Neuro-Symbolic Geometric Abstraction (NeuSOGA): From Observations to Symbolic Mathematical Representations
A paper titled 'Neuro-Symbolic Geometric Abstraction (NeuSOGA): From Observations to Symbolic Mathematical Representations' was published on arXiv on 2026-09-01. It proposes a framework that transforms observations into topological abstractions, geometric abstractions, and symbolic mathematical representations, combining topology-guided structural discovery, foundation-model perception, adaptive multi-scale geometric abstraction, and symbolic synthesis.
The paper introduces NeuSOGA, a neuro-symbolic framework for converting geometric observations into explicit symbolic mathematical representations. It uses Euclidean Distance Transforms for topology-guided structural discovery, Segment Anything for foundation-model perception, adaptive multi-scale geometric abstraction, and Implicit Area Splines for symbolic synthesis. The resulting representation is an analytical implicit model supporting arbitrary-order smoothness and additive composition.
NeuSOGA integrates topology-guided structural discovery with foundation-model perception and symbolic synthesis via Implicit Area Splines. The use of Euclidean Distance Transforms suggests a focus on shape topology, while Segment Anything provides segmentation priors. The symbolic output is an implicit analytical model with arbitrary-order smoothness, enabling differentiable and composable representations.
This research addresses the gap between perceptual AI and interpretable symbolic reasoning. If successful, such frameworks could enable more transparent and mathematically manipulable AI systems, with potential applications in CAD, scientific modeling, and automated theorem proving.
The framework could lead to tools that automatically generate editable mathematical models from images or sensor data, reducing manual modeling effort in engineering and scientific workflows.
Next observable signals include follow-up papers evaluating NeuSOGA on benchmark geometric abstraction tasks, open-source code releases, or applications in domains requiring symbolic mathematical representations from visual data.