Title: Scaling Laws Drive Divergent Selectivity in Neuron Populations

Abstract:

This study examines whether the composition of neuron populations within neural networks undergoes predictable changes as model scale increases, thereby extending the application of scaling laws beyond traditional macroscopic metrics like loss. To address this, we focus on "Rosetta Neurons," a distinct class of units identified by Dravid et al. (2023) for their consistent activation patterns across independently trained models. Through separate analyses of language models reaching 30 billion parameters and vision models up to 5 billion parameters, we find that the total count of Rosetta Neurons increases with model size according to a sublinear power law. Consequently, while their absolute numbers grow, they constitute a diminishing proportion of the total neuronal population.

We also identify a "Neuron Polarization Effect," wherein Rosetta Neurons become increasingly monosemantic and selective as models scale, effectively diverging from the expanding non-Rosetta population, which retains lower selectivity. An analytical framework that weighs feature utility against the constraints of limited neuron capacity provides a theoretical explanation for both the sublinear scaling and this polarization phenomenon. Furthermore, our findings indicate that Rosetta Neurons exhibit heightened domain specialization at larger scales. We demonstrate their specific selectivity through a case study involving targeted data filtering for continued pretraining. These results establish a scaling law for interpretable, shared neuron-level structures, connecting model size to systematic shifts in neuron universality, selectivity, and specialization.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC