Beyond Kautz Type 5 : A refined taxonomy of symbolic injection architectures

Sammanfattning

Neuro-symbolic AI studies how neural learning and symbolic reasoning can be com- bined. Kautz’s taxonomy is a widely used classification of neuro-symbolic systems, but its Type 5 category remains too broad for precise architectural comparison. It groups together frameworks that all inject symbolic knowledge into neural learning, even though they do so in different ways. This thesis refines Type 5 by separating the question of what makes a framework Type 5 from the question of how it injects symbolic knowledge. Using an established taxonomy development method, thirteen neuro-symbolic frameworks were analysed through a structured extraction template, producing a refined sub-taxonomy with three injection categories: Loss/Constraint- based, Proof/Derivation-based, and Structure/Architecture-based injection. The anal- ysis yields three findings: the categories involve a tradeoff between flexibility and coupling depth; a framework’s Kautz type can depend on how it is used; and the frameworks still rely on human-authored symbolic knowledge. A small experimental evaluation, running one framework from each category on a shared digit-addition task, supported the view that the central distinction corresponds to measurable differences in data efficiency and computational cost rather than only to differences in description. The contribution is a scoped refinement of Kautz Type 5 that gives researchers clearer language for comparing Type 5 architectures, together with an initial empirical check that the distinction is real.