Reo Tsukasa^1,2, Masatoshi Kasai^1,3, Shuntaro C. Aoki^1,2, Tadashi Isa^1,4, Yukiyasu Kamitani^1,2; ¹Kyoto University, ²ATR, ³Kitasato University School of Medicine, ⁴Institute for the Advanced Study of Human Biology (WPI-ASHBi)

The brain processes visual information through two major pathways: the cortical pathway, which includes the primary visual cortex (V1), and the subcortical pathway, which includes the superior colliculus (SC). Although the cortical pathway has been extensively characterized, the functional role of the SC in visual representation remains comparatively understudied. Recent experimental and analytical advances have begun to address this gap. In particular, progress in optical imaging has enabled single-cell–resolution recordings of SC activity (Kasai and Isa, 2016; 2021), providing new opportunities for direct comparisons between V1 and SC. In parallel, studies in humans have shown that both simple visual features and hierarchical deep neural network (DNN) features can be decoded from activity in the visual cortex (Kamitani and Tong, 2005; Horikawa and Kamitani, 2017). Building on these developments, we used neural decoding approaches to investigate how visual information is represented in the mouse SC relative to V1. We examined visual representations in mouse SC and V1 by decoding features from population activity recorded with two-photon calcium imaging in anesthetized animals. The stimulus set included natural scenes and simple artificial stimuli, allowing comparisons across different types of visual features. Using linear regression decoders, we estimated direction, spatial position, and hierarchical DNN features from neural responses. Both V1 and SC contained sufficient information to predict multiple visual attributes; for example, the direction of bar motion and the spatial position of dot stimuli were reliably decoded. Moreover, DNN-derived features—particularly those from self-supervised models trained on low-resolution images—were decoded with high accuracy. While SC and V1 exhibited broadly similar decoding performance, future work employing more fine-grained analyses—such as unit-level comparisons of decoding accuracy—will be necessary to determine whether more subtle differences in visual representations exist between these two regions.

Acknowledgements: This research was supported by AMED JP24wm0625409 and KAKENHI JP25H00450.