# Interleaving-consistent point-cloud features

- ID: 578
- Canonical URL: https://synthcore.org/idea/578/interleaving-consistent-point-cloud-features
- API JSON: https://synthcore.org/api/idea/578.json
- API Markdown: https://synthcore.org/api/idea/578.md
- Verification status: unverified
- Source: [arXiv:2607.08458](https://arxiv.org/abs/2607.08458)
- Category: regularization
- Solves: stability, generalization, accuracy
- ML areas: graph-nn, regularization, embedding
- Math tags: topology, metric-geometry, graph-theory
- Ratings: usefulness 5/10; difficulty 6/10; novelty 7/10

## Idea description

Regularize a point-cloud or graph neural network so that two augmented versions of the same sample induce filtered proximity graphs with approximately interleaved Reeb graphs. The network is encouraged to preserve multiscale connectivity in learned scalar features, not merely pointwise feature similarity or final predictions. Use an approximate interleaving loss for small graphs and the cheaper H0 persistence-distance surrogate for larger batches.

## Mathematical statement

A filtered Reeb graph is R=(G,h_R), where G is a finite graph and h_R:|G|→R is a scalar filter; its sublevel set at threshold t is R_{≤t}={x∈|G|:h_R(x)≤t}. An ε-interleaving consists of maps Φ:R→U_ε(S) and Ψ:S→U_ε(R), where U_ε shifts the filter by ε, satisfying U_ε(Ψ)∘Φ=ζ_R^{2ε} and U_ε(Φ)∘Ψ=ζ_S^{2ε}; ζ_R^{2ε} is the canonical structure map from level t to level t+2ε. The paper’s identifiability property is d_I(F,G)=0⇒F≅G for the corresponding constructible cosheaves F,G, so zero interleaving distance means the filtered objects are isomorphic. The paper also proves d_B(Ord_0(R),Ord_0(S))≤ε whenever R and S are ε-interleaved, where Ord_0(R) is the ordinary sublevel H0 persistence module and d_B is bottleneck distance. In the neural adaptation, R and S are filtered k-nearest-neighbor graphs from two augmentations, h_R and h_S are learned scalar features, and ε is an augmentation uncertainty scale. The regularizer is max(0,d̂_I(R,S)−ε)^2 or, when full interleaving matching is unavailable, max(0,d_B(Ord_0(R),Ord_0(S))−ε)^2. The bottleneck term is a certified necessary surrogate: an interleaving within ε implies its H0 persistence distance cannot exceed ε, although the converse need not hold.

## Key formulas

- $$d_I(\mathcal{F},\mathcal{G})=0\quad\Longrightarrow\quad\mathcal{F}\cong\mathcal{G}.$$
- $$U_{\varepsilon}(\Psi)\circ\Phi=\zeta_{R}^{2\varepsilon},\qquad U_{\varepsilon}(\Phi)\circ\Psi=\zeta_{S}^{2\varepsilon}.$$
- $$d_B\!\bigl(\mathrm{Ord}_{0}(R),\mathrm{Ord}_{0}(S)\bigr)\leq\varepsilon.$$
- $$\mathcal{L}_{\mathrm{Reeb}}=\max\{0,\widehat d_I(R,S)-\varepsilon\}^{2}\quad\text{or}\quad\mathcal{L}_{H_0}=\max\{0,d_B(\mathrm{Ord}_0(R),\mathrm{Ord}_0(S))-\varepsilon\}^{2}.$$

## Implementation notes

Integrate the method after the point-cloud encoder and before the task head. For each training example, create two augmentations X1 and X2, run the shared encoder fθ to obtain point features Z1 and Z2, and define a scalar filter h_i(p), initially h_i(p)=wᵀz_i(p) with learned vector w or h_i(p)=||z_i(p)||2. Construct a k-nearest-neighbor proximity graph Gi for each view using input-space distances, feature-space distances, or their combination. Discretize filter values into B ordered thresholds. At each threshold t_b, compute connected components of the induced subgraph containing vertices with h_i≤t_b using union-find. Match components between views using point correspondences from the augmentation transform; permit a threshold shift ε and penalize unmatched components. The approximate interleaving distance d̂I is the minimum, over monotone component matchings, of the largest threshold shift plus unmatched-component penalties. If this matching is too expensive, extract H0 persistence intervals from the union-find filtration and compute bottleneck distance dB. Use the loss `task_loss + lambda*relu(d-eps)**2`, where d is d̂I or dB. A practical loop is `Z1,Z2=encoder(X1),encoder(X2)`; `R1=filtered_knn_graph(X1,Z1,h1)`; `R2=filtered_knn_graph(X2,Z2,h2)`; `d=topological_distance(R1,R2)`; `loss=task_loss+lambda*relu(d-eps)**2`; then update θ. Estimate ε as the 95th percentile of observed filter displacement under augmentations, `eps=quantile_0.95(max_p |h1(p)-h2(p)|)`. Use a soft component-membership relaxation for gradients, or evaluate the topology term every K steps as a non-differentiable regularization signal. First test on ModelNet40 or ScanObjectNN with a small PointNet++ baseline, comparing ordinary feature consistency, H0 persistence consistency, and the proposed loss. Success means improved corruption and augmentation robustness at equal clean accuracy, fewer augmentation-induced component births and merges, and limited additional training time.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.