Learning Structure, Not Just the Weights: Topology-Adaptive Neural Systems

This public technical disclosure and position paper introduces and documents the concept of Neuroplastic AI: topology-adaptive neural systems in which structural change may form an integral part of learning, adaptation, or deployment. Unlike fixed-topology neural networks, where learning is primarily confined to the optimization of scalar weights within a predetermined architecture, the disclosed approach treats structural operations—such as growth, pruning, rewiring, modular isolation, or other dependency-aware modifications—as possible components of the learning and adaptation process itself. The paper situates this approach in relation to neural architecture search, neuroevolution, dynamic sparse training, mixture-of-experts systems, pruning methods, continual learning, and biologically inspired structural plasticity. It does not present a single reference implementation or benchmark result. Instead, it defines a broad technical design space covering multiple realization paths, trigger mechanisms, structural units, adaptation regimes, and control strategies. The disclosure is technically related to prior work by the author on functional data structures, variable networks, dependency-aware change handling, and graph-based computation, including the patent family around EP3896579A1 / WO2021209336A1 and the separate bubbleCalc technical disclosure. These related materials provide context for controlled structural adaptation and dependency-managed computation, but do not limit the scope of the present disclosure. This document is published to establish a dated, citable public technical disclosure of the described approach and to document authorship of the Neuroplastic AI concept. It is a strategic technical disclosure and position paper, not a grant of rights. No Creative Commons license is granted. All copyright and patent rights are reserved.