Paper deep dive
VGER: Voxel-Guided Global Event Ranking for Event Cloud Attribution
Youxin Jiang, Baoheng Fu, Hongwei Ren, Xiangqian Wu
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Event cameras produce sparse and asynchronous event streams that provide rich spatio-temporal information for efficient perception. Recent advances in event-based models have demonstrated strong performance by directly modeling asynchronous events without dense frame reconstruction. However, identifying the event-level evidence behind their predictions is crucial for improving model transparency and reliability. Directly adapting point-level saliency methods from point clouds provides fine-grained attribution but overlooks event-specific spatio-temporal structures. To address this limitation, we propose Voxel-Guided Global Event Ranking (VGER), a training-free attribution framework for point-based event cloud networks. VGER combines event-level gradient evidence with task-aware voxel perturbation evidence, transferring regional contribution into event-level attribution scores while preserving fine-grained resolution. Furthermore, VGER introduces a unified event ranking strategy, where high-ranked events are expected to be prediction-critical and low-ranked events are expected to have limited influence on predictions. We evaluate VGER on three event-based benchmarks with PointNet, PointNet++, and EventMamba. Across nine dataset-backbone settings, VGER consistently improves both high-tail and low-tail deletion performance over point-level saliency baselines.
Tags
Links
- Source: https://arxiv.org/abs/2608.01470v1
- Canonical: https://arxiv.org/abs/2608.01470v1
Trouble viewing inline? Open PDF directly →
Full Text
44,697 characters extracted from source content.
Expand or collapse full text
VGER: Voxel-Guided Global Event Ranking for Event Cloud Attribution Youxin Jiang, Baoheng Fu, Hongwei Ren*, Xiangqian Wu Abstract Event cameras produce sparse and asynchronous event streams that provide rich spatio-temporal information for efficient perception. Recent advances in event-based models have demonstrated strong performance by directly modeling asynchronous events without dense frame reconstruction. However, identifying the event-level evidence behind their predictions is crucial for improving model transparency and reliability. Directly adapting point-level saliency methods from point clouds provides fine-grained attribution but overlooks event-specific spatio-temporal structures. To address this limitation, we propose Voxel-Guided Global Event Ranking (VGER), a training-free attribution framework for point-based event cloud networks. VGER combines event-level gradient evidence with task-aware voxel perturbation evidence, transferring regional contribution into event-level attribution scores while preserving fine-grained resolution. Furthermore, VGER introduces a unified event ranking strategy, where high-ranked events are expected to be prediction-critical and low-ranked events are expected to have limited influence on predictions. We evaluate VGER on three event-based benchmarks with PointNet, PointNet++, and EventMamba. Across nine dataset-backbone settings, VGER consistently improves both high-tail and low-tail deletion performance over point-level saliency baselines. Introduction Event cameras capture sparse and asynchronous event streams, providing high temporal resolution, low latency, and high dynamic range perception (Gallego et al. 2022). Unlike conventional cameras that rely on dense frame representations, event-based methods directly exploit asynchronous changes to preserve fine-grained temporal dynamics while reducing redundant computation (Gehrig et al. 2019; Sironi et al. 2018). Recently, point-based event cloud networks have emerged as an effective paradigm for event-based recognition by representing raw events as spatio-temporal point sets and processing them with point-based architectures (Qi et al. 2017a, b; Ren et al. 2025). These methods have demonstrated promising performance across various event-based benchmarks (Amir et al. 2017; Orchard et al. 2015; Li et al. 2017), benefiting from their ability to preserve the intrinsic sparsity and temporal characteristics of event streams. However, as event cloud networks become increasingly powerful, their predictions are still largely determined by complex interactions among massive asynchronous events, making it difficult to understand which events provide critical evidence for model decisions (Ribeiro et al. 2016; Lundberg and Lee 2017). Figure 1: Event-level saliency maps produced by VGER on representative streams from the three evaluation datasets. Colors encode normalized saliency (red: high, blue: low). High-ranked events follow compact spatio-temporal structures that contribute to the corresponding predictions. Attribution methods for point-based models provide a natural starting point for interpreting event cloud networks (Zheng et al. 2019). By assigning importance scores to individual points, point-level saliency approaches can identify prediction-related inputs with fine-grained spatial resolution (Simonyan et al. 2013). However, directly transferring these methods from conventional point clouds to event streams overlooks the unique characteristics of events. Unlike static 3D points, events are asynchronously generated, and their semantic information often emerges from local spatio-temporal structures formed by correlated event groups (Gallego et al. 2022). For example, motion patterns and object boundaries are represented by groups of temporally correlated events rather than isolated observations. Therefore, event attribution requires not only identifying individual important events but also capturing regional structural evidence that contributes collectively to model predictions (Fong and Vedaldi 2017; Petsiuk et al. 2018). This motivates a new attribution paradigm that bridges event-level precision and spatio-temporal structural reasoning. A straightforward solution is to introduce regional perturbation analysis, where groups of events are removed or modified to measure their influence on model predictions (Fong and Vedaldi 2017; Petsiuk et al. 2018). Voxel-based perturbation methods provide task-aware regional evidence by evaluating prediction changes caused by perturbing local spatio-temporal regions (Deng et al. 2022). However, assigning a single importance score to all events within the same voxel inevitably sacrifices event-level discrimination. Conversely, point-level saliency preserves individual event resolution but lacks awareness of collective regional contributions. These observations reveal a fundamental granularity dilemma in event attribution: fine-grained event importance and regional structural evidence are complementary but difficult to obtain simultaneously. To address this challenge, we propose Voxel-Guided Global Event Ranking (VGER), which combines voxel-level task evidence with event-level attribution signals to generate a more faithful global ranking of asynchronous events. VGER is a training-free attribution framework designed for point-based event cloud networks. It first derives task-aware regional evidence through voxel perturbation and transfers this evidence back to individual events. Meanwhile, event-level gradient attribution provides fine-grained sensitivity information (Simonyan et al. 2013; Zheng et al. 2019). By integrating these two complementary signals, VGER produces a unified event ranking that preserves both regional structural awareness and individual event resolution. Beyond conventional attribution evaluation that only examines whether important events affect predictions, we introduce a two-tail ranking perspective: highly ranked events should be prediction-critical, while low-ranked events should have limited influence (Petsiuk et al. 2018; Hooker et al. 2019). We evaluate VGER on three event-based benchmarks with multiple point-based architectures, including PointNet, PointNet++, and EventMamba (Qi et al. 2017a, b; Ren et al. 2025, 2024a). Extensive experiments demonstrate that VGER consistently improves attribution faithfulness over point-level saliency baselines across diverse dataset-backbone settings. The main contributions of this work are summarized as follows: • We introduce VGER, the first attribution framework for point-based event cloud networks, enabling fine-grained interpretation of asynchronous event streams. • We propose a voxel-guided evidence fusion strategy that combines task-aware regional perturbation evidence with event-level gradient attribution to preserve both structural context and event-level resolution. • We develop a global event ranking evaluation strategy with complementary high-tail and low-tail assessments, providing a more comprehensive evaluation of event attribution faithfulness. Figure 2: Overview of VGER. Event-level gradient evidence and task-aware voxel perturbation evidence are transferred, normalized, and fused into one global event ordering. The same ordering is evaluated at the high and low tails to distinguish prediction-critical events from events with limited influence. Related Work Event representations and recognition. Event representations range from dense time surfaces, learned grids, and recurrent surfaces to sparse recursive, graph-based, and voxel-based forms (Sironi et al. 2018; Gehrig et al. 2019; Cannici et al. 2020; Sekikawa et al. 2019; Schaefer et al. 2022; Deng et al. 2022; Ren et al. 2026a). EvT exploits sparsity by retaining activated event-derived patches, while GET groups events by timestamp and polarity (Sabater et al. 2022; Peng et al. 2023). E2PNet, PEPNet, EventMamba, and SECNet develop task-specific spatial and temporal processing for event data (Lin et al. 2023; Ren et al. 2024b, 2025, 2026b). General point-based architectures such as PointNet, PointNet++, DGCNN, and Point Transformer provide operations that can be applied to spatio-temporal event points (Qi et al. 2017a, b; Wang et al. 2019; Zhao et al. 2021). These developments expand how event structure is processed across event-based tasks; VGER complements them by identifying the event-level evidence behind a trained model’s prediction. Fine-grained and regional attribution. Gradient Saliency, Integrated Gradients, LRP, Grad-CAM, and SmoothGrad derive attribution from input derivatives or internal activations (Simonyan et al. 2013; Sundararajan et al. 2017; Bach et al. 2015; Selvaraju et al. 2017; Chattopadhay et al. 2018; Smilkov et al. 2017). LIME fits a local interpretable surrogate, while Kernel SHAP uses a weighted linear explanation model (Ribeiro et al. 2016; Lundberg and Lee 2017). Perturbation methods measure prediction changes under controlled input interventions (Zeiler and Fergus 2014; Fong and Vedaldi 2017; Petsiuk et al. 2018). For point clouds, PCS uses a differentiable center-shifting surrogate to score individual points (Zheng et al. 2019); later work explores features, local surrogates, semantic groups, interpretable architectures, and factorized attribution maps (Levi and Gilboa 2024; Ahmadi et al. 2024; Mulawade et al. 2025; Feng et al. 2024; Liu et al. 2024). EventPoint detects event keypoints for matching and registration (Huang et al. 2023), and gradient attribution has also been studied for spiking-network inputs (Bitar et al. 2023). The attribution methods in this group span fine-grained input evidence, group-level perturbation, feature-based explanations, and inherently interpretable architectures. In contrast, VGER transfers regional task evidence to individual events before globally ranking them, directly connecting the two granularities identified in the Introduction. Evaluation of explanations. Deletion and insertion curves evaluate an attribution ordering by tracking model behavior as inputs are removed or restored (Petsiuk et al. 2018). Their interpretation depends on a controlled perturbation protocol (Hooker et al. 2019; Tomsett et al. 2020; Gomez et al. 2022), while sanity checks, input-invariance and robustness tests, and infidelity/sensitivity metrics assess complementary properties of explanations (Adebayo et al. 2018; Kindermans et al. 2019; Alvarez-Melis and Jaakkola 2018; Yeh et al. 2019). We hold deletion cardinality and backbone-specific operators fixed across attribution methods and evaluate both ends of the same global ordering. This two-tail design tests whether high-ranked events are prediction-critical and low-ranked events have limited influence under one consistent protocol. Problem Formulation We consider the attribution problem for point-based event cloud networks. An event stream is represented as a set of asynchronous events: E=eii=1N,ei=(xi,yi,ti,pi),E=\e_i\_i=1^N, e_i=(x_i,y_i,t_i,p_i), (1) where (xi,yi,ti)(x_i,y_i,t_i) denotes the spatio-temporal coordinate and pip_i denotes the event polarity. A frozen event cloud model fθf_θ takes E as input and produces prediction logits, with ℒ(fθ(E),y)L(f_θ(E),y) denoting the task loss. Given an event stream E and its label y, the objective is to assign an importance score S(ei)S(e_i) to each event while keeping θ unchanged. Unlike conventional point attribution, event attribution should capture not only individual event importance but also the contribution of local spatio-temporal structures. For a deletion ratio r and K=⌊rN⌋K= rN , the top- and bottom-ranked event subsets are defined as: Dhigh(r) D_high(r) =ei∣S(ei)∈TopK(S,K), =\e_i S(e_i) (S,K)\, (2) Dlow(r) D_low(r) =ei∣S(ei)∈BottomK(S,K), =\e_i S(e_i) (S,K)\, where TopK and BottomK denote the sets of events with the largest and smallest attribution scores, respectively. Given a unified event ranking, Eq. (2) provides two complementary evaluation perspectives. Removing DhighD_high is expected to cause larger prediction degradation, while removing DlowD_low should have limited impact on recognition. The deletion operator D removes the selected events with a fixed cardinality and follows the input representation requirements of each backbone. For fair comparison, the same deletion procedure is applied to all attribution methods under each backbone. Voxel-Guided Global Event Ranking VGER addresses the attribution challenge through four stages. It first derives fine-grained sensitivity from event-level gradients and then measures the task contribution of local spatio-temporal voxels. The voxel evidence is transferred back to individual events and combined with a structural term that organizes their relative ranks. Finally, the fused event scores define one global ordering for both high-tail and low-tail evaluation. Event-Level Gradient Evidence Event-level gradients provide the fine-grained component of VGER. Because spatial coordinates and timestamps have different units and ranges, we normalize each zi=(xi,yi,ti)⊤z_i=(x_i,y_i,t_i) using a robust stream center c=(cx,cy,ct)⊤c=(c_x,c_y,c_t) and a positive coordinate-wise scale s=(sx,sy,st)⊤s=(s_x,s_y,s_t) : z¯i=(xi−cxsx,yi−cysy,ti−ctst)⊤. z_i= ( x_i-c_xs_x, y_i-c_ys_y, t_i-c_ts_t ) . (3) This normalization prevents the temporal coordinate from dominating the score solely through its numerical range. We adapt PCS to the spatio-temporal event domain by measuring task-loss sensitivity along a differentiable center-contraction path. To distinguish temporal displacement from spatial displacement, we introduce the anisotropic metric Aρ=diag(1,1,ρ),ρ=ωtωs,A_ρ=diag(1,1,ρ), ρ= _t _s, (4) where ρ controls the relative weight of temporal displacement. The corresponding anisotropic radius is riρ=z¯i⊤Aρz¯i.r_i^ρ= z_i A_ρ z_i. (5) We define the anisotropic center-contraction path as z¯i(τ)=exp(−τAρ)z¯i,τ≥0. z_i(τ)= (-τ A_ρ) z_i, τ≥ 0. (6) Its instantaneous displacement direction is ∂z¯i(τ)∂τ|τ=0=−Aρz¯i. . ∂ z_i(τ)∂τ |_τ=0=-A_ρ z_i. (7) Therefore, the anisotropic event-level PCS score is defined as Spcsρ(ei)=−⟨∇z¯iℒ,Aρz¯iriρ+ϵ⟩(riρ)1+α,S_pcs^ρ(e_i)=- _ z_iL, A_ρ z_ir_i^ρ+ε (r_i^ρ)^1+α, (8) where ℒL is the task loss, ϵε is a small constant, and α controls the radial scaling. Eq. (8) reduces to the original isotropic PCS formulation when ρ=1ρ=1. The score in Eq. (8) measures the first-order increase in task loss when an event is suppressed along the anisotropic spatio-temporal contraction direction. Adjusting ρ balances temporal and spatial displacement while preserving a separate score for every event. Task-Aware Voxel Evidence Event-level gradients distinguish individual events but do not explicitly measure the finite task effect of a complete local spatio-temporal region. To obtain this complementary regional evidence, we partition the normalized event domain into Gs×Gs×GtG_s× G_s× G_t voxels. Equivalently, voxelization can be performed in the anisotropically transformed coordinates ui=Aρ1/2z¯i,u_i=A_ρ^1/2 z_i, (9) so that the spatial and temporal extent of each voxel follows the metric in Eq. (4). Let VjV_j denote an occupied voxel. We construct E~(j) E^(j) by moving the events in VjV_j toward the robust stream center while preserving the number, polarity, and temporal order of events. The task-aware evidence of VjV_j is Δℒ(Vj)=ℒ(fθ(E~(j)),y)−ℒ(fθ(E),y), (V_j)=L(f_θ( E^(j)),y)-L(f_θ(E),y), (10) where fθf_θ is the frozen task network and y is the ground-truth label. This voxel-level loss change is broadcast to each constituent event: Svox(ei)=Δℒ(V(ei)).S_vox(e_i)= (V(e_i)). (11) Eq. (11) performs the transfer described in the Introduction. Rather than treating a voxel as the final attribution unit, VGER assigns its measured task contribution to each constituent event. The subsequent fusion therefore retains event-level resolution while incorporating evidence about the local structure surrounding each event. Structural Ranking Term The event-level gradient and voxel perturbation terms address the two sides of the granularity dilemma. The former distinguishes individual events within a local region, whereas the latter captures the finite task effect of the region as a whole. We first combine these complementary signals into a base event score: Sbase(ei)=βpNorm(Spcsρ(ei))+βvNorm(Svox(ei)),S_base(e_i)= _pNorm\! (S_pcs^ρ(e_i) )+ _vNorm\! (S_vox(e_i) ), (12) where Norm(⋅)Norm(·) denotes stream-wise score normalization, and βp _p and βv _v control the contributions of event-level gradient evidence and voxel-level task evidence, respectively. We then organize the base scores according to their local and temporal contexts. Let ℰV(ei) _V(e_i) =ek∈E:V(ek)=V(ei), = \e_k∈ E:V(e_k)=V(e_i) \, (13) ℰT(ei) _T(e_i) =ek∈E:T(ek)=T(ei), = \e_k∈ E:T(e_k)=T(e_i) \, where V(ei)V(e_i) denotes the voxel containing eie_i, and T(ei)T(e_i) denotes its temporal-bin index. Thus, ℰV(ei)E_V(e_i) contains the events in the same spatio-temporal voxel as eie_i, while ℰT(ei)E_T(e_i) contains the events assigned to the same temporal bin. For an event subset Q containing eie_i, let r(ei)r_Q(e_i) be the rank of Sbase(ei)S_base(e_i) among the events in Q, with larger base scores receiving larger ranks. We define the normalized within-set rank as ρ(ei)=r(ei)−1||−1,||>1,12,||=1. _Q(e_i)= cases r_Q(e_i)-1|Q|-1,&|Q|>1,\\[6.0pt] 12,&|Q|=1. cases (14) The singleton case is assigned the neutral value 1/21/2. The voxel-wise and temporal ranks are consequently given by ρV(ei)=ρℰV(ei)(ei),ρT(ei)=ρℰT(ei)(ei). _V(e_i)= _E_V(e_i)(e_i), _T(e_i)= _E_T(e_i)(e_i). (15) Both quantities lie in [0,1][0,1] and preserve the relative position of an event within its corresponding local context. We additionally characterize the occupancy of each voxel. Let occV_occ denote the set of occupied voxels and let n(Vj)n(V_j) be the number of events assigned to voxel VjV_j. We define n(Vj) n(V_j) =∑k=1N[V(ek)=Vj], = _k=1^NI\! [V(e_k)=V_j ], (16) D(Vj) D(V_j) =n(Vj)maxVℓ∈occn(Vℓ). = n(V_j) _V_ _occn(V_ ). Accordingly, D(Vj)∈(0,1]D(V_j)∈(0,1] for every occupied voxel. A larger value indicates a region containing more events, whereas 1−D(Vj)1-D(V_j) represents its relative sparsity within the current event stream. Using these contextual quantities, we define the structural ranking score as Sstruct(ei)=wbSbase(ei)+wcρV(ei)+wtρT(ei)+wr(1−D(V(ei)))−wdD(V(ei)). splitS_struct(e_i)&=w_bS_base(e_i)+w_c _V(e_i)+w_t _T(e_i)\\ &+w_r (1-D(V(e_i)) )-w_dD(V(e_i)). split (17) Here, wbw_b retains the underlying attribution evidence, wcw_c and wtw_t control the voxel-wise and temporal rank contributions, and wrw_r and wdw_d calibrate the relative sparsity and occupancy of the event’s voxel. The two occupancy components are written separately to expose their structural interpretations, while their combined ranking effect is governed by wr+wdw_r+w_d. The structural term therefore refines the global event ordering without replacing the gradient and voxel attribution signals. The within-voxel rank preserves fine-grained comparisons among events receiving the same regional evidence, the temporal rank incorporates their relative temporal context, and the occupancy terms account for how events are distributed across spatio-temporal regions. Voxel-Guided Global Event Ranking The final VGER score fuses regional, event-level, and structural evidence into one scalar for every event: SVGER(ei)= S_VGER(e_i)= λvNorm(Svox(ei))+λpNorm(Spcsρ(ei)) _vNorm(S_vox(e_i))+ _pNorm(S_pcs^ρ(e_i)) +λrNorm(Sstruct(ei)). + _rNorm(S_struct(e_i)). (18) Here, Norm(⋅)Norm(·) denotes stream-wise score normalization, and λv _v, λp _p, and λr _r control the contributions of voxel-level, gradient-level, and structural evidence, respectively. VGER then performs a stable global ranking over the complete stream: π=argsortei∈ESVGER(ei).π=argsort_e_i∈ ES_VGER(e_i). (19) Experiments Datasets, Backbones, and Baselines We evaluate DVS Gesture (Amir et al. 2017), N-MNIST, and N-Caltech101 (Orchard et al. 2015), the standard event-classification benchmarks alongside CIFAR10-DVS (Li et al. 2017). The backbones are PointNet (Qi et al. 2017a), PointNet++ (Qi et al. 2017b), and EventMamba (Ren et al. 2025). The verified baselines shared by the archived 3×33× 3 matrix are Random and PCS-xyt. A separate PointNet study compares Basic Voxel with PCS-xyt on all three datasets. Random uses the archived event permutations and provides a task-agnostic deletion reference. PCS-xyt applies Eq. (8) when ρ=1ρ=1 without voxel evidence. Basic Voxel assigns the same task-loss change to all events in one voxel; its ties are resolved at event level while respecting the exact deletion budget. We restrict the quantitative comparison to these baselines because they are the methods represented by verified, common-format curves in the supplied archive. The PointNet, PointNet++ and EventMamba results use the full archived test counts: 6,623 for DVS Gesture, 10,000 for N-MNIST, and 17,869 for N-Caltech101. Two-Tail Protocol All settings are compared over ℛ=0,0.05,…,1.00.R=\0,0.05,…,1.00\. At each ratio, exactly ⌊rN⌋ rN events are deleted. Accuracy curves are summarized by trapezoidal AUC; lower is better for the high tail and higher is better for the low tail: AUC(A)=∑k=120A(rk)+A(rk+1)2(rk+1−rk).AUC(A)= _k=1^20 A(r_k)+A(r_k+1)2(r_k+1-r_k). We report exact ratio-wise wins for every setting. The two summaries answer different questions: AUC weights the magnitude of the curve difference at each ratio, whereas a win count gives every sampled ratio one vote, so a method cannot dominate either summary through a single strong region of the protocol. Method High AUC ↓ Low AUC ↑ Random 0.4977 0.4977 PCS-xyt 0.4368 0.4934 VGER 0.3836 0.5404 Table 1: VGER improves both macro AUC tails over PCS-xyt. Values are unweighted averages over nine archived settings. Backbone Dataset n VGER-H↓ PCS-H↓ VGER-L↑ PCS-L↑ PointNet DVS Gesture 6,623 0.5532 0.6351 0.8121 0.7781 PointNet N-MNIST 10,000 0.5959 0.6422 0.7339 0.6971 PointNet N-Caltech101 17,869 0.0576 0.0756 0.2665 0.2180 PointNet++ DVS Gesture 6,623 0.4665 0.5311 0.6035 0.5315 PointNet++ N-MNIST 10,000 0.2655 0.2740 0.3811 0.3489 PointNet++ N-Caltech101 17,869 0.0799 0.0817 0.0981 0.0918 EventMamba DVS Gesture 6,623 0.4429 0.5765 0.7369 0.6290 EventMamba N-MNIST 10,000 0.7729 0.8616 0.8935 0.8491 EventMamba N-Caltech101 17,869 0.2180 0.2533 0.3381 0.2972 Table 2: VGER wins both AUC tails in all nine settings. H is high-tail deletion and L is low-tail deletion; bold indicates the better AUC within each pair. Main Two-Tail Fidelity Figure 3: Two-tail deletion curves for the 3×33× 3 matrix of backbones and datasets. Each panel deletes 5%5\%–90%90\% of events in score order and reports accuracy. VGER’s high tail (VGER-H) drops below PCS-xyt-H and Random in most settings and ratios, and its low tail (VGER-L) stays above PCS-xyt-L in most settings and ratios, so the two VGER tails separate decision-critical from decision-irrelevant events across the whole protocol. Table 1 reports the macro comparison across the archived matrix.The macro comparison shows that VGER reduces high-tail AUC by 0.0532 and increases low-tail AUC by 0.0470 relative to PCS-xyt. Its macro two-tail gap, low AUC minus high AUC, is 0.1568 compared with 0.0566 for PCS-xyt. These averages weight each dataset–backbone setting equally. The rows behind them are uniformly directional and are all produced by frozen archived configurations, so the macro gains are not carried by any single backbone or dataset.Table 2 gives the corresponding setting-level AUCs. Figure 3 presents the main experiment of this paper: the complete two-tail deletion curves behind these counts. In most settings and deletion ratios, removing VGER’s top-ranked events damages accuracy earlier and deeper than removing PCS-xyt’s or Random’s, whereas removing its bottom-ranked events preserves accuracy longer than PCS-xyt’s. The appendix further averages the two-tail curves over all nine settings and condenses each setting into its two-tail separation, which VGER widens over PCS-xyt in all nine settings, showing that one global ordering simultaneously isolates decision-critical and decision-irrelevant events. Additional implementation details and supporting evidence are provided in the appendix, including the setting-specific fusion weights used in part of the experiments, the full numerical data underlying Figure 3, and the auxiliary results from hyperparameter search and related diagnostic studies. These materials are included to make the reported two-tail fidelity results traceable beyond the summarized AUC values and to facilitate independent verification of the experimental protocol. Results on PointNet PointNet uses single max-pooling, making predictions depend on a compact critical event subset. High-tail deletion therefore tests whether VGER’s top-ranked events are truly decisive. VGER lowers high-tail AUC from 0.6351 to 0.5532 on DVS Gesture and from 0.0756 to 0.0576 on N-Caltech101. Low-tail AUC also improves over PCS-xyt on all datasets, with gains of 0.0340, 0.0368, and 0.0485. This consistent two-tail behavior shows reliable importance estimates for globally pooled representations. Results on PointNet++ PointNet++ learns hierarchical features from nested spatial neighborhoods, so attribution depends on both local geometry and global structure. Although PCS-xyt already has low high-tail AUCs, VGER improves both tails on every dataset. On DVS Gesture, high-tail AUC drops from 0.5311 to 0.4665, while low-tail AUC rises from 0.5315 to 0.6035. Figure 3 further shows that top-ranked deletion disrupts recognition faster, whereas bottom-ranked deletion preserves accuracy longer. The smaller gains on N-MNIST and N-Caltech101 remain consistent. Results on EventMamba EventMamba models long-range temporal dependencies through selective state updates, making event-level attribution less direct than in point-wise networks. Still, VGER improves both tails in all three archived runs. The largest gain appears on DVS Gesture, where high-tail AUC decreases from 0.5765 to 0.4429 and low-tail AUC increases from 0.6290 to 0.7369. N-MNIST and N-Caltech101 show the same pattern: removing highly ranked events causes sharper degradation, while removing low-ranked events better preserves accuracy. Ablation Study We ablate the key evidence channels on the deterministic PointNet++/N-MNIST development subset using the same 5%–90% deletion protocol. Figure 4 compares four representative rankings: PCS-xyt, PCS-xyt with voxel evidence, the full voxel-structural fusion, and Random. The top panel reports high-tail deletion, where lower accuracy indicates that the removed events are more decision-critical. The bottom panel reports low-tail deletion, where higher accuracy indicates that the removed events are less important. Figure 4: Component ablation on the PointNet++/N-MNIST development subset. Top: high-tail deletion AUC (lower is better). Bottom: low-tail deletion AUC (higher is better). The high-tail results show that voxel evidence is the main source of decision-critical localization. Compared with PCS-xyt, adding the voxel channel causes accuracy to drop earlier across most deletion ratios. The full fusion, which further incorporates structural ranking, remains among the strongest high-tail curves and stays well below Random, indicating that it preserves the ability to identify influential events. The low-tail results reveal the complementary role of the structural component. While voxel evidence alone improves high-tail deletion, the full voxel-structural fusion gives the best low-tail preservation over most ratios, keeping accuracy higher than PCS-xyt, voxel-only, and Random before the curves converge at large deletion ratios. Together, the two panels show that voxel perturbation sharpens the head of the ranking, while structural fusion improves the tail. This supports the design of VGER: combining regional task evidence with global structural ordering yields a ranking that is effective for both explanation and safe event removal. Representative Prediction Changes Figure 5: Representative prediction changes after deleting ranked events under VGER. Red points denote deleted events; green/red text marks correct/incorrect predictions. Figure 5 shows representative prediction changes caused by deleting events selected by VGER. Each row contains the original event cloud, the deleted events highlighted in red, and the retained event cloud after deletion. Green labels indicate correct predictions, while red labels indicate incorrect predictions. The first two rows show failure-inducing high-tail deletions. Removing a small set of highly ranked events changes originally correct predictions into incorrect ones, from digit 5 to digit 9 and from face to lamp. These cases suggest that the removed events contain decision-critical evidence for the original class. The last two rows show the complementary behavior. For samples that are initially misclassified, deleting ranked events suppresses misleading evidence and recovers the correct class, changing digit 3 to digit 2 and stop sign to face. Overall, these examples connect the aggregate two-tail deletion curves with concrete event-level behavior: the ranking identifies influential events whose removal can alter or even correct the model decision. Conclusion We presented VGER, a training-free method that fuses event-level gradient evidence with task-aware voxel perturbations to produce a global event ranking for two-tail deletion. Across nine archived dataset–backbone settings, VGER improves both tails over PCS-xyt in every case. Its head identifies decision-critical events, while its tail identifies removable events that preserve accuracy, so one ranking supports both explanation and compression. The two-tail deletion protocol is architecture-agnostic and reproducible from the archived evidence trail. Beyond individual model interpretation, event cloud attribution provides a general tool for analyzing how asynchronous event streams contribute to perception decisions. Such event-level evidence may further support future research in architecture design, adversarial robustness, event selection, and efficient event processing. References J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim (2018) Sanity checks for saliency maps. Advances in neural information processing systems 31. Cited by: Evaluation of explanations.. S. M. Ahmadi, K. Aslansefat, R. Valcarce-Diñeiro, and J. Barnfather (2024) Explainability of point cloud neural networks using smile: statistical model-agnostic interpretability with local explanations. arXiv preprint arXiv:2410.15374. Cited by: Fine-grained and regional attribution.. D. Alvarez-Melis and T. S. Jaakkola (2018) On the robustness of interpretability methods. arXiv preprint arXiv:1806.08049. Cited by: Evaluation of explanations.. A. Amir, B. Taba, D. Berg, T. Melano, J. McKinstry, C. Di Nolfo, T. Nayak, A. Andreopoulos, G. Garreau, M. Mendoza, et al. (2017) A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 7243–7252. Cited by: Introduction, Datasets, Backbones, and Baselines. S. Bach, A. Binder, G. Montavon, F. Klauschen, K. Müller, and W. Samek (2015) On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10 (7), p. e0130140. Cited by: Fine-grained and regional attribution.. A. Bitar, R. Rosales, and M. Paulitsch (2023) Gradient-based feature-attribution explainability methods for spiking neural networks. Frontiers in Neuroscience 17, p. 1153999. Cited by: Fine-grained and regional attribution.. M. Cannici, M. Ciccone, A. Romanoni, and M. Matteucci (2020) A differentiable recurrent surface for asynchronous event-based data. In European Conference on Computer Vision, p. 136–152. Cited by: Event representations and recognition.. A. Chattopadhay, A. Sarkar, P. Howlader, and V. N. Balasubramanian (2018) Grad-cam++: generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), p. 839–847. Cited by: Fine-grained and regional attribution.. Y. Deng, H. Chen, H. Liu, and Y. Li (2022) A voxel graph cnn for object classification with event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 1172–1181. Cited by: Introduction, Event representations and recognition.. T. Feng, R. Quan, X. Wang, W. Wang, and Y. Yang (2024) Interpretable3d: an ad-hoc interpretable classifier for 3d point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, p. 1761–1769. Cited by: Fine-grained and regional attribution.. R. C. Fong and A. Vedaldi (2017) Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, p. 3429–3437. Cited by: Introduction, Introduction, Fine-grained and regional attribution.. G. Gallego, T. Delbrück, G. M. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis, et al. (2022) Event-based vision: a survey. IEEE transactions on pattern analysis and machine intelligence 44 (1), p. 154–180. Cited by: Introduction, Introduction. D. Gehrig, A. Loquercio, K. G. Derpanis, and D. Scaramuzza (2019) End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF international conference on computer vision, p. 5633–5643. Cited by: Introduction, Event representations and recognition.. T. Gomez, T. Fréour, and H. Mouchère (2022) Metrics for saliency map evaluation of deep learning explanation methods. In International Conference on Pattern Recognition and Artificial Intelligence, p. 84–95. Cited by: Evaluation of explanations.. S. Hooker, D. Erhan, P. Kindermans, and B. Kim (2019) A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems 32. Cited by: Introduction, Evaluation of explanations.. Z. Huang, L. Sun, C. Zhao, S. Li, and S. Su (2023) Eventpoint: self-supervised interest point detection and description for event-based camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 5396–5405. Cited by: Fine-grained and regional attribution.. P. Kindermans, S. Hooker, J. Adebayo, M. Alber, K. T. Schüt, S. Dähne, D. Erhan, and B. Kim (2019) The (un) reliability of saliency methods. In Explainable AI: Interpreting, explaining and visualizing deep learning, p. 267–280. Cited by: Evaluation of explanations.. M. Y. Levi and G. Gilboa (2024) Fast and simple explainability for point cloud networks. arXiv preprint arXiv:2403.07706. Cited by: Fine-grained and regional attribution.. H. Li, H. Liu, X. Ji, G. Li, and L. Shi (2017) Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience 11, p. 244131. Cited by: Introduction, Datasets, Backbones, and Baselines. X. Lin, C. Qiu, S. Shen, Y. Zang, W. Liu, X. Bian, M. Müller, C. Wang, et al. (2023) E2pnet: event to point cloud registration with spatio-temporal representation learning. Advances in Neural Information Processing Systems 36, p. 18076–18089. Cited by: Event representations and recognition.. S. Liu, B. Li, Z. Fang, M. Cui, and K. Huang (2024) FFAM: feature factorization activation map for explanation of 3d detectors. Advances in Neural Information Processing Systems 37, p. 80363–80382. Cited by: Fine-grained and regional attribution.. S. M. Lundberg and S. Lee (2017) A unified approach to interpreting model predictions. Advances in neural information processing systems 30. Cited by: Introduction, Fine-grained and regional attribution.. R. N. Mulawade, C. Garth, and A. Wiebel (2025) XAI for point cloud data using perturbations based on meaningful segmentation. IEEE Access. Cited by: Fine-grained and regional attribution.. G. Orchard, A. Jayawant, G. K. Cohen, and N. Thakor (2015) Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience 9, p. 437. Cited by: Introduction, Datasets, Backbones, and Baselines. Y. Peng, Y. Zhang, Z. Xiong, X. Sun, and F. Wu (2023) Get: group event transformer for event-based vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 6038–6048. Cited by: Event representations and recognition.. V. Petsiuk, A. Das, and K. Saenko (2018) Rise: randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421. Cited by: Introduction, Introduction, Introduction, Fine-grained and regional attribution., Evaluation of explanations.. C. R. Qi, H. Su, K. Mo, and L. J. Guibas (2017a) Pointnet: deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 652–660. Cited by: Introduction, Introduction, Event representations and recognition., Datasets, Backbones, and Baselines. C. R. Qi, L. Yi, H. Su, and L. J. Guibas (2017b) Pointnet++: deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30. Cited by: Introduction, Introduction, Event representations and recognition., Datasets, Backbones, and Baselines. H. Ren, Y. Jiang, T. Huang, and X. Wu (2026a) A systematic survey on event camera representation learning. arXiv preprint arXiv:2606.23078. Cited by: Event representations and recognition.. H. Ren, F. Ma, Y. Fang, H. Huang, Y. Zhou, Y. Huang, H. FU, Z. Yang, Y. Jiang, X. Wu, et al. (2026b) Scalable event cloud network for event-based classification. In Forty-third International Conference on Machine Learning, Cited by: Event representations and recognition.. H. Ren, Y. Zhou, X. Lin, Y. Huang, H. Fu, J. Song, and B. Cheng (2024a) Spikepoint: an efficient point-based spiking neural network for event cameras action recognition. In International Conference on Learning Representations, Vol. 2024, p. 27827–27846. Cited by: Introduction. H. Ren, Y. Zhou, J. Zhu, X. Lin, H. Fu, Y. Huang, Y. Fang, F. Ma, H. Yu, and B. Cheng (2025) Rethinking efficient and effective point-based networks for event camera classification and regression. IEEE Transactions on Pattern Analysis and Machine Intelligence. Cited by: Introduction, Introduction, Event representations and recognition., Datasets, Backbones, and Baselines. H. Ren, J. Zhu, Y. Zhou, H. Fu, Y. Huang, and B. Cheng (2024b) A simple and effective point-based network for event camera 6-dofs pose relocalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 18112–18121. Cited by: Event representations and recognition.. M. T. Ribeiro, S. Singh, and C. Guestrin (2016) " Why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, p. 1135–1144. Cited by: Introduction, Fine-grained and regional attribution.. A. Sabater, L. Montesano, and A. C. Murillo (2022) Event transformer. a sparse-aware solution for efficient event data processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 2677–2686. Cited by: Event representations and recognition.. S. Schaefer, D. Gehrig, and D. Scaramuzza (2022) Aegnn: asynchronous event-based graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 12371–12381. Cited by: Event representations and recognition.. Y. Sekikawa, K. Hara, and H. Saito (2019) Eventnet: asynchronous recursive event processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 3887–3896. Cited by: Event representations and recognition.. R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra (2017) Grad-cam: visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, p. 618–626. Cited by: Fine-grained and regional attribution.. K. Simonyan, A. Vedaldi, and A. Zisserman (2013) Deep inside convolutional networks: visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034. Cited by: Introduction, Introduction, Fine-grained and regional attribution.. A. Sironi, M. Brambilla, N. Bourdis, X. Lagorce, and R. Benosman (2018) HATS: histograms of averaged time surfaces for robust event-based object classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 1731–1740. Cited by: Introduction, Event representations and recognition.. D. Smilkov, N. Thorat, B. Kim, F. Viégas, and M. Wattenberg (2017) Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825. Cited by: Fine-grained and regional attribution.. M. Sundararajan, A. Taly, and Q. Yan (2017) Axiomatic attribution for deep networks. In International conference on machine learning, p. 3319–3328. Cited by: Fine-grained and regional attribution.. R. Tomsett, D. Harborne, S. Chakraborty, P. Gurram, and A. Preece (2020) Sanity checks for saliency metrics. In Proceedings of the AAAI conference on artificial intelligence, Vol. 34, p. 6021–6029. Cited by: Evaluation of explanations.. Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon (2019) Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog) 38 (5), p. 1–12. Cited by: Event representations and recognition.. C. Yeh, C. Hsieh, A. Suggala, D. I. Inouye, and P. K. Ravikumar (2019) On the (in) fidelity and sensitivity of explanations. Advances in neural information processing systems 32. Cited by: Evaluation of explanations.. M. D. Zeiler and R. Fergus (2014) Visualizing and understanding convolutional networks. In European conference on computer vision, p. 818–833. Cited by: Fine-grained and regional attribution.. H. Zhao, L. Jiang, J. Jia, P. H. Torr, and V. Koltun (2021) Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, p. 16259–16268. Cited by: Event representations and recognition.. T. Zheng, C. Chen, J. Yuan, B. Li, and K. Ren (2019) Pointcloud saliency maps. In Proceedings of the IEEE/CVF international conference on computer vision, p. 1598–1606. Cited by: Introduction, Introduction, Fine-grained and regional attribution..