Neural Networks Show Functional Modularity Beyond Individual Neurons

Summary

Neural networks often distribute multiple functions across the same individual neurons, a pattern commonly interpreted as evidence that they lack modular organization. This study proposes a nested hierarchy with three levels: unit-level modularity, where computations use disjoint neurons; orthogonal modularity, where a rotation exposes independent functional subspaces; and linear modularity, where a bounded linear map separates task-relevant computations. The authors develop a unified method for learning these decompositions and test them causally through selective ablation. Across vision models, language models, and multi-task recurrent neural networks, systems that fail the strict unit-level test frequently show clear modularity at the subspace level. In vision models, content and style separate under subspace ablation even when their neurons overlap. In language models, syntax and semantics likewise dissociate. The learned subspaces also show different relationships to brain data: content and style subspaces predict higher and lower visual-cortex activity respectively, while semantic subspaces predict language-cortex activity better than syntactic subspaces. In recurrent networks trained on multiple cognitive tasks, the method recovers sharp double dissociations between tasks that had previously appeared non-modular. The results suggest that targeted intervention on specific computations can remain possible without individually specialized neurons, with implications for mechanistic interpretability and the study of functional organization in biological and artificial neural systems.