Cortical hierarchy shows rarely categorical yet highly separable representations

Neuroscientists report a unifying framework for why the brain’s single neurons look so different from one another—yet still support powerful, readout-friendly population codes. The work, spanning cortical regions along the hierarchy, begins with a simple observation: neural selectivity profiles are highly diverse. This diversity is not a nuisance. Instead, the authors argue it is a structural resource for building population representations that are both high-dimensional and highly separable.

The key advance is a new way to quantify diversity in individual neurons’ response profiles. Rather than treating selectivity variability as qualitative “noise,” they define α-diversity, a measure grounded in the participation ratio of response profiles in the full space of regression coefficients. In other words, α-diversity captures how many effective dimensions of neuron-to-condition tuning are “occupied” by the observed response structure, integrating both amplitude differences across variables and broader organization in feature space.

The data reveal two recurring modes of structure. Some neurons show uneven selectivity, where encoding is dominated by a subset of task variables or conditions. Others exhibit categorical selectivity, where neurons cluster into functional groupings. These patterns can be illustrated in example regions: one area shows an elongated profile driven mainly by whisking, while another region distributes selectivity more uniformly across dimensions.

Crucially, α-diversity is linked to representational geometry across the cortex. Regions with higher α-diversity exhibit greater population dimensionality, providing more axes along which conditions can be separated. In the study’s empirical tests, α-diversity correlates with an increase in the number of independent conditions (Spearman’s ρ = 0.73), suggesting that richer single-neuron diversity scales up to more expressive population codes.

The authors also show that higher α-diversity aligns with less clustering in the α-diversity space, using a silhouette-score analysis (Spearman’s ρ = −0.76). This provides a geometric interpretation: when response profiles span more effective dimensions, the neural representation is less forced into a few tight groups, enabling better separability.

To connect structure to function, the study directly ties α-diversity to decoding performance. Using a cross-validated linear classifier, they demonstrate that response diversity predicts how many dichotomies of conditions can be decoded. The message is practical and theoretical at once: single-neuron tuning diversity improves population-level readouts by increasing the dimensionality available to linear decision boundaries.

The result reframes “rarely categorical” neural coding as a feature of cortical computation. Along the hierarchy, representations become more separable not by enforcing rigid neuron groups, but by spreading selectivity across many effective dimensions—making linear separation increasingly capable.