Shaping Generalization in Deep Learning: The Role of Organizational Factors

Starting from a case study of counting, I will show how the organizational factors of learning must be transformed to shape generalization in deep learning — beyond what any amount of data or model scaling can achieve. I will then introduce a novel inductive learning paradigm in which models and data co-evolve. This approach enables the discovery of generative rules — what I call “successorship” — among observations, requiring the recursive revision of hypotheses to capture progressively more complex computations.