Why Unification Is Conditional in Deep Learning
Freeborn, David Peter Wallis (2026) Why Unification Is Conditional in Deep Learning. Proceedings of the Biennial Meeting of the Philosophy of Science Association. (In Press)
Abstract
Deep learning systems often understand domains in a fragmented way, using disconnected internal components that work locally rather than a single broadly applicable structure. I call this the Fractured Understanding Hypothesis. I explain it through a conditionality thesis: fragmentation is the default because deep networks can fit data in many ways, and training usually finds patchwork solutions first, with little pressure to unify once prediction is good enough. Unification can occur, but typically only when simpler solutions are strongly favored and the domain contains discoverable regularity. A modular arithmetic case illustrates this transition from memorization to symmetry-respecting algorithmic understanding.
Actions (login required)
 |
Edit Item |