The superior colliculus response space has globally high- and locally low-dimensionality

An understanding of cell types is essential for understanding neural circuits, but only when the response of each type is clearly defined and predictable, as has been observed in the retina1. Recent work has shown that neural responses in the visual cortex are of high dimensionality, questioning the validity of defining cell types in the deeper visual system2-4. Here we investigate the dimensionality of neural responses in the midbrain using two-photon calcium imaging in superficial layers of the mouse superior colliculus (SC). Responses of individual neurons to closely related stimuli, such as ON and OFF light signals, were mutually dependent such that the response to one stimulus could be predicted from the response to the other. In contrast, individual neurons responded to brightness and motion in a statistically independent manner, maximizing functional diversity but preventing traditional cell type classification. To capture the globally high, locally low dimensionality of neural responses, we propose a multidimensional response model, in which classification of cellular responses is meaningful only in local low-dimensional structures. Our study provides a framework to investigate the processing of visual information by the SC, which likely requires a high-dimensional response space structure5,6 to perform higher-order cognitive tasks7-12.