{"id":"a1db7b17-dc13-4f48-a24a-8ba2b010ac21","arxiv_id":"2507.07734","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A high-rate two-stream spiking network with a lightweight gated fusion unit achieves 94.9% on THU EACT-50 and enables early prediction within 100 ms.","lead":"EEvAct is a two-stream spiking neural network that recognizes human actions from event-camera data, reporting 94.9% top-1 accuracy on a 50-class benchmark with far fewer parameters and operations than the previous best model. The paper also proposes a standard protocol for measuring how early a model can recognize an action, and tests it on a tennis-serve triggering task.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2% SOTA margin rests on an unmatched comparison: EEvAct is scored at a fixed 2.0 s window while the EV-ACT baseline is scored on complete 2–5 s samples (Table I, fn. 2); the paper never reports a matched baseline or EEvAct's full-sample accuracy.","rationale":"The reader's weakest assumption identifies the unmatched baseline comparison, and I agree that is the load-bearing point. My reading of Table I confirms footnote 2 marks EV-ACT as complete samples while all EEvAct rows are evaluated at 2.0 s. This is exactly where the central claim could fail: the margin is an apples-to-oranges difference, and the paper supplies no number for EEvAct on complete samples. I considered whether the direction of the mismatch makes the claim conservative, since partial observation is often harder than full observation; however, that only suggests the reported gap might survive a matched run, not that it is established without one. The efficiency comparison is also muddled because effective accumulations are compared with MACs, but the accuracy mismatch is the primary concern. The architecture and early-recognition framework are plausible, and the concern is testable rather than an internal contradiction, so the existing CONDITIONAL verdict is appropriate and should remain unchanged. No evidence of fabrication or methodological fraud appears; the issue is a missing matched evaluation protocol. The concrete test is feasible if the original code or weights are available and should be required before the SOTA claim is accepted.","tokens_in":11341,"tokens_out":5185,"duration_ms":59061,"concrete_test":"Obtain the official EV-ACT checkpoint (or retrain with the [12] code) and run it on the same THU EACT-50 test split with the paper's preprocessing (center-crop 600x600, downsample to 100x100, 2 ms bins), reporting Top-1 at S=2.0 s from sequence start. Compare this matched baseline to the 94.9% EEvAct row. Also report EEvAct's accuracy on complete samples under the [12] protocol. If the matched gap is below 2 percentage points, or if EEvAct's complete-sample accuracy is not at least 2 points above the baseline, the SOTA claim should be weakened or reworded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim (Section I, Contribution 2; Table I) compares 94.9% at S=2.0 s for the two-stream SNN with 92.72% for the EV-ACT 3D CNN, which footnote 2 says is on complete samples of 2–5 s. These are not the same evaluation condition. The direction of the mismatch is not obviously favorable: partial observation at 2 s could be harder than full sequences, so if anything the gap might survive or grow under a matched baseline; but the claim as stated is unverified because no matched evaluation is reported and no full-sample accuracy is given for EEvAct. The headline 'outperforms by more than 2% in final accuracy' requires either (a) evaluating EV-ACT under the same 2 s window and preprocessing protocol, or (b) reporting EEvAct on the same complete-sample protocol as [12]. Without one of these, the 2% margin—the paper's main quantitative SOTA claim—is not established. In addition, the efficiency comparison (0.046 G ACs vs. 29.0 G MACs) mixes 'effective accumulations' with MACs, so the '<1% synaptic operations' claim also needs a matched cost metric to be interpretable, though this is secondary to the accuracy mismatch.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EEvAct, a two-stream spiking neural network for early event-based action recognition, and a formalized evaluation framework that reports accuracy at growing observation times together with effective synaptic operation counts. The architecture combines a common convolutional spiking layer, two parallel streams with trainable PLIF or adLIF neurons, an event-based gated fusion layer (EGRU or the proposed EGU), and a leaky-integrator readout with a combined loss. On THU EACT-50, the best variant is reported to reach 94.9% Top-1 accuracy at a 2.0 s observation, outperforming the EV-ACT 3D CNN baseline by 2.18 points while using 4.46M parameters and 0.046G effective ACs. The paper also reports about 50% Top-5 accuracy within 100 ms for its best early-recognition configuration and demonstrates the method on a tennis-serve triggering use case.","tokens_in":11638,"tokens_out":5074,"duration_ms":59415,"significance":"If the central comparison were sound, the paper would make a useful contribution: it provides a concrete early-recognition benchmark for event-based cameras, a practical architecture with several ablated components, and a real-world application that illustrates the value of high-rate spiking processing. I credit the authors for moving beyond final-accuracy-only reporting, for the large-scale evaluation, and for the transparent presentation of parameter counts and operation counts in Table I. However, the two headline claims—outperforming the prior SOTA by more than 2% and using less than 1% of the synaptic operations—both rest on comparisons that are not matched in evaluation condition or cost metric. The stress-test concern raised by the reader is valid: without a same-protocol baseline or a full-sample EEvAct accuracy, the SOTA claim is not established.","major_comments":[{"comment":"The headline SOTA claim is based on unmatched evaluation conditions. The 94.9% accuracy is reported for the two-stream SNN at a fixed S=2.0 s partial observation, whereas the 92.72% accuracy for EV-ACT 3D CNN [12] is explicitly reported for complete samples of 2–5 s duration. The paper does not provide either (a) EV-ACT evaluated under the same 2.0 s partial-observation protocol, or (b) EEvAct evaluated on the complete-sample protocol. The direction of the mismatch is not obvious, so without one of these matched evaluations the central 'outperforms by more than 2%' claim in the abstract, introduction, and conclusion is unverified.","section":"Section V, Table I (footnote 2); Section I, Contribution 2"},{"comment":"The 'less than 1% of synaptic operations' claim compares 0.046 G effective ACs for the SNN against 29.0 G MACs for the 3D CNN. Effective ACs (which skip zero or spike-free operations) and MACs (dense multiply-accumulates) are not commensurable without a precise counting rule that maps both to the same operation definition. Moreover, the AC count is for a 2.0 s partial observation while the MAC count is for complete 2–5 s samples, which compounds the mismatch. Please report a matched cost comparison, e.g., both models evaluated on the same observation length and both counted as MACs, or both as effective operations under an identical counting rule.","section":"Section V, Table I (efficiency columns)"},{"comment":"The paper never specifies the surrogate gradient used to train through the Heaviside spike function in Eq. (5), nor the event threshold in Eq. (7) for the EGU. This is essential for reproducibility of any SNN result. Relatedly, Table I reports single-run accuracies without standard deviations or number of seeds; given that the claimed margin over the baseline is just over 2 percentage points, seed-to-seed variation could change the conclusion. Please specify the surrogate, the training randomness, and report error bars over multiple seeds.","section":"Section IV-B, Eqs. (3)–(5); Section V-A"},{"comment":"There is a potential mismatch between training and evaluation data sampling: models are trained on randomly cropped 1000 ms windows taken from anywhere in the 2–5 s samples, whereas early-recognition evaluation starts from the beginning of each recording. The paper should clarify whether these training crops are aligned to action onsets or include arbitrary phases of the action. If they are not aligned, the early-recognition scores could be systematically affected, and the authors should discuss how the model transfers from randomly phased crops to action-onset evaluation.","section":"Section V-A (training and evaluation protocol)"}],"minor_comments":[{"comment":"The sentence 'The best early recognition model allows for 50% Top-5 accuracy already within 100ms' does not identify which configuration is plotted; please name the exact model/loss/readout combination (e.g., the TET/Last-membrane variant in Table I).","section":"Section V-C, Figure 7"},{"comment":"The symbol L_TET is typeset with inconsistent spacing and the acronym TET is not expanded; please clarify the notation and define TET when first used.","section":"Equation (9)"},{"comment":"The terms SynOps, MACs, and ACs are not defined in the caption; please provide a sentence defining effective accumulations and how they are counted.","section":"Table I caption"},{"comment":"The phrase 'outperforming previous work by 2%' should name the exact baseline and evaluation condition, since the percentage is not interpretable without the protocol.","section":"Abstract and Section I"},{"comment":"The SpikingBody dataset description gives only total train/test counts; for reproducibility, please provide per-class sample counts and describe the recording protocol (e.g., camera placement, action segmentation).","section":"Section V-D"}],"recommendation":"major_revision","confidential_remarks":"The main obstruction is the unmatched baseline comparison and the unmatched efficiency metric; both are fixable by additional experiments or by carefully restricting the claims. The lack of surrogate-gradient details and error bars strengthens the case for major revision rather than acceptance. I do not see a circularity issue; the accuracy numbers are empirical results against an external benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful part of this paper is the early-recognition evaluation framework and the two-stream SNN design. The authors formalize accuracy-over-observation-time for event-based action recognition, add SynOps reporting, and ablate time bin, single vs. two stream, PLIF vs. adLIF, EGRU vs. EGU, and loss/readout choices. That is a solid engineering study. The EGU is a reasonable lightweight alternative to EGRU, and the tennis trigger demo is a nice touch, though the dataset is small.\n\nThe problem is the headline accuracy claim. In Table I, EV-ACT's 92.72% is reported under footnote 2 as accuracy on complete samples, while EEvAct's 94.9% is at a fixed 2.0 s partial observation. That is not an apples-to-apples comparison. The authors never report EEvAct on full samples, nor EV-ACT at 2.0 s. It is possible the 2% margin survives or even grows, but the paper does not demonstrate it. The claimed 'outperforms by more than 2%' is not established as written. The efficiency comparison is similar: 0.046 G ACs vs. 29.0 G MACs mixes different cost metrics, so 'less than 1% of synaptic operations' is not interpretable as a fair comparison.\n\nOther soft spots are minor: no error bars or repeated-run variance, the surrogate gradient is never specified, and no code or model release. The ablations are otherwise convincing, and the early-recognition plots give a genuine sense of the trade-offs.\n\nThe paper is worth a serious referee. The benchmark protocol is reusable and the architecture is sensible, but the authors need to fix the comparison. I would ask for a matched evaluation in revision: either the same 2.0 s partial protocol for EV-ACT, using the same preprocessing, or full-sequence accuracy for EEvAct. Without that, the qualitative results stand but the SOTA claim should be reworded.\n\nFor your own work: cite it for the benchmark and the EGU, but not for the 2% claim.","headline":"Useful benchmark and architecture, but the headline 2% SOTA margin is not established because the baseline is scored on full sequences while the proposed model is scored at a fixed 2.0 s window.","tokens_in":12175,"tokens_out":2610,"would_cite":true,"duration_ms":28632,"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 two-stream spiking neural network hits 94.9% Top-1 accuracy on the THU EACT-50 event-action benchmark, beating the 3D CNN baseline by two points with a fifth of the parameters and under one percent of the synaptic operations.","keywords":["early action recognition","event-based vision","spiking neural networks","two-stream architecture","THU EACT-50","neuromorphic computing","gated recurrent units","synaptic operations"],"falsifier":"Run the baseline 3D CNN under the same fixed 2.0 second partial-observation protocol and the same test split; if its Top-1 accuracy reaches or exceeds the reported 94.9%, the paper's headline superiority claim would be overturned.","tokens_in":11170,"feed_emoji":"⚡","tokens_out":7441,"duration_ms":68712,"temperature":0.7,"pith_summary":"This paper tries to show that spiking neural networks can match or beat frame-based deep networks on large-scale event-based action recognition, while keeping the low latency and sparse computation that event cameras promise. It introduces a high-rate two-stream spiking architecture and evaluates it on the THU EACT-50 dataset, reporting 94.9% Top-1 accuracy at two seconds of observation, about two percentage points above the previous 3D CNN state of the art, with only a fifth of the parameters and under one percent of the synaptic operations. The paper also formalizes an evaluation framework that reports accuracy at growing observation times, and shows the model can reach roughly 50% Top-5 accuracy within the first 100 ms. A sympathetic reader would take the central claim to be that high-rate spiking processing, rather than frame accumulation, is now a practical route to early and accurate action recognition.","feed_headline":"Spiking two-stream net hits 94.9% on event actions","feed_subtitle":"Matches or beats frame-based 3D CNNs at a fraction of the compute, and reaches 50% Top-5 accuracy within 100 ms.","key_machinery":"The load-bearing mechanism is the two-stream spiking backbone plus a gated temporal fusion. The architecture splits a shared convolutional spiking layer into a ventral stream (stride-2, growing channels) and a dorsal stream (alternating strides, slower-growing channels), each built from convolutional layers with either parametric leaky integrate-and-fire (PLIF) or adaptive LIF (adLIF) neurons whose decay, feedback, and gating parameters are trainable per channel. The streams are concatenated and fused by an event-based gated recurrent unit (EGRU), or a lighter event-based gated unit (EGU) that drops state-dependent gates and needs about a third of the EGRU parameters; this fusion emits sparse real-valued events. A non-spiking leaky-integrator readout with trainable per-class decay accumulates the fused events, and the network is trained with a combined loss that averages the membrane-potential cross-entropy over time (LCEM) and samples the potential at multiple steps (LTET), which together are designed to balance early approximate predictions with final accuracy.","core_discovery":"The central claim is that a fully trainable two-stream spiking neural network—a shared spiking convolutional layer feeding separate ventral and dorsal streams, fused by an event-based gated recurrent unit and read out by leaky-integrator neurons—achieves state-of-the-art accuracy on THU EACT-50 while operating at high temporal rate. The reported 94.9% Top-1 accuracy at a fixed 2.0 s observation window outperforms the EV-ACT 3D CNN (92.72% on complete samples) by more than two points, while requiring about 20% of the parameters and under 1% of the synaptic operations. The paper further claims that combining a mean-potential cross-entropy loss with a temporally sampled TET loss, and reading out through leaky-integrator neurons, yields both strong early recognition and strong final accuracy.","pith_inferences":["A natural next check is to re-evaluate the EV-ACT 3D CNN under the same fixed 2.0 s observation window; the paper compares its number against a full-sequence baseline, so the true gap under matched conditions remains to be measured.","The high-rate processing idea could be pushed further by testing bin durations below 2 ms with hardware-in-the-loop, since the paper finds 1 ms bins degrade accuracy; the optimum may depend on event density and class dynamics.","The EGU's removal of state-dependent gates suggests a path toward even cheaper fusion, but the paper's own results show EGRU still wins on accuracy, so the trade-off between parameter savings and recognition quality is not yet settled.","The early Top-5 capability is promising for interactive systems, but the paper does not report Top-5 accuracy for the 3D CNN baseline; a matched comparison would show whether this early-hypothesis advantage is unique to SNNs or shared by frame-based approaches."],"forward_implications":["If the reported numbers hold, SNNs become a credible alternative to frame-accumulating 3D CNNs for event-camera action recognition, with about a fifth of the parameters and over two orders of magnitude fewer effective synaptic operations.","The formalized evaluation scheme (accuracy at fixed observation times, Top-K scores, effective SynOps) gives the field a common yardstick for early event-based recognition, replacing single final-accuracy comparisons.","The 50% Top-5 accuracy within 100 ms suggests that coarse action hypotheses can be formed within the first few frames' worth of events, which is directly useful for always-on monitoring and triggering systems.","The reported transfer to tennis-serve triggering in a motion-capture setting indicates the architecture can be applied to real-world streaming tasks without retraining on dense frames.","Because the model is designed for event-by-event-like processing with small time bins, it is positioned for asynchronous neuromorphic deployment, where sparse spikes directly map to hardware operations."],"supporting_citations":[{"why":"Supplies the THU EACT-50 dataset and the EV-ACT 3D CNN baseline whose full-sequence accuracy is the comparison target for the headline claim.","marker":"[12]"},{"why":"Provides the previous spiking-neural-network baseline (Motion SNN) on THU EACT-50, which the proposed two-stream architecture outperforms.","marker":"[29]"},{"why":"Supplies the EGRU event-based gated recurrent unit used in the fusion layer of the two-stream network.","marker":"[43]"},{"why":"Supplies the adaptive LIF (adLIF) neuron model used as one of the two spiking neuron choices in the feature extractor.","marker":"[41]"},{"why":"Supplies the parametric LIF (PLIF) neuron model used as the default spiking unit in the comparative study.","marker":"[39]"},{"why":"Supplies the TET loss (average cross-entropy over sampled membrane potentials) that the proposed combined loss builds on for early recognition.","marker":"[44]"},{"why":"Supplies the MinGRU simplification that the proposed lightweight event-based gated unit (EGU) is derived from.","marker":"[48]"},{"why":"Introduces the earlier two-stream spiking network idea for event-based action recognition that this architecture extends to large-scale datasets.","marker":"[38]"}],"fun_headline_variants":["Two-stream spiking net wins early event recognition","Event SNN hits 94.9% on THU EACT-50 actions","High-rate spiking net tops action recognition benchmark","SNN outdoes frame-based 3D CNN on event actions","Early action recognition leaps with two-stream SNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy claim assumes that the 3D CNN baseline's accuracy, measured on complete action samples of 2 to 5 seconds, can be fairly compared with the proposed model's accuracy measured after exactly 2.0 seconds of observation.","fun_headline_variants_meta":{"raw":{"variants":["Two-stream spiking net wins early event recognition","Event SNN hits 94.9% on THU EACT-50 actions","High-rate spiking net tops action recognition benchmark","SNN outdoes frame-based 3D CNN on event actions","Early action recognition leaps with two-stream SNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00113,"raw_usage":{"total_tokens":4677,"prompt_tokens":908,"completion_tokens":3769,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":3686}},"tokens_in":524,"tokens_out":3769,"duration_ms":27832,"temperature":1.0,"reasoning_tokens":3686,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:33:52.556751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the baseline 3D CNN under the same fixed 2.0 second partial-observation protocol and the same test split; if its Top-1 accuracy reaches or exceeds the reported 94.9%, the paper's headline superiority claim would be overturned.","supporting_citations":[{"cited_title":"Action Recognition and Benchmark Using Event Cam- eras,","cited_arxiv_id":null,"evidence_quote":"Supplies the THU EACT-50 dataset and the EV-ACT 3D CNN baseline whose full-sequence accuracy is the comparison target for the headline claim."},{"cited_title":"Event-based Action Recognition Using Motion Information and Spiking Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Provides the previous spiking-neural-network baseline (Motion SNN) on THU EACT-50, which the proposed two-stream architecture outperforms."},{"cited_title":"Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the parametric LIF (PLIF) neuron model used as the default spiking unit in the comparative study."},{"cited_title":"Temporal Contrastive Learning for Spiking Neural Networks","cited_arxiv_id":"2305.13909","evidence_quote":"Supplies the TET loss (average cross-entropy over sampled membrane potentials) that the proposed combined loss builds on for early recognition."},{"cited_title":"Two-Stream Spiking Neural Network for Event-based Action Recognition,","cited_arxiv_id":null,"evidence_quote":"Introduces the earlier two-stream spiking network idea for event-based action recognition that this architecture extends to large-scale datasets."}],"review_version":1}