REFRAMING NEURAL SELECTIVITY THROUGH FEATURE–REPRESENTATION COMPATIBILITY

We aimed to explain why reported neural selectivity—including linear and nonlinear mixed selectivity—can vary across studies, and to propose feature–representation compatibility as a unifying criterion for interpreting “what neurons see” from finite experimental designs. In this view, selectivity is not an intrinsic label of a circuit (or area) alone, but a function of the interaction between the recorded circuit (population) and the feature description used to parameterize stimuli. We define a representation as the structure of population activity in the observed neural space, and treat features as coordinate systems over stimuli—from raw stimulus coordinates to experimenter-defined, human-interpretable descriptors—while emphasizing that inclusion in a feature description does not guarantee compatibility with the population representation. Our approach compares selectivity estimates under two feature descriptions: (i) experimenter-defined features that may be incompatible with the intrinsic geometry of population activity, and (ii) representation-compatible features constructed to match the organization of the recorded neural population (or an analogous model population). Using hidden-layer units in deep networks as controlled proxies for recorded populations, we show that the same units can appear to exhibit nonlinear mixed selectivity when analyzed in incompatible feature coordinates, yet display predominantly linear selectivity when re-expressed in compatible coordinates. These results suggest that apparent nonlinearities may reflect a mismatch between experimental feature coordinates and the intrinsic structure of neural activity, rather than a fixed coding principle, and that this possibility should be taken into account when interpreting mixed selectivity.