A Geometric Framework for Continual Learning via Holomorphic Extension and Tensor-Geodesic Dynamics

Read full text Summarise with VIRAT
Authors: Edward Meisner
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This article introduces a unified geometric framework for continual learning based on complex manifolds, holomorphic extension, curvature constraints, and tensor‑geodesic reasoning. Modern neural systems often rely on heuristic mechanisms—such as shared subspaces, basin repair, and lazy expansion—to mitigate catastrophic forgetting, but these methods lack an underlying conceptual or geometric foundation. The framework developed here replaces the flat Euclidean representation space with a complex Hermitian manifold whose curvature and phase structure define stable concept regions. Learning is modeled as holomorphic extension, which preserves these regions and ensures stability across sequential tasks. A curvature‑energy functional provides a principled novelty threshold that determines when new representational directions must be introduced. On top of this geometric substrate, the paper introduces a tensor‑logic interpretive layer, where symbolic expressions correspond to geometric constraints and reasoning corresponds to selecting minimal‑action geodesic paths. This yields a coherent account of learning, memory, and inference within a single geometric‑computational system. The framework also explains recent empirical findings—such as the convergence of LoRA update directions to shared low‑rank subspaces—as manifestations of deeper geometric invariants.

Fields of Study & Subjects

Domain Adaptation and Few-Shot Learning Neural Networks and Applications Machine Learning and Algorithms

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
30 March 2026
Language
EN