Core Concepts
DECT enables end-to-end learning of ECT, enhancing shape classification across various modalities.
Abstract
The Differentiable Euler Characteristic Transform (DECT) overcomes limitations of the ECT by enabling task-specific representations through end-to-end learning. It provides efficient and scalable performance in shape classification tasks for graphs, point clouds, and meshes. The ECT and Persistent Homology Transform (PHT) are based on multi-scale topological descriptors for shapes. DECT is differentiable with respect to directions and coordinates, allowing integration into deep neural networks. The method offers a flexible framework for shape representation and optimisation based on topological constraints.
Stats
arXiv:2310.07630v2 [cs.LG] 15 Mar 2024
Quotes
"As a motivating example, we study how learning directions affects the classification abilities of DECT."
"Our method is applicable to different data modalities—including point clouds, graphs, and meshes—and we showed its utility in a variety of learning tasks."
"DECT outperforms existing graph neural networks while requiring a smaller number of parameters."