acceptodds
Under review as a conference paper at ICLR 2027

Spiking DGCNN: Max-First Graph Spiking Networks for Efficient Point Cloud Learning

Abstract

Spiking neural networks (SNNs) offer sparse, event-driven computation for energy-efficient learning, but the naive spiking conversion of graph-based point cloud networks remains poorly understood. EdgeConv relies on neighbourhood-wise max aggregation to select discriminative local responses. Simply replacing LeakyReLU with spiking neurons in EdgeConv quantizes features before aggregation, weakening credit assignment and causing severe performance degradation. We therefore propose Spiking DGCNN, built on a Max-First Spiking Rule that performs neighbourhood selection on continuous-valued features before spike generation. We further introduce an Adaptive Pseudo-Temporal Expansion-Compression (APTEC) module that supports lightweight pseudo-temporal evidence accumulation through leaky spiking integration, adaptive thresholding, and binary spike compression. Ablation experiments show that Max-First Spiking Rule improves Spike-before-Max baseline from 79.09% to 91.77% at T=1 and from 84.11% to 92.18% at T=4 on ModelNet40. The complete model with APTEC achieves 92.38% accuracy, closely matching the 92.9% of DGCNN, while reducing the operation-level energy estimate from 11.23 mJ to 0.22 mJ.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Point Cloud Processing, Efficient Computing, Brain-inspired Computing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_pn6ROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks where spike generation should sit relative to neighbourhood max aggregation when converting DGCNN’s EdgeConv into a spiking point-cloud network. It argues that spiking each edge response and then taking a max (Spike-before-Max) destroys the neighbour competition that makes EdgeConv work, and it proposes a Max-First rule that takes the channel-wise max of continuous edge features before the spike nonlinearity. A second module, APTEC, avoids unrolling dynamic graph construction over true timesteps by repeating a feature across a short pseudo-time window, integrating it with a leaky spiking neuron, applying an input-dependent threshold adapted from membrane-potential rectification, and compressing the window with a logical OR.

Validation is on ModelNet10/40, ScanObjectNN, and ShapeNetPart, with S3DIS in the appendix. The main accuracy jump comes from Max-First: on ModelNet40, Spike-before-Max scores 79.09% (T=1) and 84.11% (T=4), while Max-First real-temporal scoring reaches 91.77% and 92.18%. The full APTEC model (1.8M parameters, T=4) reaches 92.38% on ModelNet40 and 85.1% instance mIoU on ShapeNetPart, close to ANN DGCNN (92.9% and 85.2%), and the authors estimate energy at 0.22 mJ versus 11.23 mJ for DGCNN. On S3DIS, Max-First also helps, but APTEC (42.23% mIoU) falls well behind real-temporal Max-First (50.03%, against DGCNN at 51.51%).

Critical strengths

The Max-First ablation isolates a large, recurring failure of naive spiking conversion. Replacing Spike-before-Max with neighbourhood max before the spike raises ModelNet40 from 79.09% to 91.77% at T=1 and from 84.11% to 92.18% at T=4 (Table 3), with the same direction of effect on ModelNet10 and ScanObjectNN PB-T50-RS. The appendix repeats the pattern on S3DIS: 38.38% to 48.18% mIoU at T=1, and 42.01% to 50.03% at T=4 under real-temporal processing (Table 8). This is consequential because the comparison holds the backbone fixed and changes operator order, so the gain is not an artifact of a larger Transformer or Mamba. It is the result that could justify a paper even if the rest of the system were scaled back.

The paper reports where its temporal shortcut fails. APTEC is not uniformly declared a win. Table 8 shows real-temporal Max-First within 1.5 points of DGCNN on S3DIS, while pseudo-temporal compression drops to 42.23%. That split makes the ModelNet results readable: most of the recovered accuracy is the ordering rule, and pseudo-time is a task-dependent approximation. I weigh this more heavily than the absolute ModelNet numbers, which sit on a saturated benchmark.

Critical weaknesses
  1. How is FPS-KNN deployed using asynchronous hardware during inference?
  2. Table 1 should cite the source paper for each compared method. The baselines are also not current; the paper should at least report performance against related work published at ICLR 2026.
Questions for authors

See the Weaknesses.

Reviewer_LEjxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adapts DGCNN to spiking computation by placing neighborhood max aggregation before spike generation. It also introduces APTEC, which repeats cached features across pseudo-timesteps, applies input-dependent thresholds, and compresses spikes into a binary output. Experiments show substantial improvements over Spike-before-Max and competitive object classification and part segmentation performance. Additional S3DIS experiments show that Max-First remains useful, while APTEC can substantially degrade scene segmentation performance.

Critical strengths

The ordering ablation identifies a consequential implementation choice. Max-First improves ModelNet40 accuracy from 79.09% to 91.77% at T=1 and S3DIS mIoU from 38.38% to 48.18% (Tables 3 and 8). These consistent improvements across classification and dense prediction make this a useful empirical finding.

The resulting 1.8M-parameter model retains much of DGCNN’s object-level performance: 92.38% versus 92.9% on ModelNet40 and 85.1% versus 85.2% instance mIoU on ShapeNetPart. Reporting APTEC’s substantial degradation on S3DIS also provides a valuable qualification of its applicability.

Critical weaknesses

The methodological contribution is insufficiently distinguished from existing pooling-order designs and alternative optimization explanations. Pooling continuous responses before subsequent binarization has precedents in XNOR-Net and SNN pooling-placement studies, while Eqs. (4) and (6) have identical hard outputs under a shared threshold. The single-step improvement therefore concerns gradient routing, and APTEC’s identical repeated inputs also raise the possibility of an equivalent static gate; without controls isolating these mechanisms, the methodological advance remains uncertain.

The energy comparison excludes computations directly associated with the proposed architecture. Appendix D counts convolutional and fully connected synaptic operations but excludes dynamic neighborhood construction, continuous max aggregation, and APTEC’s membrane and gating computations. The reported 30–50× reduction consequently supports a restricted synaptic-energy estimate, while the practical efficiency benefit and claimed reduction in training cost remain unverified by matched runtime or memory measurements.

Questions for authors
  1. How are ties handled when backpropagating through the max of binary responses? At T=1, does routing the Spike-before-Max backward pass through the continuous-valued argmax recover Max-First’s improvement? This would directly test the proposed credit-assignment explanation.

  2. Under zero initialization and a monotone MPR, does APTEC’s OR-compressed forward output reduce to an equivalent static gate? If so, what benefit remains compared with this collapsed gate and a tuned constant-threshold baseline?

  3. What additional energy costs arise from graph construction, neighborhood aggregation, and APTEC updates? Please provide an expanded operation breakdown and matched RT/PT timing measurements to establish the scope of the efficiency claim.

  4. Please specify the exact MPR function, decay factor, reset-gradient convention, temporal-OR backward rule, and representation used for kNN construction. Which features are cached, and which computations are repeated at each layer and timestep?

Reviewer_ZzULOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper adapts DGCNN's EdgeConv to spiking networks. Its main point is that spiking before the neighborhood max turns the max into a logical OR over neighbors and removes the magnitude ordering EdgeConv relies on. The Max-First rule takes the max over continuous edge responses and only then emits spikes. The second component, APTEC, repeats a feature over T pseudo-timesteps, integrates it with a leaky neuron under an input-dependent threshold, and ORs the spikes into one binary output, so the dynamic graph is built once. With 1.8M parameters the model reaches 92.38% on ModelNet40 (DGCNN: 92.9%) and 85.1 instance mIoU on ShapeNetPart, at a modeled 0.22 mJ against 11.23 mJ for the ANN.

Critical strengths

1.Max-First is one rule with a short, checkable argument (Eqs. 4 to 14, Fig. 1). The ablation supports it: on ModelNet40 it lifts accuracy from 79.09% to 91.77% at T=1 and from 84.11% to 92.18% at T=4 (Table 3), and the same ordering holds on S3DIS (38.38 to 48.18 mIoU at T=1, Table 8).

2.Against SPMamba (12.8M parameters) the model matches accuracy with 1.8M (Table 1).

3.The energy protocol is written out in Appendix D, and Table 7 lists per-layer MACs, ACs, and firing rates.

4.The authors report a result that goes against APTEC: on S3DIS the real-temporal variant clearly beats it (50.03 vs 42.23 mIoU).

Critical weaknesses

•APTEC's value depends on the dataset. Going from T=1 to T=4 adds 0.4 to 0.6 points on ModelNet, and 2.0 to 3.8 points on ScanObjectNN (Table 1). On S3DIS the APTEC model (42.23 mIoU, Table 8) is six points below plain Max-First at T=1 (48.18). The abstract and conclusion describe APTEC as a general mechanism, but the evidence supports "useful on some small benchmarks".

•The case that pseudo-time accumulates evidence is thin. The input is identical at every step and V_th(t) = 1 + 0.5·σ(x) depends only on x, so the threshold is constant over t. The paper's own analysis (Eqs. 18 to 20) shows that with a constant threshold the output is a thresholded, rescaled copy of x. Appendix A argues that the adaptive threshold improves gradients, which is an optimization claim and says nothing about accumulation. Most of the adaptive threshold's gain also appears at T=1, where there is no time axis (73.07 to 75.57 on PB-T50-RS, Table 6). A T=1 model whose threshold is rescaled to match the effective threshold at T=4 would show whether APTEC is more than a threshold shift.

•The energy comparison favors the SNN by construction. For APTEC, Appendix D sets the temporal factor to 1 when counting synaptic operations, so the T repeated neuron updates cost nothing, and the drop from 0.38 mJ at T=1 to 0.22 mJ at T=4 comes only from lower firing rates in the heavy layers (Table 7). The count leaves out dynamic kNN construction in every EdgeConv layer, membrane updates, the sigmoid, rectifier, and division in the adaptive threshold, the max and the OR, and memory traffic. The ANN figure (all MACs at 4.6 pJ) skips the same items, so the 30x to 50x ratio compares synaptic operations only.

•Gains over published SNN baselines are small. ModelNet40 is 92.38 against 92.30 for SPMamba. On ScanObjectNN the margins are 0.2 to 0.3 points, and the SPMamba figures are the authors' own 1024-point reproductions (Table 1 footnote). On ShapeNetPart, SPMamba has the higher class mIoU (82.29 vs 82.01). No seeds or variances are reported, so margins of this size cannot be told apart from noise. •The headline Max-First gain is measured against a baseline built to fail. Spike-before-Max is the naive conversion, so a 12-point improvement says the rule matters but not how far it moves the state of the art.

•The generality claim is untested. The conclusion presents Max-First as a principle for max-based operators, yet every experiment uses EdgeConv. PointNet and PointNet++ also rely on max pooling and would be a direct test. Sensitivity to k (fixed at 20) and to the leak λ is not studied. Table 4 takes the SPTransformer and SPMamba energy values from elsewhere, possibly under different input sizes, timesteps, and protocols. Qiu et al. (2025), a spike-driven sparse 3D network cited in the introduction, is absent from Table 1.

Questions for authors

1.Can you report mean and standard deviation over several seeds for Tables 1, 3, and 6, and say which of the 0.1 to 0.5 point differences survive?

2.Why does the T=4 APTEC model fall six mIoU below Max-First at T=1 on S3DIS? Is the OR compression responsible?

3.For a fixed input, how do membrane potentials and spikes change across pseudo-timesteps? Does a T=1 model with a matched effective threshold reach the same accuracy?

4.What do the ANN and SNN energies look like once kNN construction, T-fold neuron updates, the threshold computation, and memory access are included?

5.Does Max-First help PointNet or PointNet++ with global or grouped max pooling?

6.Where do the energy values for SPTransformer and SPMamba in Table 4 come from, and were they computed with the same 45 nm model, input size, and T?

7.How sensitive are the results to k and λ, and how does Max-First behave on partially static graphs?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.