Representation learning is central to modern machine learning, yet most research focuses on optimizing representations after a representational framework has been selected. Less attention is given to when a new representational level becomes necessary. We introduce the Bootstrap Theory of Representational Emergence (TBER), a conceptual framework in which persistent explanatory insufficiency acts as a signal for representational transition. A representation may remain descriptively useful while becoming unable to make relevant observations, relations, transformations, or organizational properties intelligible. TBER describes a recursive five-stage process: stabilized observation, anomaly detection, recognition of explanatory insufficiency, representational emergence, and provisional stabilization. The revised framework distinguishes emergence from validation: candidate representations may require problem re-representation, discriminating tests, and representational selection before stabilization. TBER concerns transitions between scientific or computational representations rather than transitions within the physical systems being observed. It provides a meta-representational framework applicable to representation learning, latent spaces, foundation models, world models, adaptive systems, and scientific discovery. A possible implication for future AI is the development of systems capable not only of learning representations, but also of detecting their explanatory limits and initiating, testing, and selecting alternative representational frameworks.

翻译：暂无翻译