Random neural networks match brain recording dimensionality, but better tests needed

A new theoretical study by Zhao, Pasek, and Nemenman (arXiv:2605.26551) challenges assumptions about how we interpret neural population recordings. Using Dynamical Mean-Field Theory, the team built a minimal random neural network model incorporating two experimentally relevant factors—finite measurement time and variability across behavioral contexts. They found that the dimensionality measured from large-scale recordings is consistent with predictions from this simple random model, suggesting that observed low-dimensional activity may not require complex circuit architectures.

However, the researchers also demonstrate that current recording durations make it difficult to use dimensionality alone to discriminate among connectivity structures. Crucially, they show that analytically predicted dimensionality varies non-monotonically with external input strength, while the orientation similarity between neural manifolds recorded under different behavioral contexts is more sensitive to network structure. This provides quantitative guidance for designing experiments that can truly infer the connectivity underlying population activity.