Published in Annual Meeting, Neuroscience, Nov 2018.
Many behaviors rely on predictions about the future state of the environment derived from recent observations. Prediction is most robust along a straight, or linear trajectory. Yet, under natural circumstances, the stream of visual input typically evolves according to highly non-linear dynamics. It thus stands to reason that computations within the visual system work to linearize the temporal dynamics of visual representations, thereby enabling more robust predictions of future states of the environment. Recent psychophysical work has provided strong evidence for this hypothesis by showing that the human visual system selectively straightens the temporal trajectories of natural image sequences, thus facilitating their extrapolation (Hénaff, Goris, Simoncelli, 2018). How does straightening emerge from the cascade of transformations performed by the visual system? We hypothesize that the successive stages of processing incrementally linearize the dynamics of visual representations. Here, we investigated the straightness (conversely, curvature) of representations of natural image sequences in the primary visual cortex (V1). We recorded neural population activity in area V1 of anesthetized macaque monkeys using multi-electrode laminar arrays while briefly presenting individual video frames taken from natural videos. As a control, we also presented stimuli taken from "unnatural" synthetic videos to determine whether neural straightening is specific to behaviorally relevant sequences. We developed a new computational analysis technique to quantify the "straightness" of a population vector's trajectory. Preliminary results from five V1 populations (ranging in size from 26-61 units) reveal that neural representations of natural videos exhibit trajectories that are on average straighter than the trajectories described by the stream of incoming light intensities. Moreover, neural straightening is limited to natural videos. For our synthetic, unnatural videos, we found an increase in the curvature of the neural representation relative to the intensity domain. Together, these results suggest that computations in the early visual system contribute to realizing a major behavioral goal: making the visual environment predictable in time.