Geometric Interpretation of Semantic Relationships: Filtering and Signature Formation for Neural Network Interoperability

Abstract

This paper proposes a method for filtering vector representations and forming universal signatures based on the geometric interpretation of semantic space as points on a hypersphere. The algorithm combines differential-geometric characteristics with statistical metrics to create a unified representation ensuring compatibility between various neural networks. The method reduces data dimensionality by selecting the most informative connections and forming compact signatures that are invariant to architectural features of models. Experiments confirm a reduction in computational complexity while simultaneously improving analysis quality. This approach potentially establishes a foundation for a universal knowledge exchange interface between heterogeneous neural network systems.