{"comment":"The header 'SFOD w/ FSN-BP(ours)' is ambiguous: it could mean the original SFOD method or the authors' reimplementation with FSN-BP. Please label the column more explicitly.","section":"Table 5"},{"comment":"There are several typos ('effcient' in the Conclusion, 'choosed' and 'Futhermore' in the Supplementary Material) that should be corrected.","section":"Conclusion and Supplementary D"},{"comment":"The figure reports energy and mAP for different time windows, but it does not indicate which K corresponds to the final default setting (K=5) of the experiments; adding a marker for the reported configuration would improve clarity.","section":"Figure 5(a)"}],"recommendation":"major_revision","confidential_remarks":"The accuracy contributions appear genuine and the paper is well positioned for the event-based vision community. The main concern for the editor is that the paper's most attention-grabbing claim, 100x energy efficiency, is derived from an energy model whose key assumption (zero FSN decay cost) is credited to the authors' own hardware paper and is not tested here. A sensitivity analysis or measured data would make the claim defensible. The missing hyperparameter values (ST-IoU weights, MESTOR filters) and the absence of error bars are fixable but need to be addressed before the paper can be considered complete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has three genuine contributions — the CLR training rule, MESTOR encoding, and ST-IoU loss — and the accuracy gains on three datasets look credible. The headline \"up to 100x energy efficiency\" does not survive scrutiny: it is the arithmetic output of an asymmetric analytical cost model, not a measurement, and should be heavily qualified.\n\nWhat's new: instead of backpropagating through spatiotemporal spikes, CLR trains a surrogate network with discrete-level activations and shares weights with the FSN-SNN. That is a sensible, practical trick, and the 3.4x training speedup plus improved deep-network accuracy (Table 5) is a concrete win. MESTOR's three-scale event integration with spatiotemporal continuity filtering is reasonable, and the ablation shows each piece helps. ST-IoU combining spike-density difference with CIoU is a simple idea that appears to improve mAP. Code is released, ablations are present, and external benchmarks are used, so the accuracy result is not circular.\n\nThe soft spot is the energy model. Eqs. 24-26 charge every non-spiking LIF timestep a full 4.6 pJ MAC for membrane decay, while charging FSN neurons zero decay cost. That asymmetry is responsible for most of the advertised 100x gap (e.g., Gen1: CREST DenseNet 3.48 mJ vs EMS-34 650.13 mJ, with roughly 624 mJ coming from the LIF decay term). No measured power data for CREST appears in the paper; the FSN property is inherited from the authors' own STELLAR hardware paper. If real hardware has static, memory, or control overhead, or if LIF leak can be done cheaply in analog, the multiple shrinks substantially. This does not damage the accuracy claims, but it means the energy headline is conditional, not established.\n\nMinor issues: no error bars on any table; final a,b weights for ST-IoU are not given (only a figure showing a=0.5,b=0.5, but the main experiments don't state which values were used); MESTOR filter values are unspecified; and Table 5 mixes self-implemented baselines, which is fine if the code is run by readers, but should be clearer.\n\nBottom line: a solid, honest engineering paper for the event-based SNN detection subfield. The training rule and representation are worth taking seriously. The energy claim needs either a hardware measurement or a sensitivity analysis under alternative cost assumptions before it can be quoted. Deserves a real peer review, but with the expectation that the energy section will be revised.","headline":"Real contributions in SNN training and event representation, but the 100x energy claim is a cost-model artifact that needs hardware validation.","tokens_in":17931,"tokens_out":2597,"would_cite":true,"duration_ms":24802,"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 conjointly-trained spike-driven framework called CREST claims state-of-the-art event-based object detection accuracy with up to 100x lower energy than prior spiking detectors.","keywords":["event-based vision","spiking neural networks","object detection","few-spikes neuron","surrogate gradient learning","energy efficiency","temporal coding","neuromorphic hardware"],"falsifier":"Measure the actual per-inference energy of CREST on an FSN-based neuromorphic processor (or a cycle-accurate simulator that includes membrane decay, memory access, and control logic) and compare it with a LIF-based SNN detector on the same task; if the measured ratio is far below 100x, the central energy claim is refuted. The accuracy claims can be checked independently by retraining the three components on the three public datasets.","tokens_in":16958,"feed_emoji":"⚡","tokens_out":9991,"duration_ms":75198,"temperature":0.7,"pith_summary":"CREST claims that a fully spike-driven neural network can surpass prior spiking detectors on event-based object recognition and detection while reducing estimated inference energy by up to two orders of magnitude. The framework combines a conjoint learning rule, which trains through a surrogate network with discrete-level activations instead of unrolling gradients over spike times, with a multi-scale spatiotemporal event integrator and a spatiotemporal IoU loss. On the NCARs, Gen1, and PKU-Vidar-DVS benchmarks, the paper reports accuracy above the SNN baselines it compares against, with the largest gains coming from the new training rule and the event representation. If the energy model holds, CREST would make high-accuracy event-based detection practical on low-power neuromorphic hardware.","feed_headline":"Conjoint training rule makes SNN detection 100x more efficient","feed_subtitle":"A surrogate discrete-level network plus multi-scale event integration lifts accuracy on three event-based datasets.","key_machinery":"The load-bearing mechanism is the conjoint learning rule. Its surrogate DL-Net defines activations by clipping the membrane potential to a bounded range and rounding to the nearest discrete level (Eqs. 9-10), then replaces the temporal accumulation over K timesteps with a single convolution (Eq. 11); backward passes approximate the round derivative as identity (Eq. 12), so error propagates only spatially. This removes the iterative spatiotemporal gradient of Eq. 4, cutting the time loop out of training and easing gradient vanishing. MESTOR supplies multi-scale spatiotemporal encoding: an FSN convolution layer accumulates fixed 3x3-filter responses over time bins and retains pixels whose spike count reaches K-1, yielding a spatiotemporal-continuous channel, combined with a spatial accumulation channel and a short-bin temporal channel. ST-IoU computes the absolute difference of spike density between predicted and ground-truth boxes and adds it to CIoU, so the loss is sensitive to the fact that objects produce denser spikes than background. FSN, the few-spikes neuron from prior work, provides the efficient temporal coding that makes layer outputs spike trains of K bits.","core_discovery":"The core discovery is that the expensive spatiotemporal backpropagation of SNNs can be replaced by a spatial surrogate gradient computed in a discrete-level network (DL-Net), without sacrificing accuracy on event-based detection. During forward propagation, the DL-Net clips and rounds membrane potentials to mimic FSN spike trains; during backward propagation, the round function is treated as an identity, so gradients flow through a single spatial pass rather than across K timesteps. This conjoint learning rule is complemented by MESTOR, which converts raw event streams into three channels—spatiotemporal-continuous events, spatial accumulation, and temporal short-bin maps—and by ST-IoU, which adds a spike-density IoU term to the standard CIoU coordinate loss. The paper reports that CREST outperforms the compared SNN detectors on all three datasets: 0.952 accuracy on NCARs, 0.360 mAP50:95 and 0.632 mAP50 on Gen1, and 0.432 mAP50 on PKU-Vidar-DVS.","pith_inferences":["The paper's energy numbers are analytical rather than measured; if real FSN chips incur non-negligible decay or memory-access overhead, the efficiency lead over LIF-based SNNs could shrink, although the accuracy results would stand.","The same conjoint-learning trick of treating a rounded activation as an identity in the backward pass could transfer to other temporal coding schemes beyond FSN, such as rank or latency coding, where unrolling over time is equally expensive.","MESTOR's filtering of time-continuous events is a general denoising step for event streams and might be reused for tracking, segmentation, or other event-based tasks that suffer from background and noise events.","Training energy on GPU/TPU is not included in the efficiency comparison; a full life-cycle accounting could alter the headline, but inference-side savings remain the relevant metric for edge deployment."],"forward_implications":["The conjoint learning rule reduces per-epoch training time by about 3.4x relative to FSN-BP on the studied settings, so directly-trained SNNs become feasible for larger detection models.","MESTOR's spatiotemporal-continuous filtering lowers firing rates while raising accuracy, meaning the same framework scales to longer event streams without proportional energy growth.","ST-IoU improves mAP over coordinate-only CIoU, showing that spike-density agreement is a reusable loss ingredient for event-based detectors.","Because the DL-Net forward pass and the FSN forward pass share weights, a model trained on GPU/TPU can be deployed on FSN hardware without re-training."],"supporting_citations":[{"why":"Supplies the few-spikes neuron (FSN) whose temporal coding is the foundation of CREST's spike-driven layers.","marker":"(Stöckl and Maass 2021)"},{"why":"Presents the STELLAR FSN hardware and the energy model that justifies the negligible-decay assumption in Eq. 26.","marker":"(Mao et al. 2024)"},{"why":"SFOD is the DenseNet-based SNN detector used as backbone and as the FSN-BP baseline in the conjoint-learning ablation.","marker":"(Fan et al. 2024)"},{"why":"EMS is the directly-trained SNN object detector that CREST outperforms on Gen1 in the main comparison.","marker":"(Su et al. 2023)"},{"why":"Provides the VoxelCube SNN baselines on NCARs and Gen1 whose energy accounting motivates the FSN approach.","marker":"(Cordone, Miramond, and Thierion 2022)"},{"why":"Supplies the per-operation energy constants (0.9 pJ per AC, 4.6 pJ per MAC) used in all energy estimates.","marker":"(Horowitz 2014)"},{"why":"Defines the NCARs recognition dataset and the HATS representation baseline.","marker":"(Sironi et al. 2018)"},{"why":"Defines the Gen1 automotive event-detection dataset used for the main detection evaluation.","marker":"(de Tournemire et al. 2020)"},{"why":"Defines the PKU-Vidar-DVS dataset used for evaluation under high-speed and low-light conditions.","marker":"(Li et al. 2022a)"}],"fun_headline_variants":["CREST: 100x energy-efficient SNN object detection","Conjoint training rule makes SNN detection 100x more efficient","CREST: conjointly trained spike-driven detector, 100x efficient","CREST: spike-driven detection with 100x energy savings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The energy-efficiency claim rests on the assumption that FSN-based hardware incurs negligible membrane-decay cost, so only spike-integration accumulations consume energy; if real chips pay significant decay, memory-access, or control overhead, the reported up-to-100x savings would shrink.","fun_headline_variants_meta":{"raw":{"variants":["CREST: 100x energy-efficient SNN object detection","Conjoint training rule makes SNN detection 100x more efficient","CREST: conjointly trained spike-driven detector, 100x efficient","CREST: spike-driven detection with 100x energy savings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00089,"raw_usage":{"total_tokens":3859,"prompt_tokens":983,"completion_tokens":2876,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":2801}},"tokens_in":599,"tokens_out":2876,"duration_ms":19866,"temperature":1.0,"reasoning_tokens":2801,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:59:28.595667+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual per-inference energy of CREST on an FSN-based neuromorphic processor (or a cycle-accurate simulator that includes membrane decay, memory access, and control logic) and compare it with a LIF-based SNN detector on the same task; if the measured ratio is far below 100x, the central energy claim is refuted. The accuracy claims can be checked independently by retraining the three components on the three public datasets.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents the STELLAR FSN hardware and the energy model that justifies the negligible-decay assumption in Eq. 26."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SFOD is the DenseNet-based SNN detector used as backbone and as the FSN-BP baseline in the conjoint-learning ablation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the NCARs recognition dataset and the HATS representation baseline."}],"review_version":1}