REACT: A Fully Spiking State-Space Model for Real-Time Event-Driven Temporal Perception
Geoffroy Keime, Nicolas Cuperlier, Benoit R. Cottereau
cs.RO · cs.AI
Quick verdict
Fully spiking SSM reads raw events one-by-one: 9.59% EvTTC error at 4.6 ms without bounding box; strong results but ablation inconsistencies need scrutiny.
Summary
The paper introduces REACT, a fully spiking state-space model that consumes event-camera streams one event at a time, without frames, voxel grids, or fixed time bins. Its core is C-SiLIF, a complex-valued spiking neuron whose discrete-time transition is computed from the physical inter-event interval, so the internal state evolves at the sensor's native timescale; no state reset is applied after spikes, keeping the recurrence linear. Training uses a parallel associative scan while deployment runs strictly sequentially. On DVS128-Gesture, REACT reaches 97.6% (97.2% INT8), matching event-by-event non-spiking SSMs with far smaller state-space blocks (0.04M vs 0.80M parameters). On EvTTC time-to-collision, without any bounding-box target prior, it attains 9.59±0.74% relative error with 4.6 ms end-to-end latency — within 0.15 points of the best learned method — versus 113 ms for Garl-TTC; zero-shot transfer to FCWD yields 25.7%, and anytime prediction is usable from 100 ms of stream. INT8 quantization cuts estimated energy from 18.5 to 2.8 mJ per 32,768 events. Limitations: energy figures are analytical 45 nm estimates with ~90% of cost in dense channel-mixing layers; preprocessing includes downsampling, a 10 ms refractory filter and 16× event decimation that qualify the 'raw events' claim; and an ablation reports 34.2% full-frame error, unexplained against the 9.59% headline. Code is promised upon publication.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- REACT processes raw event streams one event at a time with continuous-time spiking SSM dynamics driven by physical inter-event intervals, eliminating temporal accumulation windows.
- On EvTTC, without any bounding-box prior, REACT achieves 9.59±0.74% relative TTC error with 4.6 ms end-to-end latency, within 0.15 percentage points of the best learned method (Garl-TTC, 9.44%).
- The learned temporal representation transfers zero-shot to FCWD (25.7% eTTC) and supports anytime prediction usable from 100 ms of stream.
- The neuron ablation shows that removing the spike reset and adding complex oscillatory C-SiLIF dynamics are what enable temporal prediction (116.9% down to 9.3% eTTC in Table IV).
- INT8 quantization-aware training preserves accuracy while cutting estimated energy from 18.5 to 2.8 mJ per 32,768 events.
Strengths
Genuine event-by-event temporal processing
C-SiLIF's discretization is computed from the physical inter-event interval (Eq. 3-4), so state updates occur at the sensor's native timescale; end-to-end latency is 4.6 ms vs 113 ms for Garl-TTC (Table II), framed concretely as 4 cm vs 1 m of vehicle travel at the dataset's 9.22 m/s approach speed.
Informative neuron-model ablation
Table IV cleanly isolates contributions on EvTTC: LIF with reset 116.9% eTTC, no-reset LIF 84.1%, real-valued SiLIF 11.42%, complex C-SiLIF 9.3% — attributing gains to state preservation across spikes and oscillatory dynamics rather than the architecture alone.
Deployment-realistic evaluation design
Honest efficiency accounting
Careful task formulation and metrics
Weaknesses
Unexplained internal numeric inconsistencies
The raw-events, no-accumulation claim is qualified by preprocessing
Energy and efficiency advantages are estimates only
Statistical reporting is thin for the headline margins
Accuracy degrades where it matters operationally
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
REACT (arXiv 2609.19204) is a method paper in event-based neuromorphic vision. The central claim is architectural: a fully spiking state-space model can consume a raw event stream one event at a time — no frames, voxel grids, or fixed time bins — with internal dynamics driven by the physical inter-event interval Δt. The workhorse is the C-SiLIF complex-valued spiking neuron (Eq. 2-5, adopted from Fabre et al. [9] but re-driven by measured Δt instead of a fixed bin), with no post-spike state reset so the recurrence stays linear. Claims are tested on DVS128-Gesture classification and EvTTC time-to-collision regression (inverse-η parameterization, Gaussian NLL), plus zero-shot transfer to FCWD, anytime prediction, and INT8 quantization-aware training.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is REACT and how does it differ from other event-camera neural networks?
REACT is a spiking neural network that reads an event camera's output one event at a time, with no frames, voxel grids, or fixed time bins. Its complex-valued C-SiLIF neurons update their internal state using the actual time elapsed between events and are not reset after firing, so the model keeps a continuously evolving temporal state instead of integrating events into discrete snapshots.
How accurate is REACT at time-to-collision estimation on EvTTC?
On the EvTTC driving benchmark it reports a 9.59% ± 0.74% relative TTC error with 4.6 ms end-to-end latency, essentially matching the best prior learned method (Garl-TTC at 9.44% but 113 ms) even though REACT receives no bounding box and predicts from the full field of view. With INT8 quantization the error is 10.26%.
Does REACT run on neuromorphic hardware, and how much energy does it use?
Not measured on hardware yet. The paper reports analytical estimates using a 45 nm energy proxy: 18.5 mJ per 32,768 events in FP32, reduced to 2.8 mJ with INT8 quantization. About 90% of the estimated cost is in dense channel-mixing layers, and the authors state that hardware validation remains future work.
Is REACT's code or data publicly available?
The paper states that code will be made publicly available upon publication; it is not yet released. The evaluation datasets — DVS128-Gesture, EvTTC, and FCWD — are public benchmarks, and the architecture and training details are specified in the paper.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are available only via captions in the extracted text; the latency breakdown (Fig. 3) and qualitative prediction trajectories (Fig. 2) could not be verified.
- The hardware/software stack used for the 4.6 ms latency measurement and the precise definition of the held-out sequences in the Sec. IV-E.1 ablation are not stated in the extracted text.
- arXiv v1, not yet peer reviewed; no supplementary material and no released code, so all quantitative claims rest on the authors' reporting.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.