New ML Framework Focuses on Learning from Units

Researchers are proposing a fundamental shift in how machine learning tasks are formalized, moving from a sample-centric view to one where the "unit" is an explicit primitive. A unit refers to a persistent individual to which multiple events or observations are linked. This new framework explicitly declares a population of referents and a sameness criterion, with supervised learning specializing in families of unit-conditioned response laws. This contrasts with traditional sample-only conditionals, which remain ambiguous about whether observed laws stem from a homogeneous or heterogeneous world. The learning process under this framework involves learning a tokenizer that generates a contextual unit token and a shared response-law form that interprets it. This token serves as the learner's representation of the task-side unit, allowing the model to account for individual differences. The paper also addresses scenarios where unit identity is unresolved, introducing "unit abduction" to form tokens from factual evidence. It highlights how trusted same-unit pairs can help distinguish heterogeneous unit worlds from homogeneous pooled worlds, offering a more nuanced approach to learning from complex, individual-linked data.