Accessibility Plasticity: A New Principle for Adaptive AI

Modern neural networks primarily achieve adaptation by modifying parameters within their predefined computational structures. While recent advancements have introduced modularity, conditional computation, and parameter-efficient adaptation, these methods generally do not distinguish between a system's inherent computational capability and the accessibility or interaction patterns among those computations as separate adaptive variables. This research proposes "Accessibility Plasticity" as a novel principle for adaptive computation. It posits that intelligent systems can adapt not only by altering the computations that exist but also by dynamically reorganizing which existing computations are able to interact and participate. This concept establishes a "reuse-first" hierarchy of adaptation, where modifying accessibility is prioritized over more computationally expensive changes to core capabilities or structural elements. A proof-of-concept evaluation conducted on sequential learning tasks demonstrated that adapting computational accessibility can effectively reduce the need for modifying core capabilities while maintaining comparable task performance. These findings suggest that accessibility represents a distinct and valuable adaptive dimension, laying a foundation for future dynamic neural systems whose computational relationships can evolve efficiently with changing environments.