When a neural network processes text or images, it builds internal representations: high-dimensional vectors that encode meaning. But the fundamental building blocks of these representations have not been clearly identified. Common candidates like single neurons are often polysemantic, meaning they respond to several unrelated concepts, which makes interpreting them difficult.
This work formally defines Fundamental Representational Units (FRUs) as the basic atoms of representation, with two key properties: fidelity (they preserve information) and stability (they are structurally distinguishable). The authors prove that threshold-activated sparse autoencoders (TSAEs) can identify these atoms under certain conditions, and validate this with experiments across various models.
The identified atoms are more monosemantic (each corresponds to a single concept) and can be composed across layers to form higher-level concepts that align with human conceptual structure. This provides a unified framework for understanding how neural networks organize knowledge and offers a principled basis for interpretability and interventions.