RESEARCH · RESEARCH · #1851
Neural networks are more modular at the subspace level than single neurons suggest
The paper (arXiv:2610.12842v1) defines a hierarchy of modularity—unit-level, orthogonal, and linear—and presents a unified framework to learn and causally evaluate decompositions at each level via selective ablation. Across vision models, language models, and multi-task RNNs, the authors find that distinct computations often occupy separable subspaces despite overlapping neuron populations; these subspaces also differentially align with brain activity in visual and language cortex.
KEY POINTS
- The paper (arXiv:2610.12842v1) defines a hierarchy of modularity—unit-level, orthogonal, and linear—and presents a unified framework to learn and causally evaluate decompositions at each level via selective ablation.
- Across vision models, language models, and multi-task RNNs, the authors find that distinct computations often occupy separable subspaces despite overlapping neuron populations; these subspaces also differentially align with brain activity in visual and language cortex.
- This shows that functional modularity can exist at the subspace level even when individual neurons are multifunctional, enabling more targeted interpretability and interventions in artificial and biological networks.
WHY IT MATTERS
This shows that functional modularity can exist at the subspace level even when individual neurons are multifunctional, enabling more targeted interpretability and interventions in artificial and biological networks.