NervePool: A Simplicial Pooling Layer

Authors

DOI:

https://doi.org/10.57717/cgt.v5i1.70

Abstract

For deep learning problems on graph-structured data, pooling layers are important for down sampling, reducing computational cost, and to minimize overfitting. We define a pooling layer, NervePool, for data structured as simplicial complexes, which are generalizations of graphs that include higher-dimensional simplices beyond vertices and edges; this structure allows for greater flexibility in modeling higher-order relationships. The proposed simplicial coarsening scheme is built upon partitions of vertices, which allow us to generate hierarchical representations of simplicial complexes, collapsing information in a learned fashion. NervePool builds on the learned vertex cluster assignments and extends to coarsening of higher dimensional simplices in a deterministic fashion. While in practice the pooling operations are computed via a series of matrix operations, the topological motivation is a set-theoretic construction based on unions of stars of simplices and the nerve complex.

 

Downloads

Published

2026-11-16

Issue

Section

Original Research Articles

Categories

How to Cite

NervePool: A Simplicial Pooling Layer (S. McGuire Scullen, E. Röell, E. Munch, B. Rieck, & M. Hirn, Trans.). (2026). Computing in Geometry and Topology, 5(1), 11:1-11:28. https://doi.org/10.57717/cgt.v5i1.70