Functorial Neural Architectures from Higher Inductive Types
This addresses a fundamental limitation in neural networks for tasks requiring compositional reasoning, offering a novel architectural solution with formal guarantees.
The paper tackles the problem of neural networks failing at compositional generalization by showing it is an architectural issue, and introduces a method to compile Higher Inductive Type specifications into neural architectures that guarantee functoriality, resulting in performance improvements of 2-10x on various spaces.
Neural networks systematically fail at compositional generalization -- producing correct outputs for novel combinations of known parts. We show that this failure is architectural: compositional generalization is equivalent to functoriality of the decoder, and this perspective yields both guarantees and impossibility results. We compile Higher Inductive Type (HIT) specifications into neural architectures via a monoidal functor from the path groupoid of a target space to a category of parametric maps: path constructors become generator networks, composition becomes structural concatenation, and 2-cells witnessing group relations become learned natural transformations. We prove that decoders assembled by structural concatenation of independently generated segments are strict monoidal functors (compositional by construction), while softmax self-attention is not functorial for any non-trivial compositional task. Both results are formalized in Cubical Agda. Experiments on three spaces validate the full hierarchy: on the torus ($\mathbb{Z}^2$), functorial decoders outperform non-functorial ones by 2-2.7x; on $S^1 \vee S^1$ ($F_2$), the type-A/B gap widens to 5.5-10x; on the Klein bottle ($\mathbb{Z} \rtimes \mathbb{Z}$), a learned 2-cell closes a 46% error gap on words exercising the group relation.