Locomotion is associated with straighter neural trajectories for natural movies in mouse visual cortex

Sensory systems transform inputs from the environment into internal representations that capture aspects of the information useful in downstream computation and behavior. These representations are not rigid, static reflections of the external world but are dynamically modulated by fluctuating cortical and behavioral states. Locomotion, for example, is associated with facilitative changes in stimulus representation in the mouse visual cortex (Vinck et al 2015).
How does sensory processing in the visual system support downstream behavior? "Straightening" hypotheses posit that an important role of the visual system is to straighten representations across time and state space, enabling prediction through simple linear extrapolation (Hénaff et al 2019) and discrimination between objects. Whether this phenomenon depends on behavioral states has not been tested. In this work, we leverage publicly available data and infer the neural trajectories of a natural movie stimulus in the 2D embedding of the stable representation similarity space in locomoting and stationary mice. We show that locomotion influences not only encoding quality but also the structure of the encoding of natural videos, resulting in straighter neural trajectories in multiple visual areas. This observed effect is not merely the result of increased response reliability; it also holds true for single-trial analysis.
Moreover, we find that this effect is greater in the ventral visual area VISl than in the dorsal area VISal, supporting the two-stream hypothesis that attributes object manifold untangling to ventral processing (DiCarlo and Cox 2007). We demonstrate a powerful instance where an etiologically important behavior modulates the transformation of complex, naturalistic stimuli to yield a more accurate and predictive encoding.