Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Latent Brain States Improve Prediction of Neural Responses to Visual Stimuli in Mouse V1

Zijian Jiang¹, Jacob Reimer², Tatiana A Engel¹; ¹Princeton University, ²Baylor College of Medicine

Presenter: Zijian Jiang

Neural activity in the mouse primary visual cortex (V1) is strongly modulated by internal brain states, beyond what is explained by visual input alone. Behavioral variables such as pupil size and locomotion provide partial and indirect access to these states. Here, we introduce a latent-variable framework that infers low-dimensional internal brain states directly from activity of a subset of neurons and integrates these latent states with the visual input image to predict the responses of separate target neurons. These learned latent states significantly improve prediction accuracy over models that also predict neural responses from the image but rely on measured behavioral variables rather than latent internal states. The latent variables did not merely capture redundancy in neural activity, as predictive gains persisted even when the latent states were inferred from a small subset of input neurons and restricted to a very low-dimensional representation. The model revealed that neurons exhibited mixed sensitivity to visual input and internal brain state. Our results support the interpretation of latent variables as meaningful representations of internal brain states.

Topic Area: Computational Models of Vision & Visual Cortex