Motif-Grounded Uncertainty-Guided Spreading Activation for Efficient 3D Point Cloud Classification

What This Is

A novel architecture for 3D point cloud classification that combines five independently published research ideas into a single pipeline. No existing paper combines all five. That gap is the contribution.

---

The Five Theoretical Pillars

1. Dynamic Node Pruning

Grounded in: Rao et al., DynamicViT, NeurIPS 2021 (arXiv: 2106.02034)

DynamicViT prunes image tokens using attention scores. This work replaces that signal with motif participation and evidential uncertainty — adapting the pruning paradigm from 2D images to 3D point cloud graphs.

---

2. Motif-Based Graph Structure

Grounded in: Bodnar et al., CW Networks, NeurIPS 2021 (arXiv: 2106.12575)

CWN uses higher-order structures (triangles, rings) to increase GNN expressivity on full graphs. Here, motif participation scores are repurposed as a node selection criterion — nodes low in structural significance are pruned rather than retained with extra weight.

---

3. Uncertainty Estimation

Grounded in: Sensoy et al., Evidential Deep Learning, NeurIPS 2018 (arXiv: 1806.01768)

EDL fits a Dirichlet distribution over class probabilities in a single forward pass. Epistemic uncertainty per node (u = K / S) measures how little evidence the network has for any class. High uncertainty = uninformative node = pruning candidate.

---

4. Vision GNN Backbone

Grounded in: Han et al., Vision GNN, NeurIPS 2022 (arXiv: 2206.00272)

ViG treats images as graphs and processes them with alternating Grapher and FFN blocks. This work applies ViG to 3D point cloud graphs — the first such application — operating on the motif-pruned subgraph rather than the full 1024-node input.

---

5. Spreading Activation

Grounded in: Anderson, A Spreading Activation Theory of Memory, JVLB 1983 + Gasteiger et al., APPNP, ICLR 2019 (arXiv: 1810.05997)

Anderson's cognitive model describes activation propagating through associative networks weighted by link strength. APPNP implements this computationally via personalized PageRank. Here, edge weights are set by motif co-participation — activation spreads preferentially through structurally significant pathways.

---

Baseline Comparisons

Model	Paper	What It Does
PointNet++	Qi et al., NeurIPS 2017	Hierarchical point set learning, full input
DGCNN	Wang et al., ACM TOG 2019	Dynamic KNN graphs on full point cloud
ViG	Han et al., NeurIPS 2022	Graph backbone on full node set

This work targets comparable accuracy to the above at significantly reduced node count and FLOPs.

---

Benchmark

Dataset: ModelNet40 — 40-class 3D shape classification, 1024 points per cloud.